JID: FMRE

ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

Fundamental Research xxx (xxxx) xxx

Contents lists available at ScienceDirect

Fundamental Research
journal homepage: http://www.keaipublishing.com/en/journals/fundamental-research/

Review

Role of artificial intelligence in revolutionizing drug discovery

Ashfaq Ur Rehman a,b,1, Mingyu Li a,1, Binjian Wu a,1, Yasir Ali c,1, Salman Rasheed d, Sana Shaheen e, Xinyi Liu a,f,
Ray Luo b, Jian Zhang a,f,∗
a
Medicinal Chemistry and Bioinformatics Center, Shanghai Jiao Tong University School of Medicine, Shanghai 200025, China
b
Departments of Molecular Biology and Biochemistry, University of California Irvine, Irvine, CA 92697, United States
c
Institute of Chemistry, Slovak Academy of Sciences, 845 38 Bratislava, Slovakia
d
National Center for Bioinformatics, Quaid-e-Azam University, Islamabad 44000, Pakistan
e
Key Department of Biochemistry, Abdul Wali Khan University Mardan, Khyber-Pakhtunkhwa 23200, Pakistan
f
State Key Laboratory of Medical Genomics, National Research Center for Translational Medicine at Shanghai, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai
200025, China

a r t i c l e i n f o a b s t r a c t

Article history: The application of artificial intelligence (AI) in medicine, particularly through machine learning (ML), marked
Received 30 December 2023 a significant progression in drug discovery. AI acts as a powerful catalyst in narrowing the gap between disease
Received in revised form 9 April 2024 understanding and the identification of potential therapeutic agents. This review provides an inclusive summary
Accepted 24 April 2024
of the latest advancements in AI and its application in drug discovery. We examine the various stages of the
Available online xxx
drug discovery process, starting from disease identification and encompassing diagnosis, target identification,
screening, and lead discovery. AI’s capability to analyze extensive datasets and discern patterns is essential in
Keywords:
these stages, enhancing predictions and efficiencies in disease identification, drug discovery, and clinical trial
Artificial intelligence
Machine learning management. The role of AI in expediting drug development is emphasized, highlighting its potential to analyze
Disease identification vast data volumes, thus reducing the time and costs associated with new drug market introduction. The impor-
Drug design tance of data quality, algorithm training, and ethical considerations, especially in patient data handling during
Drug discovery clinical trials, is addressed. By considering these factors, AI promises to transform drug development, offering
significant benefits to patients and society.

1. Introduction As a result, there has been an increasing adoption of AI algorithms in

the drug discovery process by chemical and biological scientists.
The traditional drug discovery process is a complex and challeng- ML is extensively used in drug discovery, employing algorithms such
ing endeavor that can require up to 15 years and over $1 to 2 billion for as DL, Bayesian network (BN), random forest (RF), clustering, and sup-
each approved drug [1]. This is primarily due to rising attrition rates and port vector machine (SVM). The broad categorization of ML can be seen
extended clinical trial duration [2]. Despite considerable investment in in Fig. 1. DL models process and analyze large amounts of data in tasks
resources, almost 90 % of potential drug candidates fail even after they such as clinical imaging [8,9], virtual screening (VS) [10,11], and bioac-
have advanced to the phase-I clinical trial [3]. Advancing a drug candi- tivity predictions [12]. BNs predict toxicity or bioactivity and patients’
date to phase-I clinical trial after rigorous optimization at the preclinical response to treatments [13]. RF models are used for molecular target
stage is considered a significant milestone for both pharmaceutical com- identification and feature selection, while clustering identifies patterns
panies and academic institutions [4]. or relationships within data [14]. SVM is a supervised learning algo-
Large-scale computational screening and docking have been em- rithm used to classify data into categories, with applications such as
ployed to enhance the success rate of lead compounds in clinical trials predicting pharmacokinetic properties, VS, and toxicity prediction [15].
[5]. However, these methods have limitations such as inefficiency and Computational modelling based on AI and ML has made various drug
inaccuracy [6]. To overcome these challenges, deep learning (DL) and discovery processes achievable, including chemical compound identifi-
ML algorithms, which are subsets of AI, have been identified as potential cation, target identification, peptide synthesis, drug toxicity and phys-
solutions [7]. These AI tools possess the ability to predict macrosystem iochemical property assessment, drug monitoring, drug efficacy and
properties with high accuracy while incurring low computational costs. effectiveness assessment, bioactive agent prediction, protein-protein

∗
Corresponding author.
E-mail address: [email protected] (J. Zhang).
1 These authors contributed equally to this work.

https://doi.org/10.1016/j.fmre.2024.04.021
2667-3258/© 2024 The Authors. Publishing Services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd. This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Please cite this article as: A.U. Rehman, M. Li, B. Wu et al., Role of artificial intelligence in revolutionizing drug discovery, Fundamental Research,
https://doi.org/10.1016/j.fmre.2024.04.021
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

Fig. 1. Artificial intelligence and subtypes. Machine learning can be broadly categorized into supervised, unsupervised, and reinforcement learning, which can be
further divided into several subcategories. For supervised learning, these include classification, where algorithms categorize input data into predefined classes, and
regression, aimed at predicting continuous outcomes from input features. Additionally, unsupervised techniques such as clustering are used to grouping similar data
points. These subcategories are fundamental to the architecture of machine learning algorithms, significantly contributing to diverse applications in drug discovery.

interaction, protein folding and misfolding identification, structure Gypsum-DL, ENRI, and SPOT-ligand 2. Drug repositioning involves in-
and ligand-based virtual screening (LBVS), QSAR modeling, and drug vestigating drugs developed for one diseased condition and reposition-
repositioning [16]. ML algorithms have made it faster to identify lead ing them for other diseased conditions in drug designing and discovery.
compounds from chemical libraries that contain over 106 million This approach may be successful due to the possibility of multiple-target
compounds and solve the toxicity challenges caused by off-target involvement in multiple diseases. The emergence of AI-based tools and
interactions [17]. Furthermore, ML-based tools such as AlphaFold have algorithms in drug discovery provides a platform for future research,
made it easier to predict the 3D structure of a target protein, which is and different AI-based algorithms and web-based tools have been devel-
crucial in drug discovery process. oped in recent years, such as DRIMC, DrugNet, DPDR-CPI, PHARMGKB,
AlphaFold is a recently developed AI-based tool from Google’s Deep- PROMISCUOUS 2.0, and DRRS.
Mind. Researchers have also used AI to discover novel peptides for ther- This review aims to provide a thorough overview of AI’s role in drug
apeutic purposes. Provenzano et al. [18] created Deep-AmPEP30, a DL- discovery (Fig. 2). AI can aid in various stages of drug discovery in var-
based platform for the identification of short anti-microbial peptides ious ways, including disease identification, target acquisition, compu-
(AMPs) [18], while Yi et al. [19] devised ACP-DL, a DL-based tool us- tational screening, predicting drug toxicity, gene editing for developing
ing the LSTM algorithm, for the discovery of novel anti-cancer peptides gene therapies, and AI-based modeling for personalized drug dosing. We
(https://github.com/haichengyi/ACPDL) [20]. AI is increasingly used examine the current state-of-the-art AI technology and its potential to
in determining the proper drug dosage. Shen, Liu et al. [21] created revolutionize drug discovery.
the AI-PRS platform, a neural network-driven methodology that uses a
parabolic response curve (PRS) to associate drug combinations and dose 2. AI-based disease identification
to efficacy [22], while machine and statistical learning techniques in-
cluding k-nearest neighbor (kNN), Naive Bayesian, SVM, ANN, decision AI has shown great potential in identification of infectious diseases.
trees (DT), and RF are employed to forecast the hindrance in protein- By analyzing large amounts of data from various sources such as EHRs
protein interactions [23]. VS is an efficient method in computer-aided (electronic health records), social media, and news reports, AI can
drug design (CADD), which involves screening a promising therapeutic quickly detect outbreaks of infectious diseases and provide early warn-
compound from a pool of compounds. ing systems. AI can also assist in predicting the spread of diseases by
ML can be used for VS with a filtered dataset, employing algorithms identifying populations at high risk and tracking the movement of in-
such as SVM, RF, and DT [24,25]. After validating the trained model fected individuals. Due to its capacity for swift and precise data pro-
for accuracy, it is used on new data sets to screen for compounds that cessing of enormous amounts of data, AI can significantly improve our
have the desired activity against a target. The shortlisted compounds ability to identify and respond to infectious diseases.
undergo ADMET (absorption, distribution, metabolism, excretion and In recent years, the field of AI has shown remarkable progress in
toxicity) analysis and various bioassays before entering clinical trials. disease diagnosis, revolutionizing the way healthcare is delivered. AI
ML has the power to speed up VS, make it more robust, and even re- technologies such as ML and DL have enabled medical professionals
duce false positives in VS. PARASHIFT, HEX, USR, and Shape algorithms to develop highly accurate and reliable diagnostic models for various
have been constructed for LBVS. In recent years, with the rise of AI diseases. Early detection, precise diagnosis, and individualized treat-
algorithms in the healthcare and pharma industry, different tools and ment plans have all been made possible by the use of AI in disease
models have been developed for both LBVS and structure-based virtual diagnosis, which has improved patient outcomes. Infectious and non-
screening (SBVS), i.e., for LBVS, tools include SwissSimilarity, META- communicable disease diagnosis using AI has recently made significant
DOCK, and HybridSim-virtual screening, while for SBVS, tools include strides.

2
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

Fig. 2. Schematic representation of the integration of artificial intelligence (AI) in the drug discovery process. The diagram illustrates the workflow from
data collection, encompassing clinical sequencing, experimental text, and molecular structure analysis, to the implementation of AI algorithms and neural networks.
The applications of this process include disease diagnosis, target acquisition, and computational screening. This illustration underscores the transformative potential
of AI in advancing medical research and healthcare solutions.

AI methods have greatly advanced the diagnosis of infectious dis- During the COVID-19 pandemic, there was a particular focus on the
eases [26]. An exemplary case is the clinical decision support system development of AI models for the effective diagnosis of this disease
(CDSS) known as “Sepsis Watch” to detect Sepsis early [27], a severe [29,30]. Chest X-ray is one of the efficient indirect methods for COVID-
medical condition that occurs when an infection in the body triggers 19 diagnosis. It was used to diagnose pneumonia associated with this
a chain reaction, resulting in a life-threatening medical emergency, of- viral infection [31]. Many ML models have been developed to predict
ten originating from infections in the lungs, urinary tract, skin, or gas- the presence or absence of patterns in X-ray radiographs. For exam-
trointestinal tract [28]. Sepsis Watch utilizes a unique ML method that ple, Narin et al. proposed an automated convolutional neural network
combines multitask Gaussian processes (MGPs) with recurrent neural (CNN) based diagnostic model for detecting pneumonia caused by coro-
networks (RNNs) to identify sepsis [28]. The MGP component of the sys- navirus [9]. They developed pre-trained AI models using the X-ray ra-
tem learns the distributions of continuous functions for each dynamic diographs of healthy individuals, patients with COVID-19, patients with
variable. The system works by sampling dynamic features from the MGP viral pneumonia, and patients with bacterial pneumonia. The reported
hourly, which are then combined with static features and input into the accuracies in classification reached up to 96 %. Aapka Chikitsak is an
RNN, a form of DL particularly adept at processing time-series data and AI-powered conversational bot designed to enhance telehealth services
essential for incorporating both static and dynamic features of hospital in India by providing accessible COVID-19 information and addressing
patient encounters [28]. This process generates a sepsis risk score be- the imbalance between the demand for and the supply of human health-
tween 0 and 1 for each patient. Additionally, the system includes scripts care providers. Initially, the user’s query is converted from audio input
optimized to run every 5 min to identify patients who meet the sepsis cri- into text, and the text is used as a basis for performing Natural Language
teria, facilitating early detection and prompt medical intervention [28]. Understanding to decode the semantic meaning. Subsequently, relevant
Sepsis Watch has been trained using 50,000 patient records with both entities are identified and linked to their corresponding intents in Di-
static (eg. prehospital patient comorbidities) and dynamic (eg. medica- alogflow. The bot then generates a response, converts it to speech, and
tion administrations) input features. It has been shown to improve the delivers it back to the user [32].
care of sepsis patients [28]. However, this study had a few limitations, ML models have also been built to assist in the diagnosis of other in-
including some false positive predictions that prompted clinical action fectious diseases such as urinary tract infections (UTIs), which are often
even though the patient did not ultimately develop sepsis. Additionally, associated with high rates of diagnostic errors. Taylor et al. reported a
the study was limited to emergency department cases. retrospective cohort analysis of approximately 80,387 adults who vis-

