1. Impact of Big Data &
Artificial Intelligence in
Drug Discovery &
Development
Nick Brown
Executive Director, Imaging & Data Analytics
Clinical Pharmacology & Safety Sciences, R&D, AstraZeneca
CONFERFENCE
European Drug Discovery Innovation & Outsourcing Programme, 12th September 2023

2. Drug development process is incredibly costly
2
From drug discovery through FDA
approval, developing a new
medicine:
• at least 10 years
• Costs an average of $2.6 billion
Less than 12% of the candidate
drugs that make it into Phase I
clinical trials will be approved by
the FDA.

3. >4-fold improvement in success rates since 2012
71
66
71
88
51
59
46
72
22
15
18
43
59
60
66 70
6 4 4
19
100
0
Percentage
Preclinical Phase I Phase II Phase III Overall
2005-2010 (industry)
2005-2010 (AZ)
2013-2015 (industry)
2012-2016 (AZ)
3
Making sure our
molecule gets to the
right tissue where it
is needed
Ensuring right
safety with minimal
side effects
Selecting the
right patients
that will benefit
Defining the right
commercial value
and future viability
Identifying the
Right Target

4. Early stage Late stage
Right Dose
Clinical
Candidate
Clinical pharmacology,
pharmacometrics,
clinical bioanalysis,
regulatory filings
New indications, global
roll out of submissions,
supplemental NDAs,
product license
maintenance, impurity
assessment
Post marketing
GLP toxicology,
pathology,
bioanalysis, clinical
pharmacology,
pharmacometrics
Right Safety
TSID to LOID
Target safety
assessment,
mechanistic safety,
DMPK, ADME
Lead optimization
to CDID
Chem Tox, ADME,
DMPK, investigative
safety, pathology, CDID
safety package
4
involved end to end in drug discovery and development
Clinical Pharmacology & Safety Sciences
Embedding the right digital solutions (imaging, data and AI) to improve efficiency

5. 3. Data
Backbone
4. AI-based
Capabilities
2. Core
Systems
Machine
learning
Virtual
reality
Chat Bots
Deep learning
Angel
D360
Data Lake Data
warehouse
NLP
NGS
platforms Data
marts
Knowledge
graph
Literature
Genomic
ADME
in vivo
MedChem
Toxicology
Clinical trial EHR
1. Data
Sources
Imaging
External
federated
data
Foundational Big Data for CPSS

6. 6
Examples of AI
within CPSS
Imaging &
Data Analytics
• Al is redefining safety assessment
• AI is radically changing pathology
• AI is aiding clinical pharmacology

7. From Common Approaches To Early Safety Assessment
• Compounds of interest are tested against a
wide range of assays
• Emphasis on Cardiovascular, Hepatic & CNS
(Secondary Pharmacology)
• High quality compounds will be inactive, or
have good selectivity in the in vitro safety
assays
• Compounds with good selectivity have an
increased chance of having large safety
margins (Therapeutic Index) in vivo, or in
the clinic.
7
Target Organ Assay
Cardiovascular
hERG
NaV1.5
Iks
Kv4.3
L-type calcium channel
Cardiomyocyte
Structural Cardiovascular Tox
Hepatic
Glu/Gal Mitochondrial Assay
High Content Mitotox assay
Cytotoxicity
Hepatic Spheroid
Liver Transporters (BSEP & MRP2)
Genetic Toxicity
AMES Mutagenicity Test
In vitro Micronucleus
Various
Secondary Pharmacology Panels
Phospholipidosis
AhR (CYP1a1)

8. To Computational ML for Right Safety
What do we have today?
• >50 ML models deployed in
production and incorporated in the
DMTA cycle
• Physicochemical properties, ADME &
safety
• Greater than 4 million molecule
calculations per day
• 141 million molecules processed in
one month
8
Combining individual models to create virtual organ safety models
(such as DILI for liver or CV risk for heart) for more accurate
prediction using end point assays.

9. Building virtual in-silico screens using AI
Could we make a “digital twin” for therapeutic margins?
Build it for a single organ; expanding to other organs
We can now predict human liver injury using in vitro
or in vivo data using AEGIS (AI enabled genomics in
safety)
AEGIS: Predicts dose-dependent human
DILI assessments from in vitro data
AEGIS DILI probability
0.0 0.2 0.4 0.6 0.8 1.0
Low, Mid, High
Dose: Low, Mid, High
Alpidem
Zolpidem
AZ proprietary
genes
AI/ML
Probability of
liver injury
Neighbor drugs
Active gene
network?
AEGIS shows a dose
dependent, higher
DILI risk for Alpidem,
a drug that was
removed from
market due to DILI

10. Computer vision applications assist pathologist
10
• Specimen slides are scanned by a high-resolution scanner
• Very large slide images (WSIs) difficult to manually assess
• Using AI for fast, accurate image analysis & quantification
• Tasks include segmentation, classification, regression, clustering
Scanners
digitise the
samples

