Title Year Modality Dataset Terms Info Link

Introduced

2023 biological-k No ‘target , a dense retrieval approach (inspired by https

DrugCL
nowledge PDBBind fishing’, CLIP [31]) that computes a contrastive loss ://ar
IP:
inspired [44] between two separate pre-trained xiv.o
Contras
augmentati encoders to maximize the similarity rg/p
tive
on method, BioLip, between a protein-molecule pair, if they df/2
Protein-
HomoAug, ChEMBL have a binding affinity, and minimize it 310.
Molecul
which otherwise 063
e
creates - 67
Repres
protein-mol ●​ For encoder they use encoder
entation
ecule pairs architecture of UniMol [54],
Learnin
based on Benchmar
g for
protein k
Virtual
homology DUD-E
Screeni
evolutions. [28] and
ng
LIT-PCBA
[39],

Improving April 16, - DUD-E,

the Hit 2025 Protein-lig LIT-PCBA Traditional methods handcrafted scoring
Rates of and functions, docking scores, But these scores
Virtual interaction often don't correlate well with real
Screening s bioactivity → resulting in very low hit rates
by Active -
Learning Bioactivity
from feedback
Bioactivity (wet-lab)
Feedback - Molecular
similarity

MOLBIND: 3 Apr 2024 (1) Own - trains encoders for multiple modalities https
Multimodal Language dataset through contrastive learning, mapping all ://ar
Alignment via MolBind-M modalities to a shared feature space for xiv.o
of SciBERT 4, multi-modal semantic alignment rg/p
Language, to capture - df/2
Molecules, semantic 403.
and molecular 081
Proteins knowledge 67
; (2) 2D
Graphs
using a
pre-trained
GIN
encoder on
molecular
graphs; (3)
3D
 Molecular
Conforma
tions
using
UniMol
trained on
209M
structures;
and (4) 3D
Protein
Pockets
also using
UniMol,
pretrained
on protein
pockets.
Each
modality
has a
dedicated
encoder
and
projection
head,
producing
embedding
s aligned
via
contrastive
learning.
-

ESM 13 Jun - Residues Lot’s of https

All-Atom: 2024 (amino dataset Most existing protein language models ://ar
Multi-scale acids) from (like ESM) look at amino acids (residues), xiv.o
Protein - Atoms different not atoms. rg/p
Language (fine paper for df/2
Model for structure) both But to work with small molecules (like 403.
Unified - Position protein and drugs) that bind to proteins, you need 129
Molecular relationshi also for atom-level details too 95
Modeling ps fine tuing
Works with both residues (amino acids in
protein chains) and atoms (fine-grained
details).​

Learns relationships between atoms and

residues using a special position
encoding.​
 Is trained on data that includes both levels
(multi-scale).

Rethinking 4 Mar - Protein CrossDock Some https

Specificity 2024 pockets ed2020 preteaining In Structure-Based Drug Design (SBDD), ://ar
in SBDD: - Molecule dataset model AI models are now good at generating xiv.o
Leveraging structures (Francoeur used:​ molecules that get high docking scores rg/p
Delta - Binding et al., (i.e., they seem to bind strongly to df/2
Score and energies 2020) auto-regre proteins). That’s bad, because a good drug 403.
Energy-Gu - Delta ssive should bind strongly to one target and not 129
ided Score model to many others (to avoid side effects). 87
Diffusion (denoted
as AR) Delta Score:
(Luo et al.,
2021), ●​ A new metric to measure
Pocket2mo specificity.​
l (Peng et
al., 2022), ●​ It compares how well a molecule
DrugGPS binds to its target protein vs. how
(Zhang & well it binds to other random
Liu, 2023), proteins.
Targetdiff
(Guan et
al., 2023a)
and
Decompdiff
(Guan et
al., 2023b).

Multi-level 28 May Sidechain Baselines mutation effect prediction. https

Interaction 2024 changes​ We select ://ar
Modeling a broad When a mutation occurs in a protein, it xiv.o
for Protein Backbone range of can: rg/p
Mutational adjustment models as df/2
Effect s​ the 1.​ Change the sidechain of the 405.
Prediction baselines, mutated residue.​ 178
Binding including 02
affinity energy-bas 2.​ That can affect the backbone
shifts ed structure of the protein.​
methods
such as 3.​ Ultimately, it changes how strongly
FoldX [7], two proteins bind (binding affinity).​
Rosetta
[31] and
flex ddG Most existing models only look at
 [32]; sidechain-level changes, which misses
sequence- the full picture.
based
methods
such as Captures all three levels of mutation
ESM-1v effect:
[68] and
ESM2 [69];
an
unsupervis
ed method
ESM-IF
[33]; an
end-to-end
method
DDGPred
[30]; as
well as
pre-trainin
g methods
consisting
of ESM2*
[69],
RDE-Netw
ork [16]
and
DiffAffinity
[17].

