Maryam Taherzadeh

Computational Scientist | AI/ML for Drug Discovery

AI-DRIVEN DRUG DISCOVERY

MDM2–p53 Inhibitor Design

Interactive structural comparison of Nutlin-3a and generated MDM2 inhibitor candidates. The MDM2 receptor and Nutlin-3a reference remain fixed while selected candidates are overlaid in the binding pocket.

MDM2–Nutlin-3a Complex

PDB 5ZXF · crystallographic reference structure
Loading MDM2 structure...
Nutlin-3a Selected candidate Ligand-proximal residues (≤4.5 Å)
ITERATIVE MOLECULAR OPTIMIZATION

From Generation 1 to Generation 3

Iterative reward redesign progressively shifted the generated molecular population toward improved predicted potency, drug-like properties, and multi-objective constraint satisfaction.

01
GENERATION 1

Potency-driven exploration

Initial reinforcement-learning generation prioritized predicted MDM2 potency while maintaining molecular similarity and physicochemical constraints.

870 unique molecules
Median predicted pIC50 6.496
Shared-criteria pass 0.57%
Median QED 0.372
Median cLogP 6.074
Reward redesign
→
02
GENERATION 2

ADMET-aware optimization

The reward function was redesigned to incorporate predicted bioavailability and major ADMET liabilities alongside potency and molecular quality.

3,371 unique molecules
Median predicted pIC50 6.805
Shared-criteria pass 0.86%
Median QED 0.385
Median cLogP 5.904
Multi-objective refinement
→
03
GENERATION 3

Refined molecular population

Further reward refinement produced a population with higher predicted potency, improved QED and cLogP distributions, and greater satisfaction of the shared design criteria.

1,493 unique molecules
Median predicted pIC50 7.053
Shared-criteria pass 36.77%
Median QED 0.437
Median cLogP 5.447
↑
KEY POPULATION-LEVEL RESULT

Shared-criteria pass rate increased from 0.57% to 36.77%

Across three iterative generations, reward redesign shifted the generated population toward substantially greater simultaneous satisfaction of the predefined multi-objective design criteria.

POPULATION COMPARISON

Generation-level metrics

Summary statistics across unique generated molecules.

Metric Generation 1 Generation 2 Generation 3
Unique molecules 870 3,371 1,493
Median predicted pIC50 6.496 6.805 7.053
Shared-criteria pass rate 0.57% 0.86% 36.77%
Median QED 0.372 0.385 0.437
Median cLogP 6.074 5.904 5.447
Median Nutlin similarity 0.437 0.260 0.295
Applicability-domain support 0.443 0.297 0.323
01
INTERPRETATION

Reward design changed the molecular population, not just the top-ranked candidates.

The progression from Generation 1 to Generation 3 shows increasing median predicted potency and QED, decreasing median cLogP, and a marked increase in the fraction of molecules meeting the shared design criteria.

Structural evaluation is considered separately because population-level property improvement does not necessarily imply improved binding behavior for every selected candidate.

STRUCTURE-BASED VALIDATION

From Generated Molecules to Prioritized Candidates

Selected molecules were evaluated beyond population-level property optimization using molecular docking, MM/GBSA calculations, and short molecular-dynamics simulations.

01
DOCKING

Binding-pose screening

AutoDock Vina was used to evaluate candidate poses within the MDM2 binding site.

→
02
MM/GBSA

Energetic prioritization

Selected complexes were compared using MM/GBSA estimates derived from simulation snapshots.

→
03
MOLECULAR DYNAMICS

Short-timescale stability

Ligand RMSD and contact persistence were examined during 1 ns implicit-solvent trajectories.

01 MOLECULAR DOCKING

Candidate docking against MDM2

AutoDock Vina · PDB 5ZXF

Molecule Generation Docking score
Nutlin-3a Reference — −8.220 kcal/mol
Gen2 Candidate 1 Generation 2 −7.219 kcal/mol
Gen2 Candidate 2 Generation 2 −7.641 kcal/mol
Gen2 Candidate 3 Selected Generation 2 −8.227 kcal/mol
Gen3 Candidate 1 Generation 3 −7.179 kcal/mol
Gen3 Candidate 2 Generation 3 −6.976 kcal/mol
Gen3 Candidate 3 Generation 3 −7.051 kcal/mol
DOCKING OBSERVATION

Gen2 Candidate 3 produced a Vina score of −8.227 kcal/mol, close to the −8.220 kcal/mol Nutlin-3a reference. Generation 3 improved several population-level design properties but did not improve docking scores relative to the strongest Generation 2 candidate.