11. 11
Human assessment
Tumour cell –ve +ve Immune cell -ve +ve
+++
years
20min
10-20%
+++
days
seconds
0.65%
AI-based assessment
Complexity
Training
Time
Error rate
Complexity
Training
Time
Error rate
:
:
:
:
:
:
:
:
AI drives Digital Pathology
In focus Out-of-focus
Intermediate
Out-of-focus
Initial Scan Image QC Results
Providing quantitative data readouts upfront to
support Pathologist decision making
AutoQC applied to >40k images accelerating
delivery for pathology review

12. From Digital to Virtual Pathology
12
In the future, AI will predict patient response by
combining multi-modal imaging with patient
genomics, proteomics & spatial transcriptomics
AI model training & development will ultimately enable clinical prediction tools
Today digital pathology is a multi-step process involving IHC staining & image
analysis. We developed AI pipeline for virtual IHC that avoids the need for manual
annotation leading to faster, quantitative imaging and virtual pathology diagnosis
è AI recapitulates image but using cheaper/faster H&E of tissue samples & biopsies

13. 13
Therapeutic index is often uncertain at candidate nomination
Decreasing a safe dose is
easier than increasing it
• Efficacy = f (potency, exposure)
• Safety = f (hazard, exposure)
• Safety and efficacy of a drug is
fixed by its dose
• The dose gives a specific time
vs concentration curve
• This is solely a property of the
molecule selected
Muller & Milton (2012). Nat Rev Drug Discovery
Dose makes the poison
We use AI to predict right safety and right patient dose

14. AI Predicting The Right Patient Dose
14
• Free scientists from manual
model building
• Speed up popPK model
development
• Identify better, more stable
models
AutoPK
Pharmacometrician AutoPopPK
Automated
structure
discovery tool
Biological
understanding &
insights
High
performance
popPK models
Case study: Tested using clinical trial data from 2 phase 1 trials (184 patients)
èmatched the structural expert model and/or improved the residual error
èAI model search took <40hrs total, saving 2-3 weeks
èSaves almost 3 years in computational time annually if we had a virtual pharmacometrician
eg AutoPopPK: Discover population pharmacokinetic model structure automatically

15. AI Impacting Patient Trials
15
Trials (Product) Population
Total N
(thousands)
Diabetes N
(thousands)
PLATO (BRILINTA) ACS 18.6 4.6
PEGASUS (BRILINTA) Prior MI 21.1 6.8
SOCRATES (BRILINTA) Stroke/TIA 13.2 3.2
EUCLID (BRILINTA) PAD 13.9 5.3
CORONA (CRESTOR) Heart failure 5.0 1.5
JUPITER (CRESTOR)
Increased
CRP
17.8 0.0
AURORA (CRESTOR)
Hemodialysis/
ESRD
2.8 0.7
SAVOR (ONGLYZA)
Type 2
diabetes
16.5 16.5
EXSCEL (BYDUREON)
Type 2
diabetes
14.0 14.0
CHARM (ATACAND) Heart failure 7.6 2.2
DECLARE (FARXIGA)
Type 2
diabetes
17.2 17.2
THEMIS (BRILINTA)
Type 2
diabetes
17.0 17.0
STRENGTH
(EPANOVA)
High TG 13.0 9.0
DAPA-HF (FARXIGA) HFrEF 4.7 2.1
DAPA-CKD (FARXIGA) CKD 4.3 2.9
Future total 186.7 103
Recurrent Neural Network
Feedforward Neural Network
Case study: Tested using internal patient level data including >100k type 2 diabetes
èusing longitudinal data improved cardiovascular risk prediction by 20% compared to single visit
èFacilitate the design of future CV outcomes trials
èAid HCPs to identify earlier and more accurately patients that would benefit from AZ treatment
Optimised survival ML
models incorporate changes
in risk factors over time
eg Machine Learning For Cardiovascular Risk Prediction

16. 16
AI & Big Data enable us to quantify therapeutic index
New AI approaches allow us to
study the effects of molecules
on biological systems better
than ever before
Multi-omics & imaging
technologies generate
enormous data volumes,
need AI to support analytics
Machine learning and AI are
increasing productivity
reducing how & what our
scientists review

17. 17
What’s next for R&D Discovery & Development
Radically changing how we
create documentation through
generative AI, accessing the
right information at the right
time for the right audience
From manual assessing
patient scans to AI assistants
that automatically annotate
and track the location or
response clinically
Using big data and machine
learning to create
recommendations querying
millions of scientific records
in seconds, not months

18. 18
Questions & Answers
https://careers.astrazeneca.com/data-science-and-ai
• Richard Goodwin
• Nigel Greene
• Richard Dearden
• Jim Weatherall
• Anna Asberg
• David Greatrex
• Derek Marren
• Bino John
• Megan Gibbs
• Vignesh Subramanian
• Nikolay Burlutskiy
• Arthur Lewis
• Lars Lynne Hansen
• Pete Wolstencroft
• Fruzsina Soltesz
• Jennifer Tan
• Peter Newham
• Andrzej Nowojewski
• Richard Haworth
• Olga Obrezanova
• Scott Hoffman
• Heather Hulme
• Stefan Platz
and many more…
Acknowledgements
There’s no better time to be in science