Chapter 2: Drug Discovery-Specific MI Methods

This chapter covers Drug Discovery. You will learn 4 types of molecular representations (SMILES, QSAR principles, and ADMET 5 items specifically.

From Molecular Representation to Generative Models - Technical Foundation of AI-Driven Drug Discovery

2.1 Molecular Representation and Descriptors

The first step in drug discovery MI is converting chemical structures into formats that computers can understand. This choice of "molecular representation" significantly affects model performance.

2.1.1 SMILES Representation

SMILES (Simplified Molecular Input Line Entry System) is the most widely used format for representing molecular structures as strings.

Basic Rules:

Atoms: C (carbon), N (nitrogen), O (oxygen), S (sulfur), etc.
Bonds: Single bond (omitted), Double bond (=), Triple bond (#)
Ring structures: Marked with numbers (e.g., C1CCCCC1 = cyclohexane)
Branches: Expressed with parentheses (e.g., CC(C)C = isobutane)
Aromatic: Lowercase letters (e.g., c1ccccc1 = benzene)

Examples:

Advantages: - ✅ Compact (complex molecules in dozens of characters) - ✅ Human readable and writable - ✅ Suitable for database searches - ✅ Uniqueness (Canonical SMILES)

Disadvantages: - ❌ No 3D structural information (conformations lost) - ❌ Difficulty distinguishing tautomers - ❌ Existence of invalid SMILES (syntax errors)

2.1.2 Molecular Fingerprints

Molecular fingerprints represent molecules as bit vectors (sequences of 0/1). Widely used for similarity searches and QSAR.

ECFP (Extended Connectivity Fingerprints)

Principle: 1. Start from each atom 2. Hash neighborhood structure of radius R (typically 2-4) 3. Convert to bit vector (length 1024-4096)

Example: ECFP4 (radius 2, 4-step neighborhood)

Molecule: CCO (ethanol)

Atom 1 (C):
  - Radius 0: C
  - Radius 1: C-C
  - Radius 2: C-C-O

Atom 2 (C):
  - Radius 0: C
  - Radius 1: C-C, C-O
  - Radius 2: C-C-O, C-O

Atom 3 (O):
  - Radius 0: O
  - Radius 1: O-C
  - Radius 2: O-C-C

→ Hash these to determine bit positions
→ Set corresponding bits to 1

Types:

Tanimoto Coefficient for Similarity:

Tanimoto Coefficient = (A ∩ B) / (A ∪ B)

A, B: Fingerprint bit vectors of two molecules
∩: Bit AND (number of bits that are 1 in both)
∪: Bit OR (number of bits that are 1 in at least one)

Range: 0 (completely different) ~ 1 (perfect match)

Practical Thresholds: - Tanimoto > 0.85: Very similar (same compound class) - 0.70-0.85: Similar (possible similar pharmacological activity) - 0.50-0.70: Somewhat similar - < 0.50: Different

MACCS keys

Features: - Represents presence/absence of 166 fixed substructures - Examples: benzene ring, carboxylic acid, amino group, halogens, etc.

Advantages: - Interpretable (know which substructures are present) - Fast computation - Chemically meaningful similarity

Disadvantages: - Low information content (only 166 bits) - Difficult to handle novel scaffolds

2.1.3 3D Descriptors

3D descriptors quantify the three-dimensional structure of molecules.

Major 3D Descriptors:

1. Molecular Surface Area - TPSA (Topological Polar Surface Area): Polar surface area - Prediction: Oral absorption (favorable when TPSA < 140 Ų)

2. Volume and Shape - Molecular Volume - Sphericity - Aspect Ratio

3. Charge Distribution - Partial Charges: Gasteiger, MMFF - Dipole Moment - Quadrupole Moment

4. Pharmacophore - Hydrogen bond donor/acceptor positions - Hydrophobic regions - Positive/negative charge centers - Aromatic ring orientations

Example: Descriptors Used in Lipinski's Rule of Five

Molecular Weight (MW): < 500 Da
LogP (lipophilicity): < 5
Hydrogen Bond Donors (HBD): < 5
Hydrogen Bond Acceptors (HBA): < 10

Compounds satisfying these tend to have high oral absorption

2.1.4 Graph Representation (For Graph Neural Networks)

Molecules are represented as mathematical graphs.