FusionESP: March 4, - Protein Multimodal contrastive learning with dual https

Improved 2025 sequences encoders and projection heads ://pu
Enzyme–S (via bs.a
ubstrate language cs.or
Pair model) g/doi
Prediction - /abs/
by Fusing Molecules 10.1
Protein (via 021/
and chemical acs.j
Chemical language cim.
model) 4c02
Knowledge
357

Scaling 20 Jan - Protein BIOSNAP, https

Structure 2025 sequence BindingDB, Ultra-scalable DTI prediction using ://ar
Aware + structure DAVIS, structure-aware protein embeddings and xiv.o
Virtual tokens LIT-PCBA, vector retrieval rg/p
Screening (SaProt) DUD-E, df/2
to Billions - Molecule ENAMINE SPRINT is a vector-based virtual screening 411.
of features REAL 154
Molecules - (6.7B framework that: 18
with Co-embed molecules)
SPRINT ding space ●​ Uses structure-aware protein
language models (like SaProt)​

●​ Builds a co-embedding space for

proteins and molecules​

●​ Enables extremely fast, scalable

prediction of drug–target
interactions (DTIs)​

●​ Adds interpretability via attention

maps to show which protein
residues are important

Key Features:

●​ No need for 3D pocket prediction

(unlike DrugCLIP)​

●​ Uses attention pooling to

aggregate protein representations
(not just averaging)​

●​ Predicts top binders for the entire

human proteome in minute

DeltaDock: 16 Oct - protein structure-aware docking framework https

A Unified 2024 sidechains ://ar
Framework CPLA (Contrastive Pocket-Ligand xiv.o
for - pocket Alignment) rg/p
Accurate, flexibility. df/2
Efficient, ●​ For blind docking​ 410.
and - 3D 1122
Physically ●​ Reframes pocket prediction as 4
ligand
Reliable matching the ligand to its
structure
Molecular best-fitting pocket using contrastive
- 3D
Docking learning
protein
pocket
-
Residue- Bi-EGMN (Bi-level Energy-Guided
level and Message Network)
atom-lev
el ●​ For site-specific docking​
refineme
 nt ●​ Does coarse-to-fine refinement of
the ligand pose:​

○​ First at the residue level​

○​ Then at the atomic leve

2014 LIT-PCBA, https

Enhancing - Protein DUD-E Uni-Clip is a contrastive learning ://w
Challengin structure + framework that enhances protein-ligand ww.
g Target residue interaction modeling, particularly for biorx
Screening info challenging undruggable targets. iv.or
via g/co
Multimodal - Ligand ●​ 85% of proteins identified as nten
Protein-Lig graph + 3D clinically relevant are considered t/10.
and conformati "undruggable".​ 1101
Contrastiv on /202
e Learning ●​ These proteins have flat, flexible, 4.08
or disordered surfaces, making it .22.
hard for conventional drugs to bind.​ 609
123v
●​ Existing models often struggle to 2.full
find ligands for these targets.​ .pdf

S-MolSear 2014 ChEMBL Ligand-bas https

ch: 3D (labeled), ed virtual S-MolSearch is a semi-supervised ://pr
Semi-supe LIT-PCBA, screening contrastive learning framework that: ocee
rvised DUD-E (dense ding
Contrastiv - 3D molecu retrieval), ●​ Uses 3D molecular s.ne
e Learning - Binding af bioactive conformations and limited binding urips
for - Structural molecule affinity data.​ .cc/p
Bioactive search aper
Molecule ●​ Learns molecule embeddings from _file
Search both labeled and unlabeled s/pa
molecules.​ per/
202
●​ Employs inverse optimal 4/file
transport (IOT) to transfer /886
supervision to unlabeled data via 1183
soft labels.​ dd2
546
9d5
Dual-encoder system with IOT-driven soft 706
labeling, Uni-Mol backbone, KoLeo db9
731
 regularization aa8
6e4
9-Pa
per-
Conf
eren
ce.p
df

