How could we use Artificial intelligence in drug discovery?

Drug discovery is essentially an optimization problem under uncertainty. The goal is to find a molecule that interacts with a biological target strongly enough to treat a disease, while also being safe, stable, manufacturable, and affordable. AI is becoming powerful because almost every stage of this pipeline involves searching through enormous spaces of possibilities.

Here's the complete process and where AI fits.

The Drug Discovery Pipeline
Disease
↓
Identify biological target
↓
Validate target
↓
Find molecules that interact with target
↓
Optimize molecules
↓
Animal testing
↓
Human clinical trials
↓
FDA/EMA approval

Only about 1 in 5,000–10,000 initial compounds eventually becomes an approved drug, and the process typically takes 10–15 years.

Step 1: Understand the Disease

Scientists first ask:

What causes the disease?
Which genes are involved?
Which proteins malfunction?
Which signaling pathways break?

Example:

Parkinson's disease
Alzheimer's disease
Cancer
Diabetes

Large datasets include

Genome sequencing
RNA sequencing
Proteomics
Clinical records
Medical imaging
AI applications

Machine learning can discover

disease subtypes
hidden biomarkers
causal genes
regulatory pathways

Models:

Graph Neural Networks
Transformers
Causal inference
Foundation models for biology

Example:

Instead of studying one gene at a time,

AI analyzes millions of interactions simultaneously.

Step 2: Target Identification

A drug usually doesn't cure a disease directly.

It changes the behavior of a biological target.

Targets include

proteins
enzymes
receptors
ion channels
RNA
DNA

Examples

EGFR
HER2
KRAS
GPCRs
kinases

Question:

Which protein should we attack?

AI applications

AI predicts

protein-protein interactions
disease networks
essential genes
druggable proteins

This becomes a graph prediction problem.

Step 3: Target Validation

Even if a protein looks important,

blocking it might do nothing.

Scientists test

CRISPR knockouts
RNA interference
animal models
patient genetics

Question:

Does changing this protein actually improve disease?

AI applications

Models combine

genetic evidence
literature
experimental data

to estimate

Probability(target causes disease)

This saves years of experiments.

Step 4: Structure Determination

Now scientists need to know

"What does the target look like?"

Historically:

X-ray crystallography
Cryo-EM
NMR

Today:

AI predicts structures.

Examples

AlphaFold
RoseTTAFold

For many proteins this dramatically speeds up research.

For intrinsically disordered proteins, however, the challenge shifts from predicting a single structure to predicting a conformational ensemble, as discussed earlier.

Step 5: Hit Discovery

Now comes the search problem.

Scientists ask:

"What molecules bind this protein?"

Chemical space is unimaginably large.

Estimated possibilities:

10^60

to

10^100

possible drug-like molecules

We can only synthesize a tiny fraction.

Traditional methods

high-throughput screening
fragment screening
natural products
AI applications

This is where AI shines.

Virtual screening

Predict

Protein + Molecule

↓

Binding affinity

without performing experiments.

Models

Graph Neural Networks
Diffusion models
Equivariant neural networks

can evaluate millions of molecules in hours.

Molecular docking

Instead of physically testing compounds,

AI predicts

Protein

+

Molecule

↓

Binding pose
Generative AI

Instead of searching molecules,

generate them.

Examples

diffusion models
reinforcement learning
graph generation
language models for molecules

Input

"I want a molecule that binds KRAS."

Output

Thousands of novel candidates.

Step 6: Hit-to-Lead Optimization

Now scientists improve promising molecules.

Goals:

Increase

potency

Decrease

toxicity

Improve

solubility
stability
selectivity
metabolism

This is an enormous multi-objective optimization problem.

Medicinal chemists often make hundreds of small changes to a molecule.

AI applications

Predict

If I replace this carbon atom,

will potency improve?

Models predict

IC50
EC50
toxicity
metabolism
permeability
synthetic accessibility

This dramatically reduces trial-and-error synthesis.

Step 7: ADMET Prediction

Most drug candidates fail here.

ADMET means

Absorption
Distribution
Metabolism
Excretion
Toxicity

Example:

A molecule may bind perfectly but

never reach the brain
damage the liver
be rapidly broken down
cause heart toxicity
AI applications

Predict

liver toxicity
hERG inhibition
blood-brain barrier penetration
oral bioavailability
CYP interactions
plasma protein binding

before any animal studies.

Step 8: Preclinical Testing

Now real experiments begin.

Scientists test

cells
organoids
mice
rats
monkeys

Questions

Does it work?

Is it safe?

What dose?

AI applications

Analyze

pathology slides
imaging
behavioral data
omics data

to extract patterns that humans might miss.

Step 9: Clinical Trials

Human testing proceeds in phases:

Phase I

Healthy volunteers

Safety

Phase II

Small patient group

Does it work?

Phase III

Thousands of patients

Compare with existing treatments.

AI applications

AI can improve:

patient recruitment
trial design
endpoint prediction
adverse event detection
adaptive trial monitoring
analysis of electronic health records
Step 10: Regulatory Approval

Regulators review

efficacy
safety
manufacturing quality
risk-benefit balance

Even after approval,

monitoring continues for rare side effects.

Where AI Can Automate the Pipeline
Stage AI contribution
Disease understanding Biomarker discovery, patient stratification
Target identification Predict causal proteins and pathways
Protein structure Predict folded structures or conformational ensembles
Hit discovery Virtual screening and molecular docking
Molecule generation Design entirely new compounds
Lead optimization Suggest chemical modifications
ADMET Predict pharmacokinetics and toxicity
Preclinical Analyze experimental data
Clinical Optimize trial design and recruitment
Post-market Detect adverse events from real-world data
What AI Still Cannot Fully Replace

Despite impressive progress, AI cannot yet eliminate the need for experiments because biology is extraordinarily complex. Models still struggle with:

Protein dynamics, especially intrinsically disordered proteins
Cellular context (the same protein behaves differently in different tissues)
Off-target interactions
Long-term toxicity
Immune responses
Human metabolism
Rare adverse effects
Clinical efficacy in diverse patient populations

As a result, experimental validation remains essential.

The Long-Term Vision: A Closed-Loop AI Drug Discovery System

The most ambitious vision is not an AI that simply predicts molecules, but an autonomous research loop:

Patient Data
↓
Disease Model
↓
Target Selection
↓
Protein Structure / Ensemble Prediction
↓
Generate Candidate Molecules
↓
Predict Binding + ADMET
↓
Select Best Candidates
↓
Robotic Synthesis
↓
Automated Biological Assays
↓
Experimental Results
↓
Retrain the AI
↓
Repeat

This "design–make–test–learn" cycle allows each round of experiments to improve the model, gradually converging on better drug candidates with far fewer laboratory iterations than traditional workflows. Rather than replacing scientists, AI serves as a powerful decision-making and optimization engine, enabling researchers to explore vastly larger regions of chemical and biological space than would otherwise be possible.