Definition:

G = (V, E)

V: Vertices = Atoms
E: Edges = Bonds

Each vertex v has feature vector h_v
Each edge e has feature vector h_e

Atom (Vertex) Features: - Atomic number (C=6, N=7, O=8, etc.) - Atom type (C, N, O, S, F, Cl, Br, I) - Hybridization (sp, sp2, sp3) - Formal Charge - Aromaticity (Aromatic or not) - Hydrogen Count - Number of lone pairs - Chirality (R/S)

Bond (Edge) Features: - Bond order (Single=1, Double=2, Triple=3, Aromatic=1.5) - Bond type (Covalent, Ionic, etc.) - Part of ring or not - Stereochemistry (E/Z)

Graph Advantages: - Direct representation of molecular structure - Rotation and translation invariance - Learnable with Graph Neural Networks

2.2 QSAR (Quantitative Structure-Activity Relationship)

2.2.1 Basic Principles of QSAR

QSAR (Quantitative Structure-Activity Relationship) expresses the quantitative relationship between molecular structure and biological activity using mathematical equations.

Basic Hypothesis:

Similar molecular structures show similar biological activity (Similar Property Principle)

General QSAR Equation:

Activity = f(Descriptors)

Activity: Biological activity (IC50, EC50, Ki, etc.)
Descriptors: Molecular descriptors (MW, LogP, TPSA, etc.)
f: Mathematical function (linear regression, Random Forest, NN, etc.)

Historical QSAR Equation (Hansch-Fujita Equation, 1962):

log(1/C) = a * logP + b * σ + c * Es + d

C: Biological activity concentration (lower = higher activity)
logP: Partition coefficient (lipophilicity)
σ: Hammett constant (electronic effect)
Es: Steric parameter
a, b, c, d: Regression coefficients

2.2.2 QSAR Workflow

2.2.3 Types of QSAR Models

Classification Models (Predicting Active/Inactive)

Purpose: Predict whether a compound is active or inactive

Evaluation Metrics:

ROC-AUC: 0.5 (random) ~ 1.0 (perfect)
Target: > 0.80

Precision = TP / (TP + FP)
Recall = TP / (TP + FN)
F1 Score = 2 * (Precision * Recall) / (Precision + Recall)

TP: True Positive (correctly predict active)
FP: False Positive (incorrectly predict inactive as active)
FN: False Negative (incorrectly predict active as inactive)

Typical Problem Setting: - IC50 < 1 μM → Active (1) - IC50 ≥ 1 μM → Inactive (0)

Regression Models (Predicting Activity Values)

Purpose: Predict continuous values like IC50, Ki, EC50

Evaluation Metrics:

R² (Coefficient of determination): 0 (no correlation) ~ 1 (perfect)
Target: > 0.70

MAE (Mean Absolute Error) = Σ|y_pred - y_true| / n
RMSE (Root Mean Square Error) = √(Σ(y_pred - y_true)² / n)

y_pred: Predicted value
y_true: True value
n: Number of samples

Log Transformation: In most cases, activity values (IC50, etc.) are log-transformed before regression

pIC50 = -log10(IC50[M])

Example:
IC50 = 1 nM = 10^-9 M → pIC50 = 9.0
IC50 = 100 nM = 10^-7 M → pIC50 = 7.0

Range: typically 4-10 (10 μM ~ 0.1 nM)

2.2.4 Comparison of Machine Learning Methods

Real-World Performance (ChEMBL Data):

Task: Kinase inhibition activity prediction (classification)

Linear Regression: ROC-AUC = 0.72
Random Forest:      ROC-AUC = 0.87
SVM (RBF):          ROC-AUC = 0.85
Neural Network:     ROC-AUC = 0.89
LightGBM:           ROC-AUC = 0.90
GNN (MPNN):         ROC-AUC = 0.92

Training data: 10,000 compounds
Test data: 2,000 compounds

2.2.5 QSAR Applicability Domain

Problem: Predictions for compounds very different from training data are unreliable

Applicability Domain Definition:

AD = {x | included in chemical space of training data}

Determination methods:
1. Tanimoto distance: Similarity to nearest neighbor > 0.3
2. Leverage: h_i < 3p/n (h: hat matrix, p: number of features, n: sample size)
3. Standardized distance: d_i < 3σ (σ: standard deviation)