3
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

ited the emergency department with UTI symptoms [33]. Considering et al. developed an AI-based model for the early prediction of AD using
symptoms as well as blood and urine sample analyses, six AI algorithms positron emission tomography (PET) images [43]. The model achieved
were developed for the diagnosis of UTI: SVM, artificial neural network an accuracy of 89.5 % in predicting the onset of AD within two years.
(ANN), elastic net, adaptive boosting (AdaBoost), RF, and extreme gra- Cancer is another NCD that has benefited from the recent advancements
dient boosting (XGBoost). The models were built using a full set of 211 in AI-based algorithms. AI-based algorithms have been developed for
factors and a reduced set of 10 variables, e.g., gender, epithelial cells various cancer-related tasks, such as diagnosis, prognosis, and treatment
in the urine, history of UTI, and age. The XGBoost showed superior ac- planning. For example, Ding et al. developed a DL algorithm that in-
curacy over the other algorithms, with an area under the curve (AUC) terprets PET of the brain for the early prediction of AD. Their model
of 0.88 and 0.90 for the full and reduced XGBoost models, respectively. showed a specificity and sensitivity of 82 % and 100 %, respectively.
The sensitivity and specificity were 61.70 % and 94.90 % for the full, This model can predict AD, on average, 75.8 months before its diagnosis,
and 54.70 % and 94.70 % for the reduced models, respectively [33]. with a ROC (receiver operating characteristic) of 0.98 [8]. Li et al. devel-
Lifestyle disorders, such as diabetes, obesity, and hypertension, are as- oped a DL-based model that automatically evaluates dementia severity
sociated with the way people live, their diet, exercise levels, etc. by analyzing resting-state functional magnetic resonance imaging data
ML models have achieved substantial success in the prevention and (rs-fMRI). The study involved 133 patients with Alzheimer’s disease,
detection of HIV, a virus causing human progressive immune system and their clinical dementia rating (CDR) scores ranged from 0.5 to 3. To
failure and promoting cancers. Xiang et al. developed an ensemble ap- extract features, three-dimensional CNNs were applied to rs-fMRI data.
proach combining GCN (graph convolutional network) with LR and RF, The accuracy of the model was found to be highly satisfactory [44]. In
aiming to identify individuals at high risk of HIV infection for preven- addition, AI-based algorithms have also shown promising results in pre-
tion [34]. They built a social network in which each node stood for dicting the response to treatment in cancer patients. For example, AI
an individual participant, and the edges between them represented so- has the potential to improve the speed of analysis and the accuracy of
cial connections. Each node was assigned with a feature set, including image interpretations. Esteva and co-workers developed a CNN model
sociodemographic characteristics and sexual behavioral characteristics. trained with images of skin lesions to classify different types of skin
The ensemble methods produced promising results on HIV infection de- cancer [45]. Albayrak et al. used deep learning to extract features from
tection (GCN+LR with accuracy of 93.4 % and F1 of 88.4 %; GCN+RF breast histopathological images to detect mitosis. The proposed model
with accuracy of 96.6 % and F1 of 94.6 %). Heerden et al. created a con- extracted CNN features for SVM training and detected breast mitosis
versational agent guiding users through an HIV counseling and testing [46]. Causey et al. developed CNN model-based algorithm, NoduleX to
session utilizing NLP (nature language process). This agent encourages predict malignant lung nodules from clinical CT data. The model was
the targeted population to openly discuss their concerns with a virtual trained using over 1000 lung nodule images from LIDC and IDRI. Nod-
assistant, and its effectiveness has been confirmed through human eval- uleX predicted with a 0.99 AUC [47]. Shiri et al. tested ML methods
uation [35]. Compared with previous agents, this agent made testers using Radiomics analysis for predicting EGFR and KRAS mutation sta-
feel like the session was more private and anonymous with more gentle tus in NSCLC (non-small cell lung cancer) patients that showed AUCs of
language and more accurate detection. 0.82 and 0.83, respectively [48]. CNN models also power histological
Many AI-based algorithms have been developed for the early predic- image analysis to diagnose cancer [49-51].
tion and management of diabetes. For instance, Spänig et al. developed AI algorithms have demonstrated promising outcomes in the diag-
an interactive AI model with the capability of speech recognition and nosis, prediction, and management of non-communicable diseases in-
speech synthesis that acts as a virtual doctor, interacts directly with pa- cluding diabetes, Alzheimer’s disease, and cancer. The implementation
tients. An open-source system CMUSphinx is utilized to develop robust of AI-based algorithms can facilitate the early detection of these dis-
speech recognition capabilities. To support localized speech recognition, eases, thus enabling timely interventions and personalized treatment
the essential German language data is obtained, including a German lan- plans, ultimately leading to improved patient outcomes. By leveraging
guage model, an acoustic model, and a dictionary from the VoxForge the power of AI, healthcare professionals can potentially reduce the bur-
dataset [36], which aggregates transcribed speech specifically for use den of NCDs, enhance the quality of healthcare delivery, and optimize
in speech recognition technologies. This virtual doctor predicts Type-2 healthcare resources. The detailed overview of role of AI in disease iden-
diabetes mellitus with an AUC of 0.84 [37]. The incidence of retinopa- tification can be seen in Fig. 3.
thy is high among diabetic patients. Gulshan et al. developed a deep
CNN model that bypasses the human capacity at interpreting, evaluat- 3. Target identification
ing, and classifying retinal images. The model is trained using 128,175
retinal photographs which are evaluated by a panel of clinicians and Target identification is a critical step in the drug discovery process.
ophthalmologists. The model is demonstrated to have a high sensitiv- The traditional approach involves time-consuming and costly experi-
ity and specificity of 97.50 % and 93.40 %, respectively [38]. In 2018, mental methods, such as high-throughput screening (HTS) and X-ray
the U.S. Food and Drug Administration approved the marketing of the crystallography. However, the use of AI has revolutionized this field by
first AI-based medical device called IDx-DR [39] for detecting diabetic enabling the identification of potential targets through computational
retinopathy. The device has a retinal camera through which the retinal methods. Fig. 4 illustrates how AI-aided role in drug discovery.
image of the patient is taken and analyzed. The device is autonomous AI-based target identification involves the use of ML algorithms to
and decides on one of the following results based on the image quality analyze large datasets and identify targets with the potential to interact
(i) “more than mild diabetic retinopathy detected: refer to an eye care with a given drug. This approach utilizes various data sources, such as
professional” or (ii) “negative for more than mild diabetic retinopathy; gene expression profiles, protein-protein interaction networks, and bio-
rescreen in 12 months” [39]. logical pathways, to generate a list of candidate targets [52]. ML algo-
Alzheimer’s disease is a neurodegenerative disorder in the brain. Ge- rithms, such as SVMs and neural networks, can then be used to prioritize
netic factors and age are major risk factors associated with AD. However, these targets based on their relevance to the disease of interest.
recent research indicates that other factors, such as environmental and Furthermore, AI-based target identification can help identify novel
lifestyle factors, can also contribute to the development of AD [40]. AI- targets that were previously unknown or overlooked. By analyzing large
based algorithms have shown promising results in the early prediction datasets from various sources, ML algorithms can uncover hidden pat-
and diagnosis of AD [41]. For instance, Jo et al. proposed a hybrid model terns and relationships that may not be immediately apparent using
that combines DL -based feature extraction with ML algorithms for AD traditional methods. This can lead to the discovery of new biological
diagnosis using magnetic resonance imaging (MRI) scans [42]. The pro- pathways and targets that may have therapeutic potential. AI-based tar-
posed model achieved a classification accuracy of 96 %. Similarly, Shi get identification has the potential to revolutionize the drug discovery

4
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

standing of the structures of protein complexes in their environment,

cryo-electron tomography (cryo-ET) has shown great promise. Com-
puter algorithms are used to combine these images to create a 3D struc-
ture representation.
The software CryoDRGN, created by Zhong et al. [56], uses ML to
enable the reconstruction of proteins and protein complexes from het-
erogeneous cryo-electron microscopy data [56]. In order to embed het-
erogeneous single-particle cryo-EM images in a low-dimensional latent
space and produce 3D volumes as a function of this embedding, the
authors have proposed a technique that makes use of ML models. Cryo-
DRGN can produce an infinite number of maps from the imaged ensem-
ble and is capable of modeling complex ensembles with both continuous
and discrete heterogeneity. The software can also visualize the motion
of the protein [57].
Trypsin is one of the most important and widely used proteolytic
enzymes in mass spectrometry (MS)-based proteomic research. The di-
gestion of proteins by protease enzyme is a basic step in the protein
identification using MS. A few AI tools were developed to efficiently
predict the digestion behavior of the protease enzymes [58,59]. Deep-
Digest is the first algorithm developed using a DL method to predict the
proteolytic cleavage sites of eight different protease enzymes (trypsin,
ArgC, chymotrypsin, GluC, LysC, AspN, LysN, and LysargiNase). The DL
Fig. 3. The overview of AI-based disease identification. AI-based disease
model was trained on 19 public large-scale data sets covering the eight
identification can be divided into four categories: neurodegenerative disorders
proteases from samples of four organisms (E. coli, yeast, mouse, and
diagnosis, cancer diagnosis, infectious disease diagnosis and lifestyle disorders
diagnosis. human). The predictive ability of the tool was evaluated by the AUC,
F1 scores, and the Matthews correlation coefficients (MCCs); the val-
ues were 0.956–0.98, 0.66–0.90, and 0.65–0.84, respectively [58]. Sun
process by enabling the identification of potential targets through com- et al. developed an algorithm to predict the missed cleavage site in the
putational methods. There are multiple experimental methods available tryptic protein digestion [59]. The algorithm is demonstrated to have
for identifying drug targets, such as affinity pull-downs and genome- a high accuracy, with an AUC of 0.99. This algorithm can be incorpo-
wide knockdown screens. However, these approaches require a signifi- rated into the peptide database search in the MS analysis to facilitate
cant amount of labor, resources, and time, and are also subject to a high the identification of proteins more effectively and efficiently.
rate of failure. In contrast, computational methods have the potential to The development of computational tools, high-performance comput-
significantly decrease the effort and resources required for drug target ers, and ML algorithms is not limited to three dimensional (3D) models
identification [53]. of protein targets but also enabled a large number of drug discovery
Cryo-EM, or electron cryo-microscopy, revolutionized the investi- tools. This is a significant advancement in experimental techniques that
gation of proteins and protein complexes in the 1980s. Single-particle are fraught with challenges. For example, the X-ray diffraction tech-
analysis (SPA), electron cryo-tomography (cryo-ET), microcrystal elec- nique is limited to crystallizable samples, which is a major experimental
tron diffraction (MicroED), low-energy electron holography (LEEH), and limitation [60].
cryo-scanning transmission electron tomography (CSTET) are a few ex- In the absence of experimental data, computational techniques have
amples of cryo-EM imaging techniques [54,55]. For a deeper under- been used for decades to forecast the three-dimensional structure of

Fig. 4. The overview of the AI-based process for identifying and evaluating small molecules. Beginning with target acquisition, where protein sequences are
modeled using tools like Modeller and AlphaFold2, followed by Homology Modeling and validation of protein 3D structure through ProQ and SolvX. Next is the
pocket exploration, including the identification of both orthosteric and allosteric binding sites within the protein proteome, via static and dynamic structures. The
hit identification stage employs SBVS and LBVS to pinpoint potential compounds. Finally, the toxicity prediction stage uses pattern recognition techniques in AI/ML
to predict the ADMET properties.