SOAPI: 2025 Custom - Protein Binder

SIAMESE- curated sequences ranking designing binders that selectively
GUIDED binder–tar (binder, (top-N), engage a target protein while
GENERATI get–off-tar target, 2 conditional minimizing interactions with structurally
ON OF get off-targets) binder or functionally similar proteins remains
OFFTARG quadruplet - generation, a major challenge.
ET-AVOIDI s (details Embeddin specificity
NG in g validation
PROTEIN supplemen distances
INTERACT tary); 125 - Diffusion SOAPI leverages a Siamese protein
IONS test targets language language model with an adaptive
modeling Log-Sum-Exp Decoy Loss to enforce
specificity by embedding
fusion-specific binders close to their
target while maintaining separation
from offtargets. These optimized
embeddings then guide a diffusion
protein language model (DPLM), which
generates binders using
soft-value-based decoding (SVDD) and
Sequential Monte Carlo resampling to
iteratively refine candidates.

DrugHash: 2025-04-11 ZINC20, - Protein https

Hashing Enamine 3D Uses hash indexes to furtur improve ://ojs
REAL structure constractive learnning efficiency .aaai
Based (6.75B - Molecule .org/
Contrastiv molecules) 3D Massive Efficiency Gains: inde
e Learning ; conformati x.ph
pretraining on ●​ 32× less memory than DrugCLIP​ p/AA
for Virtual
on 3.2M - SE(3) AI/ar
Screening protein Transform ●​ 4.6× faster retrieval on ticle/
pockets er billion-scale molecule sets​ view
and 19M encodings /338
molecules ●​ Outperforms state-of-the-art 73/3
accuracy-wise, despite using 602
binary codes 8
 2025 BindingDB https
Bioptic -- A (efficacy), Pretraining: 160M molecules (PubChem ://ar
SMILES Speed + ab
Target-Agn Enamine + Enamine REAL)​ xiv.o
(molecul structure
REAL ar rg/p
ostic Space df/2
structure
Potency-Ba (scale) Supervised Fine-tuning: Uses 406.
only)
sed Small BindingDB bioactivity data​ 145
72
Molecules
Search Final Embedding Dim: 60D​
Engine
Optimizer: LARS (300 epochs)

Pharmaco 2025 https

Match: massive runtime improvement for very ://op
large-scale libraries enre
Efficient view
3D .net/
Pharmaco pdf?i
d=2
phore
Fast 3D pharmacophore screening via 7Qk
Screening neural subgraph matching 18IZ
via Neural um
Subgraph
Matching Replaces traditional alignment-based
screening with learned embedding
matching

Flow-Base 2025 Fragment 3D protein https

d screening surface This paper presents a novel ://op
data; geometry, protein-fragment encoder trained using enre
Fragment PDB-deriv molecular contrastive learning to create a shared view
Identificati ed fragments latent space between protein binding .net/
on via surfaces pockets and molecular fragments. foru
m?id
Contrastiv representing proteins as surface point =bZ
e Learning clouds and ligands as coarse fragment W1
of Binding graphs. HLT
1gI
Site-Specifi
c Latent
Representa
tions

MIN: 10 Public DTI sequence Combines https

Multi-Chan March datasets and 3 DTI prediction sota?​ ://iee
structure channels: deep learning framework for drug-target expl
nel 2025 views for 1. interaction (DTI) prediction. It uses a ore.i
Interaction accurate Structure- multi-channel strategy to combine diverse eee.
Network DTI agnostic views of protein and drug representations org/
modeling. 2. and introduces a novel C-Score Predictor abst
for to highlight critical residues for interaction.
Structure- ract/
Drug-Targe aware docu
t 3. men
Extended- t/10
Interaction C-Score residues show high overlap with
mixture 918
With true binding sites, improving explainability 833
Protein
Distillation

Escaping 2024 BindingDB, Protein https

the BioSNAP, seq (LLM), improve interpretability and generalization ://w
Human; Drug graph in Drug-Target Interaction (DTI) prediction. ww.
drug-bias 1,862 real (GCN); It directly addresses the “drug-bias trap,” biorx
trap: using complexes cross-atten where models overfit to the drug iv.or
debiasing from PDB tion fusion representation at the expense of learning g/co
meaningful protein interactions. The nten
design to architecture introduces unbalanced dual t/10.
improve branches, cross-attention aggregation, and 1101
interpretab Maximum Classifier Discrepancy (MCD) for /202
cross-domain robustness. 4.09
ility and
.12.
generalizat 612
ion of 771v
drug-target 1.full
.pdf
interaction
prediction

NaFM: 18 May Large-scal Scaffold-fo https

Pre-trainin 2025 e NP cused ://ar
databases contrastive xiv.o
ga incl. + masked rg/p
Foundation taxonomy graph df/2
Model for & BGCs learning 503.
Small-Mole Scaffold-focused contrastive learning + mas 176
56
cule
Natural
NaFM is a specialized molecular
Products
foundation model pre-trained on natural
products, addressing the limitations of
existing supervised and general-purpose
pre-trained models in capturing the
complex structures, scaffolds, and
biosynthetic logic of these molecules.
Unlike traditional SMILES- or GNN-based
models, NaFM uses a hierarchical
scaffold-centric approach with
contrastive learning and masked graph
modeling, making it tailored for capturing
both the evolutionary and structural
patterns in natural product (NP) space.