Compounds Outside AD: - Low prediction reliability - Experimental validation required - Consider model retraining

Example:

Training data: Kinase inhibitors (mainly ATP-competitive)
Novel compound: Allosteric inhibitor (different binding site)

→ Likely outside AD
→ Prediction accuracy not guaranteed

2.3 ADMET Prediction

2.3.1 Importance of ADMET

30% of clinical trial failures are due to ADMET issues

Failure Reduction Through ADMET Prediction: - Early ADMET evaluation (during lead discovery stage) - Eliminate problematic compounds - Reduce development costs by $100M-500M

2.3.2 Absorption

Caco-2 Permeability

Caco-2 Cells: Colon cancer cell line, model of intestinal epithelial cells

Measurement: Rate of permeation through cell layer (Papp: apparent permeability coefficient)

Papp [cm/s]

Papp > 10^-6 cm/s: High permeability (good absorption)
10^-7 < Papp < 10^-6: Moderate
Papp < 10^-7 cm/s: Low permeability (poor absorption)

Prediction Model:

# Simple prediction equation (R² = 0.75)
log Papp = 0.5 * logP - 0.01 * TPSA + 0.3 * HBD - 2.5

logP: Partition coefficient (lipophilicity)
TPSA: Topological polar surface area
HBD: Number of hydrogen bond donors

Machine Learning Prediction Accuracy: - Random Forest: R² = 0.80-0.85 - Neural Network: R² = 0.82-0.88 - Graph NN: R² = 0.85-0.90

Oral Bioavailability (F%)

Definition: Fraction of administered dose that reaches systemic circulation

F% = (AUC_oral / AUC_iv) * (Dose_iv / Dose_oral) * 100

AUC: Area under the plasma concentration-time curve

Target Values: - F% > 30%: Acceptable range - F% > 50%: Good - F% > 70%: Excellent

Predictive Factors: 1. Solubility 2. Permeability 3. First-pass metabolism (hepatic first-pass metabolism) 4. P-glycoprotein efflux

BCS Classification (Biopharmaceutics Classification System):

2.3.3 Distribution

Plasma Protein Binding

Definition: Percentage of drug bound to plasma proteins (albumin, α1-acid glycoprotein)

Binding rate = (Bound / Total) * 100%

Bound: Bound drug concentration
Total: Total drug concentration

Clinical Significance: - High binding (> 90%): Low free drug (active form) - Risk of drug-drug interactions (binding site competition) - Affects volume of distribution

Prediction Models: - Random Forest: R² = 0.65-0.75 - Deep Learning: R² = 0.70-0.80

Blood-Brain Barrier (BBB) Permeability

LogBB (Brain/Blood Ratio):

LogBB = log10(C_brain / C_blood)

LogBB > 0: Concentrates in brain (favorable for CNS drugs)
LogBB < -1: Does not penetrate brain (avoid CNS side effects)

Predictive Factors: - Molecular weight (< 400 Da favorable) - TPSA (< 60 Ų favorable) - LogP (2-5 optimal) - Number of hydrogen bonds (fewer is better)

Machine Learning Prediction: - Random Forest: R² = 0.70-0.80 - Neural Network: R² = 0.75-0.85

2.3.4 Metabolism

CYP450 Inhibition

CYP450 (Cytochrome P450): Major metabolic enzyme in liver

Major Isoforms:

Problem with Inhibition: - Drug-drug interactions (DDI) - Increased blood concentration of co-administered drugs → toxicity risk

Prediction (Classification Problem):

Inhibitor: IC50 < 10 μM
Non-inhibitor: IC50 ≥ 10 μM

Prediction accuracy:
Random Forest: ROC-AUC = 0.85-0.90
Neural Network: ROC-AUC = 0.87-0.92

Metabolic Stability

Measurement: Percentage remaining after incubation with liver microsomes

t1/2 (half-life) [min]

t1/2 > 60 min: Stable (slow metabolism)
30 < t1/2 < 60: Moderate
t1/2 < 30 min: Unstable (fast metabolism)

Prediction: - Random Forest: R² = 0.55-0.65 (somewhat difficult) - Graph NN: R² = 0.60-0.70