5
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

proteins. Modeller, a homology modeling software, predicts a protein’s models, such as XGBoost [71], can now classify allosteric sites with
structure based on its alignment to one or more proteins of known greater accuracy. XGBoost implements the gradient boosting algorithm
structure (templates) [61]. AlphaFold, the DL algorithm developed by with regularized terms to reduce overfitting and has demonstrated su-
DeepMind, a UK based company, predict the 3D structure of proteins perior predictive performance in protein-protein interactions [72] and
from their amino acid sequences [62]. The development of this AI- tool hot spots [73] compared to SVM and RF.
is claimed to be a breakthrough in drug discovery as it was used to Tian et al. developed a webserver called PASSer (Prediction of Al-
solve the structure of nearly 200 million proteins, which is ∼98.5 % losteric Sites Server), combining the results of XGBoost and GCN to pre-
of the proteins in the human body. In July 2021, the predicted three- dict the allosteric sites on proteins using an ensemble learning method.
dimensional models for the whole human proteome generated using Al- A total of 1946 entries information of allosteric sites from Allosteric
phaFold, were made available to the public, as recently reported in Na- Database were collected for training with 19 descriptors extracted for
ture [63]. Together with the European Bioinformatics Institute (EMBL- each site by Fpocket. For a given pocket, physical properties are calcu-
EBI), a database called AlphaFold DB (https://alphafold.ebi.ac.uk/) was lated and fed into the XGBoost model while an atomic graph is fed into
created to store all the structures solved so far with AlphaFold. How- the GCNN model. The final result is the averaged probability of these
ever, the effect of mutation on the folding of proteins is beyond the two models. The model showed an accuracy of 0.97, precision of 0.73,
capability of AlphaFold [64]. It is also limited to predicting only a sin- and specificity of 0.98. The aforementioned model can acquire knowl-
gle state of a given protein, it does not consider the dynamic nature of edge about both the physical traits and topology of allosteric pockets and
protein structures [65]. Recently, new methods based on natural lan- has been demonstrated to outperform the XGBoost and GCN models in-
guage processing (NLP) which only uses the amino acid sequences from dividually. The findings are consistent with earlier research and have a
sequence databases to learn structural, functional and evolutionary pat- greater likelihood of placing allosteric sites in top positions. The online
terns and predict structural conformation. In 2022, two methods gained server is equipped with an easily navigable interface. Protein structures
attention i.e. ESMFold and EMBER3D. ESMfold, developed by Lin et al. and top pockets are displayed in an interactive window on the result
in 2022, utilizes a masked transformer protein language model with a page [66].
deep understanding of biological properties, trained with 15 billion pa- Protein structural fluctuations generate novel cryptic pockets
rameters. While it falls short of AlphaFold2 in overall performance, as [74-77], which offer druggable sites beyond experimentally determined
indicated by lower TM-scores (0.68 compared to AlphaFold2′s 0.85 on structures. Cryptic pockets regulate protein functions allosterically.
CASP14), ESMFold outperforms AlphaFold2 when evaluating the amino These pockets are hidden protein cavities that open up when ligands or
acid sequence alone without multiple sequence alignment (MSA) (0.68 protein partners bind [78]. Targeting these cryptic pockets offers drug
versus 0.37 on CASP14). ESMFold demonstrates comparable accuracy development opportunities. Cryptic pockets can target undruggable pro-
to AlphaFold2 for structures predicted with high confidence, exhibiting teins [79]. An algorithm to predict which proteins have cryptic pock-
a median all-atom RMSD (root-mean-square deviation) of 1.91Å and ets could help prioritize proteins to target in cases where proteins lack
a backbone RMSD of 1.33Å—achieving accuracy levels akin to experi- ground state pockets or modulators are difficult to design. CryptoSite is
mental results. Additionally, ESMFold shows a substantial improvement an excellent supervised ML algorithm that predicts ligand-binding cryp-
in prediction speed, eliminating the need for MSA construction. The tic pockets from protein structures [80]. CryptoSite accurately predicts
authors leverage this approach to introduce the ESM Metagenomic At- cryptic pocket participation for amino acid residues (ROC-AUC = 0.83).
las, predicting over 617 million structures from metagenomic databases. Miller et al. developed PocketMiner, a graph neural network (GNN), to
Among these, 225 million structures are predicted with high confidence, predict cryptic pockets in protein structures. The model is trained us-
including novel ones [63]. While Ember3D falls short of surpassing Al- ing residues likely to form cryptic pockets from 2400 simulations of 35
phaFold in performance, it exhibits significantly faster processing speeds proteins. Model AUC was 0.87. This supports molecular dynamics sim-
compared to both AlphaFold and ESMFold. Notably, AlphaFold2 strug- ulations for cryptic pocket identification [78].
gles to efficiently investigate the impact of single amino acid variants on Madhukar et al. developed BANDIT, a Bayesian ML platform for
protein structure. In contrast, Weissenow et al. [64] demonstrated that drug target prediction that integrates multiple data types to achieve
Ember3D’s predicted distance maps show a strong correlation with na- greater predictive power. BANDIT utilizes over 20 million data points
tive and mutant 3D structures obtained through deep mutational scan- from six distinct data types, including drug efficacies, post-treatment re-
ning, outperforming ESMFold in this regard. The researchers also de- sponses, bioassay results [81,82], and known targets [83,84]. The plat-
veloped a tool that highlights differences between native and mutant form achieves high accuracy at identifying shared target interactions
predicted structures for all possible amino acid exchanges at each po- and uncovers novel targets for cancer treatment. BANDIT was tested on
sition in a protein sequence. The tool utilizes the similarity between ∼2000 compounds and can quickly pinpoint potential therapeutics with
the native and mutated amino acids to identify exchanges that may novel mechanisms of action to accelerate drug development.
have a significant impact on the protein structure [64]. Cheng et al. In 2021, Kozlovskii and Popov developed a DL approach to pre-
introduced AlphaMissense, a DL model that extends the capabilities of dict the binding site for small molecules on nucleic acids, DNA, and
the AlphaFold2 protein structure prediction tool. This model undergoes RNA, based on their 3D structures [85]. Their approach is called
training with population frequency data and incorporates both sequence BiteNetN (https://sites.skoltech.ru/imolecule/tools/bitenet/) having a
information and predicted structural context to enhance its overall per- large dataset of 1933 nucleic acid-ligand complexes, including 1065
formance. AlphaMissense effectively categorizes 32 % of all missense DNA and 886 RNA structures (18 structures contain both DNA and
variants as likely pathogenic and identifies 57 % as likely benign, achiev- RNA) of different type. It was the first 3D CNN to learn features directly
ing a 90 % precision cutoff on the ClinVar dataset. This outcome ensures from nucleic acid structures. They validated the model using two differ-
a robust prediction for the majority of human missense variants [65]. ent protein structures, HIV-1 transactivation response element RNA and
A specific and detailed review [66] on all the recent advances in pro- ATP-aptamer structures. The model showed an AUCROC of ca. 0.87,
tein structure prediction are reviewed by Peng et al. Some breakthrough proved to be top-performing for protein-small molecule and protein-
protein structure prediction models are tabulated in Table 1. peptide binding site detection [86,87].
In addition, AI algorithms can predict the likelihood of allosteric reg- DeepDTnet is a network-based DL method developed by Zeng et al.
ulation in a protein based on its structure and ligand presence, and iden- to aid the target identification process [88]. The model is trained with
tify allosteric modulators that may modify its activity [66]. ML methods, chemical, genomic, and cellular network data for the accurate predic-
like Allosite; developed by our group [67], and AlloPred [68], used SVM tion of molecular targets. The model is shown to have a high accuracy
with optimized features for protein pocket classification [69], while oth- in prediction with an AUC of 0.96 [88]. In another study, Mamoshina
ers used RF [70] to build a three-way predictive model. Advanced ML et al. developed ML techniques to analyze human muscle transcriptomic

6
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

Table 1
List of the methods for protein structure prediction with their standlone availability links.

S. No Method Web links

1 AlphaFold2 https://github.com/deepmind/alphafold
2 ColabFold https://github.com/sokrypton/ColabFold
3 DMPFold2 https://github.com/psipred/DMPfold2
4 ESMFold https://github.com/facebookresearch/esm
5 EMBER3D https://github.com/kWeissenow/EMBER3D
6 HelixFold-Single https://paddlehelix.baidu.com/app/drug/protein-single/forecast
7 Openfold https://github.com/aqlaboratory/openfold
8 OmegaFold https://github.com/HeliXonProtein/OmegaFold
9 RoseTTAFold https://github.com/RosettaCommons/RoseTTAFold
10 RNG https://github.com/aqlaboratory/rgn
11 RNG2 https://github.com/aqlaboratory/rgn2

data to discover biomarkers associated with muscle-related diseases and The conventional drug discovery process involves synthesizing and
to identify tissue-specific drug targets [89]. The authors collected tran- testing thousands of compounds, which is both time-consuming and
scriptomic data from human muscle tissues and used a combination of costly, requiring large amounts of protein supplies and established lab-
unsupervised and supervised ML algorithms to identify differentially oratory methods for bioactivity testing. In contrast, VS has emerged as a
expressed genes and gene modules that are associated with muscular cost-effective approach to scan millions of commercially available com-
dystrophy and sarcopenia. They further investigated the identified gene pounds and prioritize those for further testing, synthesis, or purchase. VS
modules and pathways using functional annotation and network anal- methods are classified into two categories: structure-based and ligand-
ysis tools to identify potential drug targets. Their best model showed a based methods. However, despite the potential benefits, it still takes an
Pearson correlation of 0.80. average of 10 to 15 years and over $2 billion to develop a single drug
After identifying the protein targets, corresponding novel lead com- [94].
pounds can be further generated from scratch, leveraging the pocket
features and ligand topology. Zhang et al. developed a pocket-aware 4.1. Structure-based methods
ligand generator ResGen. A two-level autoregression protocol for molec-
ular generation is introduced in the model to better incorporate the To utilize structure-based methods, 3D structural information of the
geometry of protein pockets. The global autoregression is to generate protein-ligand complex or at least the protein’s binding site is needed.
atoms in pockets, and the atomic autoregression is to produce the co- Molecular docking is a commonly used technique that generates multi-
ordinates and topology of the newly added atoms. ResGen could gener- ple possible binding poses of a ligand in the target protein structure and
ate more physically sensible molecules with tighter binding [90]. Wei ranks them using a scoring function (SF) [95] to estimate their binding
et al. developed a fragment-based generation model Lingo3DMol built affinity [96]. Recently, machine/ DL-based SFs [21] have been intro-
on the transformer-based structure [91]. A new molecule representation duced as a new group of SFs. On the other hand, ligand-based methods
FSMILES is introduced, enabling the generation of 3D molecules with such as QSAR modeling, molecular similarity search, and ligand-based
reasonable conformations and topology. Additionally, non-covalent in- pharmacophores, are more established technologies that require only
teractions and ligand–protein binding patterns are also considered dur- ligand information [97], unlike structure-based methods. AI techniques
ing the generation. Lingo3DMol demonstrates excellent performance in can also be employed to improve the efficiency of computer-aided drug
terms of drug likeness, synthetic accessibility, pocket binding mode and discovery processes, which often need extensive high-performance com-
molecule generation speed. puting resources and significant computation time.
Gentile et al. reported an open-source protocol for AI-enabled
4. AI-enabled virtual screening in drug discovery: opportunities VS methods to screen libraries with billions of molecules. They
and challenges used a screening platform called Deep Docking (https://github.com/
jamesgleave/DD_protocol) which can accelerate the SBVS by 100 folds.
The initial phase of drug discovery usually involves computational The input data consists of the molecule’s SMILES with its Morgan finger-
screening of numerous compounds to identify those with the desired prints as descriptors and the target’s structure. Deep Docking performs
cellular or biochemical effects. To enhance the speed, efficiency, and molecular docking for a small subset of a large library based on DNN to
cost-effectiveness of this process, new methods are constantly being de- infer the ranking of the yet-unprocessed remainder, followed by ligand-
veloped. A positive response during the first round of screening in a based prediction of the docking for the rest of the library. In this way,
biochemical assay identifies primary “hit” compounds. Subsequently, Deep Docking discards undockable molecular structures without wast-
additional screening is performed to assess whether the physicochem- ing computational resources. A key advantage of this protocol is that it
ical and pharmacological properties of the hit compounds are suitable can be used in conjunction with other docking programs such as Glide,
for developing a medicine. If they pass this filter, they are designated Autodock-GPU, and FRED from OpenEye. Although, the deep docking
as “leads”. These leads are then refined chemically and subjected to bi- method provides faster screening, it is limited to I) the availability of
ological screening in subsequent rounds before proceeding to clinical graphical processing units (GPU) and ii) the quality and accuracy of the
testing. With some luck, a lead may ultimately receive drug approval, docking program used [10].
a process that may take 12–15 years from the beginning of testing Various ML techniques, including Naïve Bayesian (NB) classifiers,
[92]. kNNs, SVMs, RFs, and ANNs, can be used for VS. While SVMs and ANNs
Despite significant advancements in drug discovery and medicinal are commonly regarded as the most accurate, each technique has its own
chemistry technologies, drug development still remains a slow and ex- strengths and weaknesses. For instance, NB excels at identifying favor-
pensive process. The current standard process involves HTS, where in able scaffold fragments, RFs can be parallelized and boosted, and kNN
vitro assays are conducted on thousands of compounds to identify hit is simple to implement and can utilize MTL. Combining an ensemble of
compounds that can be optimized to lead compounds with desirable ML models is often preferred as it can enhance performance [98].
properties such as increased potency, solubility, and reduced toxicity For LBVS, NB classifiers are widely utilized and exhibit excellent per-
and off-target effects [93]. formance. Wang et al. investigated four classification models to identify