A 6 Mar 2025 Protein-lig Mixture-of- https

Generalist and Domain-Ex BIT is a generalist, cross-domain ://ar
complexes perts Transformer-based foundation model for xiv.o
Cross-Dom , small (MoDE) + structure-based drug discovery (SBDD). It rg/p
ain molecules, Mixture-of- jointly learns from small molecules, df/2
Molecular proteins Structure- proteins, and protein-ligand complexes 503.
(2D & 3D) Experts across 2D and 3D modalities. It 043
Learning incorporates two novel architectures:
(MoSE) 62
Framework Mixture-of-Domain-Experts (MoDE) for
for biochemical diversity and
Mixture-of-Structure-Experts (MoSE) for
Structure-B
geometric positional modeling. This makes
ased Drug it highly adaptable across multiple
Discovery biochemical tasks.

MultiT2: A January 27, Genomics, Bacterial https

Tool 2025 proteomics T2PK-relat Demonstrates first successful multimodal ://pu
, ed fusion in NP discovery; restructured NP bs.a
Connectin metabolom multimodal knowledge graph cs.or
g the ics, datasets g/doi
Multimodal spectrosco (genomic, first multimodal AI algorithm designed to /full/
py, enzyme proteomic, integrate fragmented and diverse data 10.1
Data for sequences etc.) types related to bacterial aromatic 021/
Bacterial , BGCs polyketides (T2PKs)—a critical class of acso
(biosynthet medically relevant natural products. meg
Aromatic ic gene a.4c
Polyketide clusters) 1126
Natural 6
Products

A 23 May 3 public Dual-chan SOTA https

Cross-Field 2024 DTI nel (DTI + accuracy SiamDTI introduces a siamese ://ar
benchmark PPI); for novel double-channel neural network for xiv.o
Fusion s (with protein-lev drugs/targ predicting drug–target interactions (DTIs) rg/p
Strategy known and el global + ets; by fusing local (binding pockets, active df/2
for novel local fusion ~85.7% sites) and global (3D shape, tertiary 405.
DTIs) accuracy structure) protein information. This 145
Drug–Targ cross-field fusion from both DTI and PPI
vs ~62.3% 45
et for perspectives significantly improves the
Interaction baselines ability to predict novel DTIs, a key
on novel challenge in real-world drug discovery.
Prediction
DTIs
Outperforms prior methods by large margin
on AUROC for novel DTI prediction

PHARMAC 14 Mar Virtual 3D ~10x https

OMATCH: 2025 screening pharmacop speedup; PharmacoMatch redefines 3D ://ar
sets hores, zero-shot pharmacophore screening as a neural xiv.o
EFFICIENT (ligand-que neural pharmacop subgraph matching task using a rg/p
3D ry) subgraph hore contrastive-learning GNN and order df/2
PHARMAC (GNN) matching; embeddings. It significantly accelerates 409.
efficient screening across billion-scale chemical 063
OPHORE libraries by avoiding costly alignment steps.
billion-scal 16
SCREENIN e library Trained self-supervised, it enables fast,
G VIA screening scalable, and zero-shot pharmacophore
matching via precomputed embeddings.
NEURAL
SUBGRAP
H
MATCHIN
G

3DMOLFO 7 Feb 2025 Mixed data Dual: https

RMER: A (PDBbind Discrete dual-channel transformer framework ://ar
+ tokens designed for both protein-ligand docking xiv.o
DUAL-CHA pocket-liga (atoms) + (predictive) and pocket-aware 3D drug rg/p
NNEL nd + 3D design (generative). It introduces a df/2
FRAMEWO synthetic coordinate parallel sequence format that jointly 502.
ligands) s models discrete atom/protein tokens and 051
RK FOR their 3D coordinates, overcoming the
STRUCTUR long-standing challenge of integrating 07
E-BASED sequence and structural modalities.
DRUG
DISCOVER
Y

MoleculeC 13 Jun * Possible https

LA: 2024 benchmark MoleculeCLA is a large-scale, ://ar
for computation-based molecular xiv.o
Rethinking molecules benchmark specifically designed to rg/p
Molecular overcome the known flaws of existing df/2
Benchmark benchmarks like MoleculeNet. It contains 406.
~140,000 molecules evaluated across 9 177
via molecular properties related to 97
Computati ligand-target interactions using
onal computational docking (Glide) instead of
wet-lab experiments. This allows for
Ligand-Tar
high-throughput, scalable, noise-free
get Binding benchmarking with strong
Analysis physicochemical and biological
interpretability.