02 MM/GBSA

Comparative binding-energy estimates

More negative values indicate more favorable computational estimates.

REFERENCE

Nutlin-3a

−40.34 kcal/mol
GEN 2

Candidate 1

−29.34 kcal/mol
GEN 2

Candidate 2

−35.33 kcal/mol
GEN 2

Candidate 3

−39.51 kcal/mol
GEN 3 · POST-SUBMISSION

Candidate 1

−46.53 kcal/mol
GEN 3 · POST-SUBMISSION

Candidate 2

−52.79 kcal/mol
GEN 3 · POST-SUBMISSION

Candidate 3

−55.49 kcal/mol
Interpretation note

MM/GBSA values are comparative computational estimates, not experimental binding free energies. Generation 3 MM/GBSA calculations were performed as a post-submission extension of the original analysis.

03 SHORT MOLECULAR DYNAMICS

Generation 3 stability analysis

1 ns implicit-solvent trajectories

REFERENCE

Nutlin-3a

Reference
Mean ligand RMSD 1.16 Å
Mean contacts 13.54
GENERATION 3

Candidate 1

Mean ligand RMSD 2.83 Å
Mean contacts 1.38
GENERATION 3

Candidate 2

Balanced profile
Mean ligand RMSD 2.97 Å
Mean contacts 5.43
GENERATION 3

Candidate 3

Mean ligand RMSD 5.96 Å
Mean contacts 1.70
02
STRUCTURAL INTERPRETATION

Energetic ranking alone was insufficient for candidate prioritization.

Generation 3 Candidate 3 produced the most negative MM/GBSA estimate, but also showed the largest ligand RMSD and weak contact persistence during the short trajectory. Candidate 2 retained more ligand–pocket contacts while maintaining a substantially more favorable MM/GBSA estimate than the Nutlin reference.

These short implicit-solvent simulations are used for computational prioritization rather than as evidence of experimental binding stability.

MULTI-STAGE PRIORITIZATION

Different computational stages answer different questions.

Population optimization

Evaluates whether reward redesign shifts the overall generated chemical population toward desired properties.

Docking

Evaluates plausible binding poses and docking scores within the MDM2 binding site.

MM/GBSA

Provides comparative energetic estimates for selected protein–ligand complexes.

Molecular dynamics

Examines short-timescale structural behavior, ligand displacement, and contact persistence.

END-TO-END COMPUTATIONAL WORKFLOW

From Experimental Data to Candidate Prioritization

An integrated computational pipeline connects experimental activity data, machine-learning potency prediction, reinforcement-learning molecular generation, ADMET-aware optimization, and structure-based evaluation.

01
DB
DATA

ChEMBL

Curated MDM2 bioactivity data for target CHEMBL5023 used to construct the potency-prediction dataset.

4,146 curated compounds
→
02
ML
POTENCY MODEL

QSAR

ExtraTrees regression with molecular descriptors and fingerprints for predicted MDM2 inhibitory potency.

R² 0.732 test RMSE 0.612
→
03
RL
GENERATIVE DESIGN

REINVENT4

Reinforcement-learning molecular generation with iterative multi-objective reward redesign.

Gen 1 → Gen 3 iterative optimization
→
04
AD
PROPERTY OPTIMIZATION

ADMET

Multi-objective scoring incorporated predicted bioavailability and key ADMET liabilities alongside potency and molecular quality.

36.77% Gen3 shared-criteria pass
Candidate prioritization
↓
05
SEL
SELECTION

Candidate Selection

Generated molecules were prioritized using predicted properties, chemical constraints, and generation-level ranking.

Selected candidate set
→
06
DK
STRUCTURAL SCREENING

Molecular Docking

AutoDock Vina evaluation of generated molecules within the MDM2 binding pocket.

−8.227 kcal/mol · Gen2 C3
→
07
ΔG
ENERGETIC ANALYSIS

MM/GBSA

Comparative energetic estimates were calculated for selected protein–ligand complexes.

Comparative energetic ranking
→
08
MD
DYNAMIC EVALUATION

Molecular Dynamics

Short trajectories examined ligand RMSD and protein–ligand contact persistence.

1 ns implicit-solvent MD
03
WORKFLOW DESIGN

Candidate decisions emerge from multiple computational signals rather than a single score.

The workflow progressively narrows chemical space: experimental bioactivity data support potency modeling, the model guides generative optimization, ADMET-aware objectives reshape the molecular population, and structure-based analyses provide complementary evidence for candidate prioritization.

COMPUTATIONAL STACK
Python RDKit scikit-learn PyTorch REINVENT4 ADMET-AI AutoDock Vina AmberTools 3Dmol.js