7
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

inhibitors of methicillin-resistant Staphylococcus aureus. These models cient of determination, a statistical measure within regression analysis)
include NB, SVM, recursive partitioning (RP), and kNN algorithms, in- values of 0.85 and 0.78 for AChE and BACE1, respectively. Dhamodha-
corporating various sets of physicochemical descriptors and fingerprints. ran et al., [113] used a target-specific scoring model to identify potential
Among these models, the NB classifier displayed the highest perfor- inhibitors for 12 targets from the SAM MTase family. 446 actives and
mance in the testing set of, achieving an accuracy of 0.891 and a speci- 1294 decoys were docked using Glide and the DUD-E website. The MLP
ficity of 0.920 [99]. Likewise, Lian et al. [100] showed that NB models outperformed other docking tools in a binary classification experiment.
outperformed SVM models with an accuracy of 0.975 and a specificity of Atomwise, Inc. developed AtomNet, one of the earliest CNNs uti-
0.989 in classifying the inhibitors and non-inhibitors of neuraminidase. lized for VS [113]. Unlike most ML-based VS, AtomNet employs an SBVS
They further identified nine effective inhibitors of neuraminidase using and its architecture includes input and logistic-cost layers, four convo-
an enhanced ensemble model comprising NB models and SVMs. lutional layers, and two fully connected layers. The filters in AtomNet’s
Another simple ML classification method is kNN. Unlike NB, kNN is convolutional layers correspond to chemical functions, allowing the
intuitive. In searching for G-Protein Coupled Receptor ligands, Luo et al. network to identify features that aid in binding. AtomNet consistently
found that kNN-QSAR with variable selection outperformed LBVS ap- achieves AUC scores >0.74 on various benchmark datasets, outperform-
proaches without ML [101]. Compared to other ML methods, kNN’s per- ing numerous previous docking models. These features and capabilities,
formance is usually in the middle [99,102]. kNN is less popular than the previously exclusive to NB and RF classifiers, now enable ANNs to iden-
ML methods below because of this. In screening for Estrogen Receptor- tify features that aid in binding, making them more accurate and less of
mediated endocrine disruptors, kNN has been used for VS [103]. a black box in VS.
Vapnik et al., introduced SVMs for the first time [104]. They func- Alzheimer’s disease is characterized by amyloid-beta (A) plaques
tion by representing input data as feature vectors and plotting them and neurofibrillary tangles (NFT) of hyperphosphorylated tau protein.
within the same-dimensional space. Although SVM can be used for un- Acetylcholine (ACh) levels are lower in AD brains [114-116]. Tau may
supervised learning, supervised learning is preferred for VS because it be phosphorylated by GSK-3, CDK5, and other Alzheimer’s disease-
guarantees that a compound will eventually be categorized as active or related enzymes and targets [117-119]. Side effects from polypharmacy
inactive. Chandra and colleagues conducted a study to identify poten- are possible. Drugs that target one protein frequently have side effects
tial inhibitors for PTP1B, a treatment for Type 2 diabetes [105]. Multiple on other proteins as well. Through the inhibition of multiple targets
ML models were developed, and the best model that utilized SVM was in complex diseases like Alzheimer’s, polypharmacology can increase a
applied to an external database. This model successfully identified five drug’s effectiveness. With this technique, MTDL (Multi-Target Directed
inhibitory compounds, two of which demonstrated significant activity Ligands) screening was increased [43]. Using NB and RP classifiers in
in vitro. In a separate study, Deshmukh et al., [106] discovered that an LBVS on MTDLs, Fang et al. discovered compounds that bound to 25
their SVM model was able to identify almost half of the known FEN1 in- targets, including BACE1, the M1 subtype of mAChR, APP, CDK5, and
hibitors in a test set, and also identified previously unknown inhibitors GSK-3 [120]. The model was validated using current AD medications,
from the Maybridge small molecule database, which were experimen- and it was used to forecast upcoming MTDLs. As scientists become more
tally validated. Baba et al. found that SVM models with regression were aware of its effectiveness, ML in VS for drug discovery will expand. The
more effective than RF in predicting a compound’s ability to permeate effectiveness and cost of drug discovery will be enhanced by computer
the skin[107]. Lee et al., [108] employed RF-QSAR to study compound science and medicinal chemistry.
polypharmacology, resulting in the creation of a targetfishing server that ML-based screens are computationally efficient and successful in
identifies possible targets for a given compound. Their method achieved modern CADD, which requires vast computational resources to screen
an overall AUC score of 0.97 and outperformed NB-based methods in ex- expanding chemical libraries due to automated synthesis and robotics.
ternal testing [109]. OpenEye GigaDocking and VirtualFlow are supercomputing platforms
for docking large libraries that have screened billions of molecules us-
4.2. Ligand-based virtual screening (LBVS) ing thousands of CPUs/GPUs in relatively short periods [121]. However,
they are resource-intensive compared to Deep Docking (DD) [122]. DD
LBVS is based on selecting, from databases, molecules that share sim- requires fewer computational resources, making it an attractive alterna-
ilar structural features with an active ligand. Pharmacophore-based VS tive for large-scale VS.
is one of the LBVS techniques. It involves building 2D fingerprints of one Gentile et al. developed an open-source protocol for AI-enabled
or more active ligands using molecular descriptors such as hydrogen- VS of billion-molecule libraries. Deep Docking (https://github.com/
bond donors, hydrogen-bond acceptors, and aromatic rings. These 2D jamesgleave/DD_protocol), a screening platform that accelerates SBVS
fingerprints are then used to identify molecules, from large chemical li- by 100-fold, was used. One of the fastest AI-enabled docking platforms
braries, which have matching pharmacophoric features. ML also helps and the only one tested on 1B+ libraries is DD. The DD protocol does not
to study the correlation between molecular descriptors (or even atomic require a docking program, so it can be used with emerging large-scale
descriptors [110,111] and the biological activity of a ligand. This is a docking methods to improve throughput. technical limitation of Deep
broad category of research known as Quantitative Structure-Activity Re- Docking is GPU acceleration, which is needed for optimal performance
lationship (QSAR), where the activity of a ligand depends on its phar- unlike CPU-based docking platforms. The protocol only provides dock-
macophoric features. Melge et al. developed hybrid inhibitors using ing details for top-scoring molecules, ignoring large fractions of chem-
the pharmacophore fingerprint of two well-known anti-cancer drugs ical libraries for fast VS. Docking campaigns assessing hit rate variabil-
Ponatinib and Vorinostat [112]. They developed a supervised ML ap- ity with docking scores [123] or rescoring low-ranked molecules [124]
proach for 2D-QSAR and 3D-pharmacophore studies to predict the in- should consider a bruteforce approach. A docking program’s ability to
hibitory activity of novel hybrid molecules. The model had AUC values prioritize active molecules from an ultra-large library also determines
of 0.98 and 0.94 for the two different cancer targets, BCR-ABL and Hi- the quality of DD results [123]. Bender et al.’s guidelines for large-scale
stone deacetylase (HDAC), respectively. Based on in-vitro evaluations, docking benchmarking are helpful [122].
the identified novel hybrid molecules showed the potential to develop A wave of DL methods and applications has improved affin-
into lead compounds. Dhamodharan et al. developed three AI models ity—and other properties like ADMETox—prediction in VS over the
based on genetic function approximation (GFA), SVM, and ANN, to pre- last decade(s). Models learn characteristics rather than using human-
dict the activity of acetylcholinesterase (AChE) and Beta-Secretase 1 designed descriptors. Novel encodings like voxels (where physicochem-
(BACE1) dual inhibitors for AD treatment [113]. The predictive power ical atomic properties are pinned to locations in 3D space) and graphs
of the models was evaluated on a test set of 11 inhibitors of AChE and (which describe bonded and non-bonded connectivity between atoms)
BACE1. The ANN model had the best predictive power with R2 (coeffi- appear to capture the variety of information needed for ligand-binding.

8
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

DeepAtom [125], a 3D grid-based method that assigns physicochemi- Hsiao et al. [139] and Arab et al. [140] also used the RF model with
cal properties to each grid cell, may be suitable for modeling protein- high accuracy. Ogura et al. developed an SVM model using ECFP_4 struc-
ligand binding complexity. The study [126] and DGraphDTA model tural fingerprints and 72 NSGA-II-selected descriptors, outperforming
[127] show that encoding chemical and biological objects as graphs other predictors with 98.4 % accuracy and 0.733 kappa statistic [141].
works well. Despite this, several challenges remain, and new ones Konda et al. [142] generated hERG classification models with 2D de-
have emerged, including chemical encoding precision, generalization scriptors using RF, SMO (sequential minimal optimization), and MLP
of chemical space, lack of (large and high-quality) data, model compa- (multilayer perceptron) algorithms, with their consensus model outper-
rability, and interpretability. forming other predictors with 92 % accuracy. DNN, ANN, RNN, CNN,
Finding new therapeutic uses for already-approved drugs is a cru- GNN, GCNN, and GAT are used in developing hERG predictive mod-
cial aspect of drug repurposing. Due to their previously tested phar- els. Shan et al., [143] generated a directed message-passing neural net-
macokinetics and toxicity, repurposed medications have a higher suc- work (DMPNN) model with moe206 descriptors that outperformed other
cess rate [128,129]. Drug repurposing skips lead optimization and pre- models with an accuracy of 80 % [143]. Zhang et al., [144] developed
clinical studies to enter phase-II trials. Due to complex disease patho- HergSPred, which outperformed other models, achieving an accuracy
physiology, many drugs have off-target binding and are excluded from of 98.3 %. Ryu et al., [145] developed DeepHit, which outperformed
pre-clinical trials [130]. Reker and colleagues developed a method to other tools in terms of accuracy, MCC, and SE (sensitivity). Wei et al.
predict molecular targets of known drugs and computer-generated de created Interpretable-ADMET, achieving the highest accuracy (91.9 %)
novo small molecules, including key-target and off-target proteins. Self- and ROC-AUC (78.2 %) on 8672 compounds. ADMETLab 2.0 achieved
organizing map-based prediction of drug equivalence relationships (SPi- the highest accuracy and ROC-AUC of 88.9 % and 94.3 %, respectively,
DER) is this method. The software was trained on 12,661 manually cu- on a large hERG data set containing 13,845 compounds, utilizing a mul-
rated active molecules [131]. The ROC ranged from 0.86 to 0.93 in a titask graph attention framework (MGAF) to predict ADMET properties.
10-fold cross-validation of SPiDER’s predictive ability.
5.2. LD50 prediction
5. Prediction of drug toxicity with AI
In Toxicologists use the LD50 (median lethal dose) to determine a
Proportion of drug candidates being discarded during clinical trials substance’s toxicity. The dose needed to kill 50 % of test animals in a
due to unexpected adverse effects. Predicting drug toxicity during pre- particular period of time is referred to as the LD50 value of a chemical.
clinical stages is a crucial step to reduce the failure rate and improve The LD [146] is used as the first step in the drug screening process. Rats’
the efficiency of drug discovery. Traditional methods of predicting drug acute oral toxicity was evaluated using LD50 . Interspecies variability and
toxicity are limited by their reliance on small datasets and simplistic ethical issues render conventional LD50 testing obsolete [147]. Acute
models. However, AI-based approaches have emerged as promising al- animal toxicity tests using tissue culture and in silico LD50 prediction are
ternatives, leveraging large and diverse data sources, including chemical gradually being replaced [148]. The binary classification model divides
structures, biological pathways, and clinical data. By utilizing ML algo- substances into two categories: toxic (LD50 = 2000 mg/kg) and nontoxic
rithms, AI-based approaches can improve the accuracy and efficiency (LD50 = 50 mg/kg).
of predicting the potentially toxic effects of new compounds, helping to Compounds are divided into multiple classes by the Globally Har-
mitigate risks associated with clinical trials, reduce drug development monized System of Classification and Labeling of Chemicals and the US
costs, and ultimately lead to better patient outcomes. Environmental Protection Agency [149]. Recent releases of the ADMET
Recently, the use of AI-based computational models to forecast drug prediction programs FP-ADMET [138] and Interpretable-ADMET [150],
toxicity has grown in popularity [132]. Large drug and toxicity data sets in these both applications, LD50 prediction models are applied.
have been analyzed in a number of studies using ML and DL algorithms, Ballabio et al. developed LD50 prediction models using the binary fin-
such as neural networks, to identify potential toxic effects during drug gerprints of NB, N–Nearest Neighbors, Binned-Neighbors, and Extended
development. By identifying toxicities early on, these models can speed Connectivity [151]. They found that their models were 84 % sensitive
up the development of new drugs. Additionally, AI-based toxicity pre- and 81 % specific for 8992 chemicals. An integrated QSAR model for
diction models can prioritize compounds for testing and find new drug 8448 compounds was created by Gadaleta et al., [152], by combining
targets and toxicity mechanisms. AI toxicity prediction has been the sub- balanced RF, regression, aiQSAR, istkNN, SARpy, RF with hyperparam-
ject of several reviews [132-137]. A single review is challenging due to eter tuning, and a general linear model [153]. The ideal model’s RMSE
the wide field of AI-based toxicity prediction’s numerous toxicity prop- (root-mean-squared error) was 0.477 %, and its accuracy balance was
erties. Recent AI-based toxicity prediction models and in-depth studies above 70 % for multiclass endpoints and 80 % for binary endpoints.
of toxicity properties are required to develop, optimize, and improve a Jain et al., [154] created a multitasking DL consensus model using
model. For four important toxicity properties, recent ML and DL-based RF, DNN, CNN, and GCNN on a dataset of 80,081 compounds that out-
AI-based drug toxicity prediction methods are presented. performed other models with RMSE of 0.65 and R2 of 0.5. Using the
same dataset of 7413 compounds, BTAMDL predicted LD50 with a higher
degree of accuracy than MolGIN (RMSE = 0.557, R2 = 0.662). Using
5.1. Cardiac side effect prediction
the 7413-compound TopTox data set from the ECOTOX aquatic toxicity
database, Karim et al., [155] proposed QuantitativeTox, a DL framework
ML-based methods, including RF, SVM, NB, SVR, kNN, DT, GB (gra-
utilizing FFNN, CNN, GCNN, and baseline feature representations. This
dient boosting), PLS (partial least squares), and XGB, are commonly used
model outperformed others (R2 = 0.687), and it was adopted.
in predicting hERG toxicity (a gene related to a potassium channel in the
heart, and is used to evaluate cardiac side).
Venkatraman developed an ADMET prediction model using ECFP6- 5.3. Drug induced liver injury (DILI) prediction
based RF models [138]. A total of 7889 compounds were assembled
from 4 well-defined experimental assays with experimental hERG block- Drug or chemical toxicity causes DILI [156]. It causes 32 % of drug
ing bioactivities for training, where compounds with experimental val- recalls, which worries researchers and doctors [157]. Predicting human
ues less than or equal to 10 𝜇M were regarded as positive samples (4355 DILI with in vitro or animal studies is difficult. RF, kNN, SVM, deep
in total) and the others as negative samples (3534 in total). RF builds neural network (DNN), CNN, and GNN can predict compound properties
multiple DT on the data and merges them together to get hERG toxicity. from chemical structure. Recent reviews have examined AI methods for
This model achieved an 80 % accuracy and 88 % ROC-AUC. DILI in silico prediction [158-160].