The benchmark spans 10 protein targets

and evaluates models across traditional
and deep learning-based molecular
representation learning (MRL) pipelines.
MoleculeCLA improves data balance,
label quality, and interpretability, making
it ideal for drug-target interaction tasks
and general molecular property learning.

Scaling 2025 Jan LIT-PCBA, Protein tokens, https

Structure 20 BindingDB; sequence molecule scalable, fast, and interpretable virtual ://p
screens + structural embedding screening framework designed for mc.n
Aware 6.7B mols tokens, s​ drug-target interaction (DTI) prediction. It cbi.n
Virtual across full molecule SOTA on introduces a vector-based co-embedding lm.ni
Screening proteomes embedding DTI, approach using structure-aware protein h.go
s LIT-PCBA, language models (PLMs) and v/arti
to Billions self-attention architectures. Unlike
binding cles/
of affinity; docking, SPRINT performs DTI prediction PM
Molecules ultra-fast via dot products in embedding space, C11
retrieval; making it billions of times faster while 838
with
interpretabl retaining high accuracy. 698/
SPRINT e via
attention
SCORCH2: April 16, DUD-E, Docked https
a 2025 DEKOIS poses, Strong hit a next-generation machine learning ://w
2.0, ligand-rece enrichment framework for high-enrichment virtual ww.
generalise VSDS-vd; ptor ; screening, aimed at improving early-stage biorx
d no interaction interpretabl hit identification in drug discovery. iv.or
heterogene crystal-qua s, e via SCORCH2 advances beyond its g/co
lity descriptors SHAP; predecessor (SCORCH) by enhancing both nten
ous robust on generalization and interpretability using a
structural t/10.
consensus dependenc novel consensus-based ensemble of 1101
model for y targets interaction-focused models. /202
5.03
high-enrich
.31.
ment 646
interaction- 332v
based 2.full
.pdf
virtual
screening

GenSPARC 07 May, CPI AlphaFold Strong https

: 2025 datasets 2-derived generalizat a deep learning framework for predicting ://w
with protein ion; compound-protein interactions (CPI) that ww.r
Generalize structure-in structure, validated overcomes the limitations of sequence-only esea
d formed Foldseek on multiple models and data scarcity. It incorporates rchs
Structure- features 3D CPI structure-aware protein embeddings quar
from alphabet, benchmark from AlphaFold2 and Foldseek’s 3D e.co
and interaction alphabet, and combines these
AlphaFold GCN s; virtual m/ar
Property-A 2 and compound screening with graph neural network (GCN)-based ticle/
ware Foldseek encoding, compound representations and pretrained rs-6
attention chemical language models. 326
Representa
235/
tions of v1
Language
Models for
Compound
-Protein
Interaction
Prediction

MOLCOMA 2025 MoleculeN Outperfor https

: et, QM9 ms SOTA MolCoMA is a multi-modal pretraining ://op
Atom-lev
on both framework designed to create enre
COMPLEM el fusion
2D/3D & fine-grained molecular representations view
of 2D
ENTARY multi-moda by fusing 2D topological and 3D .net/
MASKING topologic l tasks; geometric information. Unlike earlier pdf?i
STRATEGY al tested on methods that align or generate across d=y
graphs bio and modalities in a coarse or isolated fashion, 7Ud
FOR quantum MolCoMA introduces a complementary 3RA
and 3D
PROMOTI datasets masking strategy combined with PT8
molecula
NG r cross-modal reconstruction to capture
geometr nuanced atomic-level relationships
ATOM-LEV
y via between modalities.
EL
complem
MULTI-MO entary
DAL masking
MOLECUL
AR
REPRESEN
TATION

Geometric graph nN, Unlike generic graphs, geometric graphs often exhibit
physical symmetries of translations, rotations, and reflections, making
them ineffectively processed by current Graph Neural Networks (GNNs).
https://link.springer.com/article/10.1007/s11704-025-41426-w