Navigating the Transition:

From Models to Deployed AI

Swetabh Pathak
CTO & Co-Founder, Elucidata
Emerging Leaders for AI in Life Sciences R&D
..with Differentiated Technology & Expertise

#1 30+ customer proof points across

discovery, development and trials

#2 Cloud-first data platform and ML that can

seamlessly complement your ecosystem.

#3 With a team that can quickly assimilate

domain expertise into high impact services.
AI Shows Promise, But Opinions Vary..

AI is currently only
It's too early to used for solving AI is a new hype -
AI is the future - it Al is somewhat
recognize the true simple investors buy into
will help us explore valuable.
impact of Al - we will problems. The the
areas that have never In our work, AI has
only be able to see InSilico desire to be hip, cash
been explored before. helped make a lot of
the true impact once screen would only is raised, Pharma cos
One day AI will help molecules
we can see the have do deals to be in the
us understand biology synthesizable faster &
productivity over had a 4% failure rate, news. There's lots of
so deeply that we can cheaper.
time. even without AI. noise in this field but
form new scientific
it has not been
laws and drug design
proven yet.
principles.
Computational
Head of Data and Translational Deputy Director,
Biologist, Chief Executive,
Platform Strategy, Scientist, Academia Global Health, Non-
Research Institute Data Consortium
‘AI-first’ biotech profit Organisation

Innovators Early Adopters Early Majority Late Majority Laggards

(2.5%) (13.5%) (34%) (34%) (16.5%)

Source:BCG Report on AI, 2023

Biggest Barriers to Broader AI Adoption

Limiting Belief 1 Limiting Belief 2 Limiting Belief 3

‘We don’t have ‘We just couldn’t ‘We don’t have

a relevant use get enough the expertise or
case for AI’ training data to the millions of
solve the dollars needed
problem we’re to build a team’
working on’

Translational Leader, Computational CSO, Early Stage

Mid-Stage Biologist, Research Therapeutics
Pharmaceutical Institute
Company
 Limiting Belief 1

“We don’t have a relevant use

case for AI”
Good AI use cases are not rare events

Access to ‘high quality and relevant’ data

Human supervision is available & possible

What do good
use-cases look Biological rules can be framed

like?
Hypothesis generation rather than testing

Explainability is not necessary

Right use case is the biggest predictor of success

‘An Early Stage Therapeutics company wanted to develop and train Classifier Models
to segment patients in AML’

2+ Targets
Identified
and
Validated

10k Public Multi- Patient Prioritized list of

Segmented Cohorts
omics samples Stratification Model gene targets

AI was used to assist domain experts, to solve a well-defined problem with

clear outcomes: ‘Segment the patient cohorts in AML as per their prognosis’
LLMs were Deployed to Advance Target Identification

8
 Limiting Belief 2

“We just couldn’t get enough

training data to solve the
problem we’re working on”
Make the Shift to Data-Centric AI

Fine-tune existing models with high quality and relevant data. Especially useful for predicting long-
tail problems with limited data points (<10,000)

Use Cases

Cell Type
Annotation
Pre-training Data

Available Single Cell Fine-tuned with task Dataset Integration

Gene Expression Fed into Foundational
specific, high quality
profiles (33 million) Model
data Gene Perturbation
Response
Prediction

Gene Regulatory
Network Inference
How do we define High Quality Data?

● Up-to-date with patient information, domain specific

● Metadata annotations, processing & ontologies are custom to
Highest
the use case / task
Quality Data
● QC-ed by experts

● Annotated with critical metadata

● Relevant to the biological domain
Context Specific,
Relevant (scGPT,
GenePT)

● Ingested & transformed into Machine readable

formats Machine Curated from Public Sources
● Structured into tabular files (CSV, JSON) (BERT, GPT)

11
How well does scGPT perform after Fine-Tuning with High Quality Data?

scGPT can perform reference based cell type annotation in a zero shot setting.
However, fine-tuning with high quality data improves model performance by 20% (avg)

Experimental Design

● Training Dataset: 25k immune

cells from HCA to fine-tune

● Testing Dataset: 13k immune

cells from Tabula Sapiens

● All datasets were cleaned and

linked with Elucidata’s
Harmonization Engine.
Elucidata’s Harmonization Engine
Cleans the Data you Need

50 Million
10X FASTER Samples harmonized to support use cases
curation with in drug discovery, development & trials
LLM powered
annotation tools Techn
ology

25+ Data Types

Supported including RWE, Omics and
Clinical
Proces
People
s

100+ Experts
30+
99.99% Accurate
In curation, NLP, data
Data delivered with Data Pipelines built and maintained on
engineering, & Elucidata infrastructure to process data
robust QA/QC
bioinformatics
 Limiting Belief 3

“We don’t have the expertise

or millions needed to build a
team”
Case Study: Building Production-ready AMDET Pipelines

SCENARIO

‘This mid-cap pharmaceutical company wanted to develop an end-to-end ADMET prediction pipeline
that would support 5 lead development programs across neurology and oncology’.

THEIR NEEDS
● Collect and prepare all the data generated across assays in a meaningful and scalable way.
● Productionize existing models on the cloud, so that they can run at scale.
● Develop an ML-ops infrastructure to manage the data & models across stakeholders, multiple
sites and different types of users.

Doing this in-house needs a team of 7 FTEs and cloud resources.

Costs drum upto ~$2 Million and projects could take 1+ years to kick start.
Scaling up AI in production, Set up in 1.5 Months

Compound Screening & Evaluation Accelerated by 2X with Production-Ready ADMET pipeline

Polly

Load & Split Data, Feature Selection

Train & Fine-Tune Models

Track Experiments, Select Models,

Version

Relevant Public Harmonization Validated Model to Dashboard

and In-house Predict endpoints Predicted Endpoints
Engine
Datasets Ingest, process, annotate

Add inputs & Run Workflow

ADMET Model Deployment Workflow

Significant R&D Productivity Unlocked

Projects could be kickstarted within a month, at 4X Lower Costs.

Productivity Areas Improvement with Elucidata Rationale

● Dedicated team to perform searches

Data Acquisition 4X Faster
in Public Databases

Data Preparation, ● LLM-powered Harmonization reduced

4X Faster
Annotation, QC manual effort in data preparation.

● Key bottleneck steps in the process

Model Development /
Reduced by 30% (ingestion, cleaning, ML Model
Deployment Cycle
Deployment, Versioning) automated.
Any Questions?

Reach out at elucidata.io to know more!

"The value is in the data, it is not in
the tools. That is the one thing, it’s a
bit of a hobby horse for me. One
Polly Platform &
Services thing always point to in these
discussions around data, don’t
underestimate the amount of time
ML-ready and value in doing what is really
Data Access often difficult and not so rewarding
directly work, like cleaning data sets
isn’t always fun, but it is often the
most valuable thing you can do."

ML
Initiatives Dr. Jeffrey Reid,
Regeneron's Chief Data Officer