9
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

Recent ML techniques, such as RF, LR, NB, SVM, kNN, AB, GDBT to increase sensitivity while preserving a high positive predictive value
(gradient boosting decision trees), and ET (equivariant transformer) for a variety of illnesses, such as multiple sclerosis and inflammatory
[138,161-165], have been applied to the development of accurate DILI bowel disease [181].
prediction models. Mora et al. created an ensemble model based on A study on phenotyping rheumatoid arthritis using structured data
QuBiLS-MAS Features and Shallow Learning using k-NN, MLP, RF, NB, (ICD codes and medication data) and clinical concepts derived from NLP
SVM, LR (logistic regression), classification tree, Fisher’s linear dis- shows how ML has been used to create phenotyping models. The SVM
criminant analysis (FLDA) [164], and Bayes network algorithms, which models outperformed rule-based techniques in accuracy [182], proving
achieved an 84 % accuracy rate over 10-fold cross-validation using a that a high-performing classifier can be built without the use of feature
training set of 1075 compounds from a previous study [166]. Using a engineering. ML techniques are more scalable, work with less standard-
dataset of 2608 chemicals and SVM and hybrid quantum particle swarm ized datasets than rule-based approaches, and can capture more complex
optimization algorithms, Wang and Chen constructed five consensus phenotypes.
models with an 80 % accuracy rate [165]. However, manually labeled gold standard training and test datasets
ADMETLab 2.0 [167], FP-ADMET [138] and InterpretableADMET are crucial for developing and validating supervised ML models. How-
[150], are ADMET prediction software with over 79 % accurate DILI pre- ever, creating them requires considerable time and expertise. To address
diction models. DL improves the accuracy of DILI prediction. Li et al.’s this, unsupervised learning techniques have been proposed to gener-
[168] DeepDILI model outperformed five ML algorithms and two ad- ate patient clusters for specific medical conditions without human su-
vanced ensemble methods. Using Transcriptional Response Data and pervision. Ho et al. used the “Limestone” non-negative tensor factor-
GNN, Hwang et al. [169] developed GLIT, a DILI prediction model that ization technique to automatically generate multiple phenotype candi-
outperformed baseline models with 77.3 % accuracy. With a multiview dates without predefined definitions [183]. A medical professional de-
GNN approach, Ma et al. [170] increased accuracy from 78.8 % to termined that only 40 of the top 50 candidates are necessary for higher
81.4 % and ROC-AUC from 86.6 % to 88.8 %. Using CNN algorithms predictive accuracy of patients at risk of heart failure. Two upgraded
and molecular fingerprint-embedded features, Nguyen-Vo et al. [171] variations of Limestone, Marble, and Granite, showed improved perfor-
developed a DILI prediction model with 89 % accuracy and 96 % ROC- mance [184,185]. Numerous phenotyping methods have been reported
AUC. CNN’s ResNet18DNN achieved a 95.8 % success rate on 1446 com- in line with the use of DL approaches, including Gehrmann et al., [186]
pounds [172]. These methods may reduce drug recalls and improve the and Yang et al., [187] use of discharge summaries or clinicians’ notes,
accuracy of DILI prediction. and Miotto et al., [188] de-noising auto-encoders for auto-encoding.
In genomics research, DL has emerged as a prominent class of algo-
5.4. Carcinogenesis prediction rithms owing to its capability to effectively handle large datasets with
high dimensionality. A plethora of DL-based models have been devel-
Potentially carcinogenic substances must be identified in order to oped and utilized for designing gRNA (guide RNA), incorporating fea-
prevent environmental cancers [173]. Many FDA-approved medications tures from both sequence and secondary structure data. Additionally,
have been withdrawn due to their carcinogenic characteristics. Short- transfer learning, a ground-breaking innovation in computer vision, has
term biological studies and theoretical models have been tested to find been used in genomics research as well to take advantage of trained
such compounds. ML and DL techniques can be used to replace, scale models and only need small sample sizes [189]. Moreover, BERT mod-
back, and enhance animal studies. els have been specially created for NLP tasks in the clinical domain
Recently, a number of AI-based models and tools for compound car- [190]. ClinicalBERT and Discharge Summary BERT were developed by
cinogenicity prediction were created [174-178]. Using hybrid neural Alsentzer et al., [191] and pre-trained on millions of clinical notes from
networks (NN), Limbu and Dakshanamurthy developed carcinogenic- the MIMIC-III database [192,193]. These MIMIC notes were split into
ity prediction models with an average accuracy of 74.3 and an average sections followed by sentences extraction as input into the model, which
ROC-AUC of 80.1 [174]. Due to sparse data sets, DL models have low captured contextual relationships between words bidirectionally. These
predictive accuracy. Wang et al. developed CapsCarcino, a new DL ar- models were pre-trained and fine-tuned for downstream tasks, outper-
chitecture, to address this problem [175]. On a set of external validation forming BERT and BioBERT on three clinical NLP tasks. In order to pre-
data, CapsCarcino had an average accuracy of 74.5 percent and an accu- dict hospital readmission, Huang et al. also created ClinicalBERT, but
racy prediction rate of 83 %. On sparse training data that was arbitrarily they also emphasized the drawbacks of using data from a single health-
reduced to 20 %, 40 %, 60 %, and 80 % of the full training data, Cap- care institution. Retraining on bigger databases of clinical notes is ad-
sCarcino performed better than other models. Li et al. developed the vised as a result for better performance. These advancements demon-
DeepCarc model, which has an average improvement rate of 37.0 % strate how ML and its related fields can enhance genomics and clinical
and an accuracy of 75.4 %, using 863 compounds and three descriptors research.
[176]. Other reviews [173,179] cover additional models.
7. AI-based modeling for personalized drug dosing
6. AI and gene editing technologies for developing gene therapies
Traditionally, clinical practice has been based on the concept of
With the growing accumulation of genomic and clinical data, data “one therapy fits all’. However, drug molecules may undergo differ-
scientists face both challenges and opportunities when attempting to ex- ent metabolic activities in different patients. For example, a drug that
tract biologically or clinically relevant information from massive geno- works well for a group of people may not be as effective or may have
type and phenotype datasets. In genomics, AI-based technologies and adverse side effects for others. These differences in drug metabolism are
data science techniques have been utilized effectively over the past two mostly attributed to the differences in the genetic profile of individu-
decades. als. Thus, a more futuristic approach is the personalized treatment also
A significant amount of phenotypic information can be found in the known as precision medicine, where patients are treated based on their
clinical notes, discharge summaries, radiology, and pathology reports genetic profile. The target is to maximize treatment outcomes while
that make up about 80 % of the unstructured data in EHRs [180]. Clin- minimizing adverse effects per individual. Thus, different therapies and
ical NLP methods like cTAKES can parse semantic relationships and ex- doses are customized per individual (or per group of patients that share
tract structured concepts from free text to extract this information. The similar genome profiles). AI has fostered considerable improvements in
accuracy of phenotyping has significantly improved with the use of NLP the development of personalized medicine [194]. For example, the AI-
techniques in combination with structured and unstructured data. Ac- derived platform, CURATE.AI, predicts the optimal dosing along with
cording to Liao et al.’s analysis of structured data, NLP can be added to it the treatment outcomes based on the patients’ individual data. It gen-

10
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

erates a profile for each patient using their own medical records, and In general, AI-based technologies offer promising solutions to im-
it dynamically recalibrates the predicted profile over time based on the prove the diagnosis and management of rare diseases. However, the
progression or recession of the disease. CURATE.AI can optimize doses ethical, legal, and social implications of AI in healthcare must be care-
of not only single drugs, but also combinations of drugs [195,196]. This fully considered to ensure the safety and effectiveness of AIMDs. With
is helpful given that, nowadays, therapies are becoming more sophisti- careful consideration and collaboration, AI has the potential to revo-
cated with emphasis on combination (or multimodal) treatments. These lutionize the diagnosis and treatment of rare diseases, leading to im-
involve more than one drug or treatment offered either simultaneously proved patient outcomes and a better quality of life for those affected
or sequentially. Combination therapy is proven to have more efficacy by RDs.
compared to single drug regimen especially in the treatment of com-
plex diseases like cancer [197,198]. To predict the efficacy of a chosen
9. Conclusion
treatment, Kureshi et al. developed an AI decision tree to establish a
link between the characteristics of the patient and the tumor response
The use of AI technology in drug design has grown rapidly due to
in NSCLC [199]. They used four classifiers (histology, mutation in epi-
its predictive ability and accuracy. This review highlights the numerous
dermal growth factor receptor, targeted drugs, and smoking habits) for
applications of AI in all phases of drug development, from disease diag-
predicting the response of NSCLC patients to the EGFR tyrosine kinase
nosis to post-marketing analysis. AI helps in the early prediction of dis-
inhibitors. The method showed 76.6 % accuracy and it can support the
eases, the development of personalized medicine, optimization of drug
clinician in choosing the correct treatment for NSCLC patients. One of
doses, and the prediction of treatment outcomes. Additionally, AI assists
the drawbacks of the study is the small training set used (n = 355).
in target and lead identification through the prediction of protein struc-
This resulted in the omission of rare patterns such as duplication, dele-
tures and biological activities of small molecules. AI technology can also
tions, insertions, and point mutations. Using a larger training set could
predict drug-like properties and off-target effects of new compounds,
further improve the predictive accuracy of this decision support model.
reducing the need for experimental validation. Furthermore, AI-driven
The ‘IBM Watson for oncology’ software made a large impact on person-
approaches improve patient stratification, recruitment, monitoring, and
alized treatment plans for cancer patients [200]. The software is trained
follow-ups in clinical trials, and can even assist in FDA approvals and
on thousands of clinical and health records of cancer patients from the
pharmacovigilance. The integration of AI in drug design has resulted
medical journals, textbooks, and literature curated by Memorial Sloan
in faster drug discovery, cost savings, reduced resource and manpower
Kettering. This software makes accurate diagnoses and treatment rec-
usage, and decreased attrition rates in clinical trials. Additionally, AI
ommendations by identifying related cases from databases of worldwide
helps to minimize the use of in vivo bioassays, reducing animal sacri-
clinical trials (http://www.clinicaltrials.gov) [201].
fice. AI has far-reaching applications beyond medicine, including health-
care management, surgeries, mRNA vaccination, preventive treatments,
8. The role of AI in rare disease research
and nutrigenomics. However, it is important to note that AI models are
meant to complement human intelligence, not replace it. AI models may
Rare diseases (RDs) are a significant health issue that affects almost
have comparable or better predictive ability than human researchers,
1 in 10 individuals in US [202]. Despite their prevalence, the diagno-
but they still lack human intuition. Predictions made by AI machines
sis of RDs is often challenging due to the complexity of symptoms and
must be verified by humans, as AI models can provide false positive and
the rarity of the conditions. The delay in diagnosis can be as long as 7
false negative results, compromising the sensitivity and specificity of
years, leading to significant delays in treatment and management [203].
the model. Additionally, resource sustainability needs holistic solutions
Hence, there is a need for new approaches to enhance the diagnosis and
like cost-aware cross-layer co-design, integrating hardware, algorithms,
treatment of RDs. AI has the potential to transform the diagnosis and
and models for efficient exploration of resource-sustainable configura-
management of rare diseases [204-207] based on NB, RF, XGBoost, CNN,
tions. Consensus-based distributed learning is suggested to fully utilize
AE (autoencoder), RNN, GAN (generative adversarial network), etc. Fer-
existing and future computing infrastructures, incorporating Internet-
nández et al. developed a deep DL-based approach to detect tubers in
of-Things devices and edge servers for data sharing while ensuring pri-
selected MRI (magnetic resonance imaging) images for the diagnosis of
vacy. Stable infrastructures with AI-enhanced resource allocation are
tuberous sclerosis complex (TSC) [204]. This model adopts a unique
recommended, involving dedicated healthcare AI infrastructures com-
InceptionV3 CNN architecture to recognize whether an MRI image has
pliant with evolving government regulations. Lastly, interpretable self-
tubers in it or not, showing promising performance (accuracy: 95 %)
supervised learning is proposed to address the sustainability issue in
in the detection of a rare neurological disorder. Founta introduced a
domain expertise, enhancing trust by extracting clinically useful fea-
semi-automated preprocessing gene selection methodology to identify
tures and providing human-interpretable evidence in healthcare appli-
causal amyotrophic lateral sclerosis (ALS) genes [205], with which they
cations. There are numerous challenges associated with AI, including
developed a classifier based on XGBoost and RF to diagnose ALS and
the explainability of models, the quality and suitability of data used to
its specific subtypes. This methodology achieved 88.89 % accuracy for
train models, avoiding bias and overfitting, resource sustainability and
the classification of sporadic ALS motor neuron samples. Additionally,
more. It is crucial to remain aware of the limitations and risks asso-
AI-based PET is a promising tool for early detection and diagnosis of
ciated with AI technology. Opportunities for improvement in AI tech-
RDs [208].
nology include minimizing dependence on supercomputing power, ad-
However, the implementation of AI in healthcare requires careful
dressing ethical concerns surrounding data collection, and implement-
consideration of ethical, legal, and social implications [209]. AI med-
ing AI in a controlled manner in the healthcare sector to limit nega-
ical devices must be developed with the active involvement of patient
tive consequences. It is possible that the future of AI-assisted drug dis-
advocacy groups to ensure that the technology is designed to meet the
covery lies in developing a virtual human with complete complexity,
specific needs of rare disease patients. The datasets used to train these
allowing for accurate predictions of all possible interactions between
algorithms must be diverse and augmented to ensure that they repre-
molecules and exploring all therapeutic potentials and adverse side
sent the end-user population accurately. Furthermore, the safety and
effects.
effectiveness of AI-based medical devices (AIMDs) must be thoroughly
evaluated to avoid potential harm to patients [210]. AIMDs must be RD-
aware at every stage of their conceptualization and life cycle to avoid Declaration of competing interest
potential harm and unsustainable deployment of AIMDs into clinical
practice. This requires a multidisciplinary approach involving clinicians, The authors declare that they have no conflicts of interest in this
computer scientists, and patient advocacy groups. work.

11
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

Acknowledgments [30] L. Yu, X. Shi, X. Liu, et al., Artificial intelligence systems for diagnosis and clinical
classification of COVID-19, Front Microbiol 12 (2021) 729455.
[31] A.A. Castro, T.D. Antonio, E.C. Martínez, et al., Usefulness of chest X-rays for eval-
This work was supported by the National Key R&D Program of China uating prognosis in patients with COVID-19, Radiologia 63 (2021) 476–483.
(2023YFF1205103), National Natural Science Foundation of China [32] W. Xie, F. Wang, Y. Li, et al., Advances and challenges in de novo drug design
(81925034, 22237005), the Key Research and Development Program using three-dimensional deep generative models, J Chem Inf Model 62 (2022)
2269–2279.
of Ningxia Hui Autonomous Region (2022CMG01002), the Starry Night [33] W.T. Li, J. Ma, N. Shende, et al., Using machine learning of clinical data to diagnose
Science Fund of Zhejiang University Shanghai Institute for Advanced COVID-19: a systematic review and meta-analysis, BMC Med Inform Decis Mak 20
Study (SN-ZJU-SIAS-007), the innovative research team of high-level (2020) 1–13.
[34] L. Capoferri, A. Lodola, S. Rivara, et al., Quantum mechanics/molecular
local universities in Shanghai (SHSMU-ZDCX20212700).
mechanics modeling of covalent addition between EGFR-cysteine 797 and
N-(4-anilinoquinazolin-6-yl) acrylamide, J Chem Inf Model 55 (2015) 589–599.
[35] A. Chatzigoulas, Z. Cournia, Rational design of allosteric modulators: challenges
References and successes, WIREs Comput Mol Sci 11 (2021) e1529.
[36] Q. Zhang, Y. Chen, D. Ni, et al., Targeting a cryptic allosteric site of SIRT6 with
[1] I.V. Hinkson, B. Madej, E.A. Stahlberg, Accelerating therapeutics for opportunities small-molecule inhibitors that inhibit the migration of pancreatic cancer cells, Acta
in medicine: a paradigm shift in drug discovery, Front Pharmacol 11 (2020) 770. Pharm Sin B 12 (2022) 876–889.
[2] R. Barker, A flexible blueprint for the future of drug development, Lancet 375 [37] S. Spänig, A. Emberger-Klein, J.-P. Sowa, et al., The virtual doctor: an interactive
(2010) 357–359. clinical-decision-support system based on deep learning for non-invasive prediction
[3] D. Sun, W. Gao, H. Hu, et al., Why 90% of clinical drug development fails and how of diabetes, Artif Intell Med 100 (2019) 101706.
to improve it? Acta Pharm Sin B 12 (2022) 3049–3062. [38] V. Gulshan, L. Peng, M. Coram, et al., Development and validation of a deep learn-
[4] H. Dowden, J. Munro, Trends in clinical success rates and therapeutic focus, Nat ing algorithm for detection of diabetic retinopathy in retinal fundus photographs,
Rev Drug Discov 18 (2019) 495–496. JAMA 316 (2016) 2402–2410.
[5] P. Hassanzadeh, F. Atyabi, R. Dinarvand, The significance of artificial intelligence [39] A.A. Van Der Heijden, M.D. Abramoff, F. Verbraak, et al., Validation of automated
in drug delivery system design, Adv Drug Deliv Rev 151-152 (2019) 169–190. screening for referable diabetic retinopathy with the IDx-DR device in the Hoorn
[6] J.D. Bolcer, R.B. Hermann, The development of computational chemistry in the Diabetes Care System, Acta Ophthalmol 96 (2018) 63–68.
United States, Rev Comput Chem (1994) 1–63. [40] X. Liu, K. Chen, T. Wu, et al., Use of multimodality imaging and artificial intelli-
[7] F. Bianconi, M. Filippucci, Digital Wood Design: Innovative Techniques of Repre- gence for diagnosis and prognosis of early stages of Alzheimer’s disease, Transl Res
sentation in Architectural Design, Springer, 2019 Vol. 24. 194 (2018) 56–67.
[8] Y. Ding, J.H. Sohn, M.G. Kawczynski, et al., A deep learning model to predict a [41] S. Basheer, S. Bhatia, S.B. Sakri, Computational modeling of dementia predic-
diagnosis of Alzheimer disease by using 18F-FDG PET of the brain, Radiology 290 tion using deep neural network: analysis on OASIS dataset, IEEE Access 9 (2021)
(2019) 456–464. 42449–42462.
[9] A. Narin, C. Kaya, Z. Pamuk, Automatic detection of coronavirus disease (covid-19) [42] T. Jo, K. Nho, A.J. Saykin, Deep learning in Alzheimer’s disease: diagnostic classi-
using x-ray images and deep convolutional neural networks, Pattern Anal Appl 24 fication and prognostic prediction using neuroimaging data, Front Aging Neurosci
(2021) 1207–1220. 11 (2019) 220.
[10] F. Gentile, J.C. Yaacoub, J. Gleave, et al., Artificial intelligence–enabled virtual [43] X.H. Ma, Z. Shi, C. Tan, et al., In-silico approaches to multi-target drug discovery:
screening of ultra-large chemical libraries with deep docking, Nat Protoc 17 (2022) computer aided multi-target drug design, multi-target virtual screening, Pharm Res
672–697. 27 (2010) 739–749.
[11] K.A. Carpenter, D.S. Cohen, J.T. Jarrell, et al., Deep learning and virtual drug [44] I. Beheshti, H. Demirel, H. Matsuda, et al., Classification of Alzheimer’s disease and
screening, Future Med Chem 10 (2018) 2557–2567. prediction of mild cognitive impairment-to-Alzheimer’s conversion from structural
[12] M. Bule, N. Jalalimanesh, Z. Bayrami, et al., The rise of deep learning and trans- magnetic resource imaging using feature ranking and a genetic algorithm, Comput
formations in bioactivity prediction power of molecular modeling tools, Chem Biol Biol Med 83 (2017) 109–119.
Drug Des 98 (2021) 954–967. [45] A. Esteva, B. Kuprel, R.A. Novoa, et al., Dermatologist-level classification of skin
[13] C. Berzuini, R. Bellazzi, S. Quaglini, et al., Bayesian networks for patient monitor- cancer with deep neural networks, Nature 542 (2017) 115–118.
ing, Artif Intell Med 4 (1992) 243–260. [46] A. Albayrak, G. Bilgin, in: Proceedings of the 2016 IEEE 17th Interna-
[14] S. Kapsiani, B.J. Howlin, Random forest classification for predicting lifespan-ex- tional symposium on computational intelligence and informatics (CINTI), 2016,
tending chemical compounds, Sci Rep 11 (2021) 1–13. pp. 000335–000340.
[15] K. Heikamp, J. Bajorath, Support vector machines for drug discovery, Expert Opin [47] J.L. Causey, J. Zhang, S. Ma, et al., Highly accurate model for prediction of lung
Drug Discov 9 (2014) 93–104. nodule malignancy with CT scans, Sci Rep 8 (2018) 9286.
[16] F. Zhong, J. Xing, X. Li, et al., Artificial intelligence in drug design, Sci China Life [48] I. Shiri, H. Maleki, G. Hajianfar, et al., Next-generation radiogenomics sequencing
Sci 61 (2018) 1191–1204. for prediction of EGFR and KRAS mutation status in NSCLC patients using mul-
[17] N. Brown, P. Ertl, R. Lewis, et al., Artificial intelligence in chemistry and drug timodal imaging and machine learning algorithms, Mol Imaging Biol 22 (2020)
design, J Comput Aided Mol Des 34 (2020) 709–715. 1132–1148.
[18] C. Provenzano, M. Cappella, R. Valaperta, et al., CRISPR/Cas9-mediated deletion [49] C.L. Srinidhi, O. Ciga, A.L. Martel, Deep neural network models for computational
of CTG expansions recovers normal phenotype in myogenic cells derived from my- histopathology: a survey, Med Image Anal 67 (2021) 101813.
otonic dystrophy 1 patients, Mol Ther–Nucleic Acids 9 (2017) 337–348. [50] Ş. Öztürk, B. Akdemir, HIC-net: a deep convolutional neural network model for
[19] H.-C. Yi, Z.-H. You, X. Zhou, et al., ACP-DL: a deep learning long short-term memory classification of histopathological breast images, Comput Electr Eng 76 (2019)
model to predict anticancer peptides using high-efficiency feature representation, 299–310.
Mol Ther–Nucleic Acids 17 (2019) 1–9. [51] Z. Hameed, S. Zahia, B. Garcia-Zapirain, et al., Breast cancer histopathology image
[20] H. Yi, Z.H. You, X. Zhou, L. Cheng, X. Li, T.H. Jiang, ACP-DL: A deep learning classification using an ensemble of deep learning models, Sensors 20 (2020) 4373.
long short-term memory model to predict anticancer peptides using high-efficiency [52] B.S. Yadav, V. Tripathi, Recent advances in the system biology-based target iden-
feature representation, Mol Ther–Nucleic Acids 17 (2019) 9. tification and drug discovery, Curr Top Med Chem 18 (2018) 1737–1744.
[21] C. Shen, J. Ding, Z. Wang, et al., From machine learning to deep learning: advances [53] C. Nantasenamat, C. Isarankura-Na-Ayudhya, V. Prachayasittikul, Advances in
in scoring functions for protein–ligand docking, WIREs Comput Mol Sci 10 (2020) computational methods to predict the biological activity of compounds, Expert
e1429. Opin Drug Discov 5 (2010) 633–654.
[22] Y. Shen, T. Liu, J. Chen, et al., Harnessing artificial intelligence to optimize [54] A. Assaiya, A.P. Burada, S. Dhingra, et al., An overview of the recent advances in
long-term maintenance dosing for antiretroviral-naive adults with HIV-1 infection, cryo-electron microscopy for life sciences, Emerg Top Life Sci 5 (2021) 151–168.
Adv Ther (Weinh) 3 (2020) 1900114. [55] S.G. Wolf, E. Shimoni, M. Elbaum, et al., STEM tomography in biology, in: Cellular
[23] M. Zhang, Q. Su, Y. Lu, M. Zhao, Application of machine learning approaches for Imaging: Electron Tomography and Related Techniques, 2018, pp. 33–60.
protein-protein Interactions prediction, Med Chem 13 (2017) 506–514. [56] E.D. Zhong, T. Bepler, B. Berger, et al., CryoDRGN: reconstruction of heterogeneous
[24] F. Yu, H.H. Ip, Semantic content analysis and annotation of histological images, cryo-EM structures using neural networks, Nat Methods 18 (2021) 176–185.
Comput Biol Med 38 (2008) 635–649. [57] L.F. Kinman, B.M. Powell, E.D. Zhong, et al., Uncovering structural ensembles from
[25] G.B. Goh, N.O. Hodas, A. Vishnu, Deep learning for computational chemistry, J single-particle cryo-EM data using cryoDRGN, Nat Protoc 18 (2023) 319–339.
Comput Chem 38 (2017) 1291–1307. [58] J. Yang, Z. Gao, X. Ren, et al., DeepDigest: prediction of protein proteolytic diges-
[26] M. Singer, C.S. Deutschman, C.W. Seymour, et al., The third international consensus tion with deep learning, Anal Chem 93 (2021) 6094–6103.
definitions for sepsis and septic shock (Sepsis-3), JAMA 315 (2016) 801–810. [59] B. Sun, P. Smialowski, T. Straub, et al., Investigation and highly accurate prediction
[27] M.P. Sendak, W. Ratliff, D. Sarro, et al., Real-world integration of a sepsis deep of missed tryptic cleavages by deep learning, J Proteome Res 20 (2021) 3749–3757.
learning technology into routine clinical care: implementation study, JMIR Med [60] K. Murata, M. Wolf, Cryo-electron microscopy for structural analysis of dynamic
Inform 8 (2020) e15182. biological macromolecules, Biochimica et Biophysica Acta 1862 (2018) 324–334.
[28] G. Husabø, I.L. Teig, J.C. Frich, et al., Promoting leadership and quality improve- [61] B. Webb, A. Sali, Comparative protein structure modeling using MODELLER, Curr
ment through external inspections of management of sepsis in Norwegian hospitals: Protoc Bioinformatics 54 (2016) 5.6. 1-5.6. 37.
a focus group study, BMJ Open 10 (2020) e041997. [62] J. Jumper, R. Evans, A. Pritzel, et al., Highly accurate protein structure prediction
[29] N. Alballa, I. Al-Turaiki, Machine learning approaches in COVID-19 diagnosis, mor- with AlphaFold, Nature 596 (2021) 583–589.
tality, and severity risk prediction: a review, Inform Med Unlocked 24 (2021) [63] K. Tunyasuvunakool, J. Adler, Z. Wu, et al., Highly accurate protein structure pre-
100564. diction for the human proteome, Nature 596 (2021) 590–596.

12
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

[64] G.R. Buel, K.J. Walters, Can AlphaFold2 predict the impact of missense mutations [101] M. Mochizuki, S.D. Suzuki, K. Yanagisawa, et al., QEX: target-specific druglikeness
on structure? Nat Struct Mol Biol 29 (2022) 1–2. filter enhances ligand-based virtual screening, Mol Divers 23 (2019) 11–18.
[65] A. Perrakis, T.K. Sixma, AI revolutions in biology: the joys and perils of AlphaFold, [102] Y. Li, L. Wang, Z. Liu, et al., Predicting selective liver X receptor 𝛽 agonists using
EMBO Rep 22 (2021) e54046. multiple machine learning methods, Mol Biosyst 11 (2015) 1241–1250.
[66] H. Tian, X. Jiang, P. Tao, PASSer: prediction of allosteric sites server, Mach Learn [103] L. Zhang, A. Sedykh, A. Tripathi, et al., Identification of putative estrogen receptor–
Sci Technol 2 (2021) 035015. mediated endocrine disrupting chemicals using QSAR-and structure-based virtual
[67] W. Huang, S. Lu, Z. Huang, et al., Allosite: a method for predicting allosteric sites, screening approaches, Toxicol Appl Pharmacol 272 (2013) 67–76.
Bioinformatics 29 (2013) 2357–2359. [104] C. Cortes, V. Vapnik, Support-vector networks, Mach Learn 20 (1995) 273–297.
[68] J.G. Greener, M.J. Sternberg, AlloPred: prediction of allosteric pockets on proteins [105] S. Chandra, J. Pandey, A.K. Tamrakar, et al., Multiple machine learning based
using normal mode perturbation analysis, BMC Bioinformatics 16 (2015) 1–7. descriptive and predictive workflow for the identification of potential PTP1B in-
[69] J.A. Suykens, J. Vandewalle, Least squares support vector machine classifiers, Neu- hibitors, J Mol Graph Modell 71 (2017) 242–256.
ral Process Lett 9 (1999) 293–300. [106] A.L. Deshmukh, S. Chandra, D.K. Singh, et al., Identification of human flap en-
[70] A. Liaw, M. Wiener, Classification and regression by randomForest, R news 2 (2002) donuclease 1 (FEN1) inhibitors using a machine learning based consensus virtual
18–22. screening, Mol Biosyst 13 (2017) 1630–1639.
[71] T. Chen, C. Guestrin, in: Proceedings of the Proceedings of the 22nd ACM [107] H. Baba, J.-i. Takahara, H. Mamitsuka, In silico predictions of human skin perme-
SIGKDD International Conference On Knowledge Discovery and Data Mining, 2016, ability using nonlinear quantitative structure–property relationship models, Pharm
pp. 785–794. Res 32 (2015) 2360–2371.
[72] A.H. Basit, W.A. Abbasi, A. Asif, et al., Training host-pathogen protein–protein [108] M. AlQuraishi, End-to-end differentiable learning of protein structure, Cell Syst 8
interaction predictors, J Bioinform Comput Biol 16 (2018) 1850014. (2019) 292–301 e293.
[73] K. Li, S. Zhang, D. Yan, et al., Prediction of hot spots in protein–DNA binding [109] K. Lee, M. Lee, D. Kim, Utilizing random Forest QSAR models with optimized pa-
interfaces based on supervised isometric feature mapping and extreme gradient rameters for target identification and its application to target-fishing server, BMC
boosting, BMC Bioinformatics 21 (2020) 1–10. Bioinformatics 18 (2017) 75–86.
[74] R.E. Amaro, Will the real cryptic pocket please stand out? Biophys J 116 (2019) [110] C.F. Matta, A.A. Arabi, Electron-density descriptors as predictors in quantitative
753–754. structure–activity/property relationships and drug design, Future Med Chem 3
[75] C.R. Knoverek, G.K. Amarasinghe, G.R. Bowman, Advanced methods for accessing (2011) 969–994.
protein shape-shifting present new therapeutic opportunities, Trends Biochem Sci [111] A.M. Osman, A.A. Arabi, Quantum and classical evaluations of carboxylic acid
44 (2019) 351–364. bioisosteres: from capped moieties to a drug molecule, ACS Omega 8 (2022)
[76] C.C. Witt, Y. Ono, E. Puschmann, et al., Induction and myofibrillar targeting of 588–598.
CARP, and suppression of the Nkx2. 5 pathway in the MDM mouse with impaired [112] A.R. Melge, S. Parate, K. Pavithran, et al., Discovery of anticancer hybrid molecules
titin-based signaling, J Mol Biol 336 (2004) 145–154. by supervised machine learning models and in vitro validation in drug resistant
[77] M.A. Cruz, T.E. Frederick, U.L. Mallimadugula, et al., A cryptic pocket in Ebola chronic myeloid leukemia cells, J Chem Inf Model 62 (2022) 1126–1146.
VP35 allosterically controls RNA binding, Nat Commun 13 (2022) 2269. [113] G. Dhamodharan, C.G. Mohan, Machine learning models for predicting the activity
[78] A. Meller, M.D. Ward, J.H. Borowsky, et al., Predicting the locations of cryptic of AChE and BACE1 dual inhibitors for the treatment of Alzheimer’s disease, Mol
pockets from single protein structures using the PocketMiner graph neural network, Divers (2022) 1–17.
Biophys J 122 (2023) 445a. [114] J. Hardy, D.J. Selkoe, The amyloid hypothesis of Alzheimer’s disease: progress and
[79] S.A. Hollingsworth, B. Kelly, C. Valant, et al., Cryptic pocket formation under- problems on the road to therapeutics, Science (1979) 297 (2002) 353–356.
lies allosteric modulator selectivity at muscarinic GPCRs, Nat Commun 10 (2019) [115] M. Rapoport, H.N. Dawson, L.I. Binder, et al., Tau is essential to 𝛽-amyloid-induced
3289. neurotoxicity, Proc Natl Acad Sci 99 (2002) 6364–6369.
[80] P. Cimermancic, P. Weinkam, T.J. Rettenmaier, et al., CryptoSite: expanding the [116] N.J. Woolf, The critical role of cholinergic basal forebrain neurons in morpholog-
druggable proteome by characterization and prediction of cryptic binding sites, J ical change and memory encoding: a hypothesis, Neurobiol Learn Mem 66 (1996)
Mol Biol 428 (2016) 709–719. 258–266.
[81] Q. Li, T. Cheng, Y. Wang, et al., PubChem as a public resource for drug discovery, [117] M.P. Mazanetz, P.M. Fischer, Untangling tau hyperphosphorylation in drug design
Drug Discov Today 15 (2010) 1052–1057. for neurodegenerative diseases, Nat Rev Drug Discov 6 (2007) 464–479.
[82] B. Chen, D.J. Wild, PubChem BioAssays as a data source for predictive models, J [118] W. Sun, H.Y. Qureshi, P.W. Cafferty, et al., Glycogen synthase kinase-3𝛽 is
Mol Graph Modell 28 (2010) 420–426. complexed with tau protein in brain microtubules, J Biol Chem 277 (2002)
[83] J. Lamb, The Connectivity Map: a new tool for biomedical research, Nat Rev Cancer 11933–11940.
7 (2007) 54–60. [119] A. Arif, Extraneuronal activities and regulatory mechanisms of the atypical cy-
[84] J. Lamb, E.D. Crawford, D. Peck, et al., The Connectivity Map: using gene-expres- clin-dependent kinase Cdk5, Biochem Pharmacol 84 (2012) 985–993.
sion signatures to connect small molecules, genes, and disease, Science 313 (2006) [120] J. Fang, Y. Li, R. Liu, et al., Discovery of multitarget-directed ligands against
1929–1935. Alzheimer’s disease through systematic prediction of chemical–protein interac-
[85] I. Kozlovskii, P. Popov, Structure-based deep learning for binding site detection in tions, J Chem Inf Model 55 (2015) 149–164.
nucleic acid macromolecules, NAR Genom Bioinform 3 (2021) lqab111. [121] C. Gorgulla, A. Boeszoermenyi, Z.-F. Wang, et al., An open-source drug discovery
[86] I. Kozlovskii, P. Popov, Spatiotemporal identification of druggable binding sites platform enables ultra-large virtual screens, Nature 580 (2020) 663–668.
using deep learning, Commun Biol 3 (2020) 618. [122] B.J. Bender, S. Gahbauer, A. Luttens, et al., A practical guide to large-scale docking,
[87] I. Kozlovskii, P. Popov, Protein–peptide binding site detection using 3D convolu- Nat Protoc 16 (2021) 4799–4832.
tional neural networks, J Chem Inf Model 61 (2021) 3814–3823. [123] J. Lyu, S. Wang, T.E. Balius, et al., Ultra-large library docking for discovering new
[88] X. Zeng, S. Zhu, W. Lu, et al., Target identification among known drugs by deep chemotypes, Nature 566 (2019) 224–229.
learning from heterogeneous networks, Chem Sci 11 (2020) 1775–1797. [124] W.L. Jorgensen, The many roles of computation in drug discovery, Science 303
[89] P. Mamoshina, M. Volosnikova, I.V. Ozerov, et al., Machine learning on human (2004) 1813–1818.
muscle transcriptomic data for biomarker discovery and tissue-specific drug target [125] Y. Li, M.A. Rezaei, C. Li, et al., in: Proceedings of the 2019 IEEE International
identification, Front Genet 9 (2018) 242. Conference on Bioinformatics and Biomedicine (BIBM), 2019, pp. 303–310.
[90] S. Dara, S. Dhamercherla, S.S. Jadav, et al., Machine learning in drug discovery: a [126] J. Lim, S. Ryu, K. Park, et al., Predicting drug–target interaction using a novel
review, Artif Intell Rev 55 (2022) 1947–1999. graph neural network with 3D structure-embedded graph representation, J Chem
[91] W. Feng, L. Wang, Z. Lin, et al., Generation of 3D Molecules in Pockets via Language Inf Model 59 (2019) 3981–3988.
Model, arXiv preprint (2023) arXiv:2305.10133. [127] M. Jiang, Z. Li, S. Zhang, et al., Drug–target affinity prediction using graph neural
[92] A. Smith, Screening for drug discovery: the leading question, Nature 418 (2002) network and contact maps, RSC Adv 10 (2020) 20701–20712.
453–455. [128] M. Rudrapal, S.J. Khairnar, A.G. Jadhav, Drug repurposing (DR): an emerging ap-
[93] M. Butkiewicz, Y. Wang, S.H. Bryant, et al., High-throughput screening assay proach in drug discovery, Drug Repurposing-Hypothesis, Molecular Aspects and
datasets from the pubchem database, Chem Inform 3 (2017) 1. Therapeutic Applications (2020) 10.
[94] N. Berdigaliyev, M. Aljofan, An overview of drug discovery and development, Fu- [129] S.H. Sleigh, C.L. Barton, Repurposing strategies for therapeutics, Pharmaceut Med
ture Med Chem 12 (2020) 939–947. 24 (2010) 151–159.
[95] N.S. Pagadala, K. Syed, J. Tuszynski, Software for molecular docking: a review, [130] R.K. Harrison, Phase II and phase III failures: 2013–2015, Nat Rev Drug Discov 15
Biophys Rev 9 (2017) 91–102. (2016) 817–818.
[96] N. Brooijmans, I.D. Kuntz, Molecular recognition and docking algorithms, Annu [131] D. Reker, T. Rodrigues, P. Schneider, et al., Identifying the macromolecular targets
Rev Biophys Biomol Struct 32 (2003) 335–373. of de novo-designed chemical entities through self-organizing map consensus, Proc
[97] A. Tropsha, Best practices for QSAR model development, validation, and exploita- Natl Acad Sci 111 (2014) 4067–4072.
tion, Mol Inform 29 (2010) 476–488. [132] X. Yang, Y. Wang, R. Byrne, et al., Concepts of artificial intelligence for comput-
[98] K.A. Carpenter, X. Huang, Machine learning-based virtual screening and its ap- er-assisted drug discovery, Chem Rev 119 (2019) 10520–10594.
plications to Alzheimer’s drug discovery: a review, Curr Pharm Des 24 (2018) [133] A.H. Vo, T.R. Van Vleet, R.R. Gupta, et al., An overview of machine learning and
3347–3358. big data for drug toxicity evaluation, Chem Res Toxicol 33 (2020) 20–37.
[99] L. Wang, X. Le, L. Li, et al., Discovering new agents active against methicillin-re- [134] J. Liu, W. Guo, F. Dong, et al., machine learning for predicting organ toxicity, in:
sistant Staphylococcus aureus with ligand-based approaches, J Chem Inf Model 54 H. Hong (Ed.), Machine Learning and Deep Learning in Computational Toxicology,
(2014) 3186–3197. Springer International Publishing, Cham, 2023, pp. 519–537.
[100] W. Lian, J. Fang, C. Li, et al., Discovery of Influenza A virus neuraminidase in- [135] A.O. Basile, A. Yahi, N.P. Tatonetti, Artificial intelligence for drug toxicity and
hibitors using support vector machine and Naïve Bayesian models, Mol Divers 20 safety, Trends Pharmacol Sci 40 (2019) 624–635.
(2016) 439–451. [136] E. Pérez Santín, R. Rodríguez Solana, M. González García, et al., Toxicity prediction

13
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

based on artificial intelligence: a multidisciplinary overview, WIREs Comput Mol using convolutional neural network and molecular fingerprint-embedded features,
Sci 11 (2021) e1516. ACS Omega 5 (2020) 25432–25439.
[137] A. Bassan, V.M. Alves, A. Amberg, et al., In silico approaches in organ toxicity haz- [172] Z. Chen, Y. Jiang, X. Zhang, et al., ResNet18DNN: prediction approach of drug-in-
ard assessment: current status and future needs in predicting liver toxicity, Comput duced liver injury by deep neural network with ResNet18, Br Bioinform 23 (2022)
Toxicol 20 (2021) 100187. bbab503.
[138] V. Venkatraman, FP-ADMET: a compendium of fingerprint-based ADMET predic- [173] A. Golbamaki, E. Benfenati, A. Roncaglioni, In silico methods for carcinogenic-
tion models, J Cheminform 13 (2021) 1–12. ity assessment, in: Silico Methods for Predicting Drug Toxicity, Springer, 2022,
[139] Y. Hsiao, B.-H. Su, Y.J. Tseng, Current development of integrated web servers for pp. 201–215.
preclinical safety and pharmacokinetics assessments in drug development, Brief [174] S. Limbu, S. Dakshanamurthy, Predicting environmental chemical carcinogenicity
Bioinformatics 22 (2021) bbaa160. using a hybrid machine-learning approach, bioRxiv (2021) 2021.05.03.442477.
[140] I. Arab, K. Barakat, ToxTree: descriptor-based machine learning models for both [175] Y.-W. Wang, L. Huang, S.-W. Jiang, et al., CapsCarcino: a novel sparse data deep
hERG and Nav1. 5 cardiotoxicity liability predictions, arXiv preprint (2021) learning tool for predicting carcinogens, Food Chem Toxicol 135 (2020) 110921.
arXiv:2112.13467. [176] T. Li, W. Tong, R. Roberts, et al., DeepCarc: deep learning-powered carcinogenicity
[141] K. Ogura, T. Sato, H. Yuki, et al., Support vector machine model for hERG inhibitory prediction using model-level representation, Front Artif Intell (2021) 176.
activities based on the integrated hERG database using descriptor selection by NS- [177] P. Fradkin, A. Young, L. Atanackovic, et al., A graph neural network approach for
GA-II, Sci Rep 9 (2019) 1–12. molecule carcinogenicity prediction, Bioinformatics 38 (2022) i84–i91.
[142] L.S.K. Konda, S.K. Praba, R. Kristam, hERG liability classification models using [178] X. Xiang, Y. Chen, J. Gao, et al., in: Proceedings of the 2021 16th International
machine learning techniques, Comput Toxicol 12 (2019) 100089. Conference on Computer Science & Education (ICCSE), 2021, pp. 864–869.
[143] M. Shan, C. Jiang, J. Chen, et al., Predicting hERG channel blockers with directed [179] R.R. Tice, A. Bassan, A. Amberg, et al., In silico approaches in carcinogenicity haz-
message passing neural networks, RSC Adv 12 (2022) 3423–3430. ard assessment: current status and future needs, Comput Toxicol 20 (2021) 100191.
[144] X. Zhang, J. Mao, M. Wei, et al., HergSPred: accurate classification of hERG [180] F. Martin-Sanchez, K. Verspoor, Big data in medicine is driving big changes, Yearb
blockers/nonblockers with machine-learning models, J Chem Inf Model 62 (2022) Med Inform 23 (2014) 14–20.
1830–1839. [181] K.P. Liao, T. Cai, G.K. Savova, et al., Development of phenotype algorithms using
[145] J.Y. Ryu, M.Y. Lee, J.H. Lee, et al., DeepHIT: a deep learning framework for pre- electronic medical records and incorporating natural language processing, BMJ 350
diction of hERG-induced cardiotoxicity, Bioinformatics 36 (2020) 3049–3055. (2015) h1885.
[146] R. De Jesus, N. Vicuña-Fernández, A. Osorio, et al., Determination of medium lethal [182] R.J. Carroll, A.E. Eyler, J.C. Denny, Proceedings of the AMIA Annual Symposium
dose (LD50) and acute toxicity of formulation Cytoreg®, an ionic mixture of strong Proceedings, 2011, p. 189
and weak acids, Latin Am J Dev 3 (2021) 1121–1126. [183] J.C. Ho, J. Ghosh, J. Sun, in: Proceedings of the Proceedings of the 20th ACM
[147] V. Bhat, J. Chatterjee, The use of in silico tools for the toxicity prediction of poten- SIGKDD International Conference on Knowledge Discovery and Data Mining, 2014,
tial inhibitors of SARS-CoV-2, Alternat Lab Anim 49 (2021) 22–32. pp. 115–124.
[148] K. Morris-Schaffer, M.J. McCoy, A review of the LD50 and its current role in hazard [184] J.C. Ho, J. Ghosh, S.R. Steinhubl, et al., Limestone: high-throughput candidate phe-
communication, ACS Chem Health Saf 28 (2020) 25–33. notype generation via tensor factorization, J Biomed Inform 52 (2014) 199–211.
[149] J. Kramer, Label Review Manual Chapter 7: Precautionary Statements Label Review [185] J. Henderson, J.C. Ho, A.N. Kho, et al., in: Proceedings of the 2017 IEEE Interna-
Manual, (2014). tional Conference on Healthcare Informatics (ICHI), 2017, pp. 214–223.
[150] Y. Wei, S. Li, Z. Li, et al., Interpretable-ADMET: a web service for ADMET prediction [186] S. Gehrmann, F. Dernoncourt, Y. Li, et al., Comparing deep learning and concept
and optimization based on deep neural representation, Bioinformatics 38 (2022) extraction based methods for patient phenotyping from clinical narratives, PLoS
2863–2871. One 13 (2018) e0192360.
[151] D. Ballabio, F. Grisoni, V. Consonni, et al., Integrated QSAR models to predict acute [187] Z. Yang, M. Dehmer, O. Yli-Harja, et al., Combining deep learning with token se-
oral systemic toxicity, Mol Inform 38 (2019) 1800124. lection for patient phenotyping from electronic health records, Sci Rep 10 (2020)
[152] K. Vukovic, D. Gadaleta, E. Benfenati, Methodology of aiQSAR: a group-specific 1–18.
approach to QSAR modelling, J Cheminform 11 (2019) 1–9. [188] R. Miotto, L. Li, B.A. Kidd, et al., Deep patient: an unsupervised representation to
[153] D. Gadaleta, K. Vuković, C. Toma, et al., SAR and QSAR modeling of a large col- predict the future of patients from the electronic health records, Sci Rep 6 (2016)
lection of LD50 rat acute oral toxicity data, J Cheminform 11 (2019) 1–16. 1–10.
[154] S. Jain, V.B. Siramshetty, V.M. Alves, et al., Large-scale modeling of multispecies [189] E. Guarnera, I.N. Berezovsky, Structure-based statistical mechanical model ac-
acute toxicity end points using consensus of multitask deep learning methods, J counts for the causality and energetics of allosteric communication, PLoS Comput
Chem Inf Model 61 (2021) 653–663. Biol 12 (2016) e1004678.
[155] A. Karim, V. Riahi, A. Mishra, et al., Quantitative toxicity prediction via meta en- [190] B. Mieth, A. Rozier, J.A. Rodriguez, et al., DeepCOMBI: explainable artificial in-
sembling of multitask deep learning models, ACS Omega 6 (2021) 12306–12317. telligence for the analysis and discovery in genome-wide association studies, NAR
[156] R.J. Andrade, N. Chalasani, E.S. Björnsson, et al., Drug-induced liver injury, Nat Genom Bioinform 3 (2021) lqab065.
Rev Dis Prim 5 (2019) 58. [191] E. Alsentzer, J.R. Murphy, W. Boag, et al., Publicly available clinical BERT embed-
[157] S. Babai, L. Auclert, H. Le-Louët, Safety data and withdrawal of hepatotoxic drugs, dings, arXiv preprint (2019) arXiv:1904.03323.
Therapies 76 (2021) 715–723. [192] J. Listgarten, M. Weinstein, B.P. Kleinstiver, et al., Prediction of off-target activities
[158] O.J. Béquignon, G. Pawar, B. van de Water, et al., Computational approaches for for the end-to-end design of CRISPR guide RNAs, Nat Biomed Eng 2 (2018) 38–47.
drug-induced liver injury (DILI) prediction: state of the art and challenges, Syst [193] J.G. Doench, N. Fusi, M. Sullender, et al., Optimized sgRNA design to maximize
Med 2 (2021) 308–329. activity and minimize off-target effects of CRISPR-Cas9, Nat Biotechnol 34 (2016)
[159] V.M. Lauschke, Toxicogenomics of drug induced liver injury–from mechanistic un- 184–191.
derstanding to early prediction, Drug Metab Rev 53 (2021) 245–252. [194] F. Boniolo, E. Dorigatti, A.J. Ohnmacht, et al., Artificial intelligence in early
[160] J. Liu, W. Guo, S. Sakkiah, et al., Machine learning models for predicting liver drug discovery enabling precision medicine, Expert Opin Drug Discov 16 (2021)
toxicity, in: In Silico Methods for Predicting Drug Toxicity, 2022, pp. 393–415. 991–1007.
[161] M.G. de Lomana, F. Svensson, A. Volkamer, et al., Consideration of predicted [195] A. Blasiak, J. Khong, T. Kee, CURATE, AI: optimizing personalized medicine with
small-molecule metabolites in computational toxicology, Digit Discov 1 (2022) artificial intelligence, SLAS Technol 25 (2020) 95–105.
158–172. [196] A.J. Pantuck, D.K. Lee, T. Kee, et al., Modulating BET bromodomain inhibitor
[162] X. Liu, D. Zheng, Y. Zhong, et al., Machine-learning prediction of oral drug-induced ZEN-3694 and enzalutamide combination dosing in a metastatic prostate cancer
liver injury (DILI) via multiple features and endpoints, Biomed Res Int 2020 (2020) patient using CURATE. AI, an artificial intelligence platform, Adv Ther 1 (2018)
4795140. 1800104.
[163] Y. Wang, Q. Xiao, P. Chen, et al., In silico prediction of drug-induced liver injury [197] P. Kumar, R. Benedict, F. Urzua, et al., Combination treatment significantly en-
based on ensemble classifier method, Int J Mol Sci 20 (2019) 4106. hances the efficacy of antitumor therapy by preferentially targeting angiogenesis,
[164] J.R. Mora, Y. Marrero-Ponce, C.R. García-Jacas, et al., Ensemble models based on Lab Investig 85 (2005) 756–767.
QuBiLS-MAS features and shallow learning for the prediction of drug-induced liver [198] T.M. MacDonald, B. Williams, D.J. Webb, et al., Combination therapy is superior
toxicity: improving deep learning and traditional approaches, Chem Res Toxicol to sequential monotherapy for the initial treatment of hypertension: a double-blind
33 (2020) 1855–1873. randomized controlled trial, J Am Heart Assoc 6 (2017) e006986.
[165] Y. Wang, X. Chen, Joint decision-making model based on consensus modeling tech- [199] N. Kureshi, S.S.R. Abidi, C. Blouin, A predictive model for personalized therapeutic
nology for the prediction of drug-induced liver injury, J Chem 2021 (2021) 1–20. interventions in non-small cell lung cancer, IEEE J Biomed Health Inform 20 (2014)
[166] C.Y. Liew, Y.C. Lim, C.W. Yap, Mixed learning algorithms and features ensemble 424–431.
in hepatotoxicity prediction, J Comput Aided Mol Des 25 (2011) 855–871. [200] J. Fu, A. Gucalp, M.G. Zauderer, et al., Steps in Developing Watson for Oncol-
[167] G. Xiong, Z. Wu, J. Yi, et al., ADMETlab 2.0: an integrated online platform for ogy, A Decision Support System to Assist Physicians Choosing First-Line Metastatic
accurate and comprehensive predictions of ADMET properties, Nucleic Acids Res Breast Cancer (MBC) therapies: Improved performance With Machine learning. Ed-
49 (2021) W5–W14. itor, Book Steps in Developing Watson for Oncology, a Decision Support System to
[168] T. Li, W. Tong, R. Roberts, et al., DeepDILI: deep learning-powered drug-induced Assist Physicians Choosing First-Line Metastatic Breast Cancer (MBC) therapies:
liver injury prediction using model-level representation, Chem Res Toxicol 34 Improved performance With Machine learning, Series Steps in Developing Watson
(2020) 550–565. for Oncology, A Decision Support System to Assist Physicians Choosing First-Line
[169] D. Hwang, M. Jeon, J. Kang, in: Proceedings of the 2020 IEEE International Con- Metastatic Breast Cancer (MBC) Therapies: Improved performance With Machine
ference on Big Data and Smart Computing (BigComp), 2020, pp. 323–329. Learning, American Society of Clinical Oncology, 2015.
[170] H. Ma, W. An, Y. Wang, et al., Deep graph learning with property augmentation [201] P. Bach, M.G. Zauderer, A. Gucalp, et al., Beyond Jeopardy!: Harnessing IBM’s Wat-
for predicting drug-induced liver injury, Chem Res Toxicol 34 (2020) 495–506. son to Improve Oncology Decision making. Editor, Book Beyond Jeopardy!: Har-
[171] T.-H. Nguyen-Vo, L. Nguyen, N. Do, et al., Predicting drug-induced liver injury nessing IBM’s Watson to Improve Oncology Decision making, Series Beyond Jeop-

14
JID: FMRE
ARTICLE IN PRESS [m5GeSdc;March 19, 2025;21:4]

A.U. Rehman, M. Li, B. Wu et al. Fundamental Research xxx (xxxx) xxx

ardy!: Harnessing IBM’s Watson to Improve Oncology Decision Making, American [210] C. Pang, J. Qiao, X. Zeng, et al., Deep generative models in De Novo drug molecule
Society of Clinical Oncology, 2013. generation, J Chem Inf Model 64 (2024) 2174–2194.
[202] C. Isert, K. Atz, G. Schneider, Structure-based drug design with geometric deep
learning, Curr Opin Struct Biol 79 (2023) 102548. Author profile
[203] J. Jin, D. Wang, G. Shi, et al., FFLOM: a flow-based autoregressive model for frag-
ment-to-lead optimization, J Med Chem 66 (2023) 10808–10823. Jian Zhang (BRID:09778.00.20899) received his BS in pharmacology from Peking Univer-
[204] H. Li, L. Zou, J.A.H. Kowah, et al., A compact review of progress and prospects of sity in 2002 and PhD in medicinal chemistry from Chinese Academy of Sciences in 2007.
deep learning in drug discovery, J Mol Model 29 (2023) 117. He joined Shanghai Jiao Tong University School of Medicine as a professor in 2009 and
[205] J. Li, A. Fu, L. Zhang, An overview of scoring functions used for protein-ligand was rewarded the Changjiang Scholars Program in 2017. He is the director of the Medici-
interactions in molecular docking, Interdiscip Sci 11 (2019) 320–328. nal Chemistry & Bioinformatics Center at Shanghai Jiao Tong University. He has published
[206] S. Lu, X. He, D. Ni, et al., Allosteric modulator discovery: from serendipity to struc- 120+ peer-reviewed papers (e.g. Nat Chem Biol, Chem), invented 20 issued patents, and
ture-based design, J Med Chem 62 (2019) 6405–6421. made hundreds of international invited talks. He was elected the 2017 Top-ten Science
[207] D. Mucs, R.A. Bryce, The application of quantum mechanics in structure-based drug and Technology Young Scientist in China, member of the Academic Degrees Committee
design, Expert Opin Drug Discov 8 (2013) 263–276. of the State Council in 2020, the Fellow of Royal Society of Chemistry and Excellent Re-
[208] D. Ni, Z. Chai, Y. Wang, et al., Along the allostery stream: recent advances in com- search Advisor of American Chemical Society in 2022. He is the associate editor for RSC
putational methods for allosteric drug discovery, WIREs Comput Mol Sci 12 (2021) Medicinal Chemistry, editorial board member for Medicinal Research Review and other seven
e1585. international journals.
[209] D. Ni, Y. Liu, R. Kong, et al., Computational elucidation of allosteric communication
in proteins for allosteric drug design, Drug Discov Today 27 (2022) 2226–2234.

15