2.3.5 Excretion

Renal Clearance

Definition: Rate of drug removal by kidneys

CL_renal [mL/min/kg]

CL_renal > 10: High clearance (rapidly excreted)
1 < CL_renal < 10: Moderate
CL_renal < 1: Low clearance

Influencing Factors: - Molecular weight (smaller = easier excretion) - Polarity (higher = easier excretion) - Renal transporter substrate

Half-Life (t1/2)

Definition: Time for plasma concentration to decrease by half

t1/2 = 0.693 / (CL / Vd)

CL: Clearance (systemic)
Vd: Volume of distribution

Clinical Significance: - t1/2 < 2 h: Frequent dosing required (3-4 times/day) - 2 < t1/2 < 8 h: Twice daily dosing - t1/2 > 8 h: Once daily dosing possible

Prediction Accuracy: - Random Forest: R² = 0.55-0.65 (difficult) - Complex combination of pharmacokinetic parameters

2.3.6 Toxicity

hERG Inhibition (Cardiotoxicity)

hERG (human Ether-à-go-go-Related Gene): Cardiac potassium channel

Result of Inhibition: QT prolongation → fatal arrhythmia (Torsades de pointes)

Risk Assessment:

IC50 < 1 μM: High risk (development termination)
1 < IC50 < 10 μM: Medium risk (careful evaluation needed)
IC50 > 10 μM: Low risk (safe)

Prediction Accuracy (Highest Accuracy in ADMET Predictions): - Random Forest: ROC-AUC = 0.85-0.90 - Deep Learning: ROC-AUC = 0.90-0.95 - Graph NN: ROC-AUC = 0.92-0.97

Structural Alerts (Structures Prone to hERG Inhibition): - Basic nitrogen atom (pKa > 7) - Hydrophobic aromatic rings - Flexible linker

Hepatotoxicity

DILI (Drug-Induced Liver Injury): Drug-induced liver damage

Mechanisms: 1. Formation of reactive metabolites 2. Mitochondrial toxicity 3. Bile acid transport inhibition

Prediction (Classification Problem):

Hepatotoxic drug classification:

Random Forest: ROC-AUC = 0.75-0.85
Neural Network: ROC-AUC = 0.78-0.88

Challenge: Data shortage (many cases discovered post-approval)

Mutagenicity (Ames Test)

Ames Test: Mutagenicity test using bacteria

Prediction:

Positive: Mutagenic (carcinogenicity risk)
Negative: Non-mutagenic

Random Forest: ROC-AUC = 0.80-0.90
Deep Learning: ROC-AUC = 0.85-0.93

Structural Alerts: - Nitro group (-NO2) - Azo group (-N=N-) - Epoxide - Alkylating agents

2.4 Molecular Generative Models

2.4.1 Need for Generative Models

Traditional Virtual Screening: - Select from existing compound libraries (10^6-10^9 compounds) - Limitation: Cannot discover compounds not in library

Generative Model Approach: - Generate novel molecules directly - Freely explore chemical space (10^60 molecules) - Design molecules with desired properties

Types of Generative Models:

2.4.2 VAE (Variational Autoencoder)

Architecture:

SMILES → Encoder → Latent space (z) → Decoder → SMILES'

Encoder: Compresses SMILES string to low-dimensional vector (z)
Latent space: Continuous chemical space (typically 50-500 dimensions)
Decoder: Reconstructs SMILES string from vector (z)

Training Objective:

Loss = Reconstruction Loss + KL Divergence

Reconstruction Loss: Error in reconstructing input SMILES
KL Divergence: Regularization of latent space (approach normal distribution)

Molecule Generation Workflow: 1. Train VAE on known molecules (training data) 2. Sample in latent space (z ~ N(0, I)) 3. Generate novel SMILES with Decoder 4. Validity check (parseable by RDKit?) 5. Property prediction (QSAR models, ADMET models) 6. Select favorable molecules

Advantages: - ✅ Optimization possible in continuous latent space - ✅ Explore around known molecules (generate similar molecules) - ✅ Generate intermediate molecules via interpolation

Disadvantages: - ❌ High rate of invalid SMILES generation (30-50%) - ❌ Difficult to generate novel scaffolds (depends on training data)

Example: ChemVAE (Gómez-Bombarelli et al., 2018)

Training data: ZINC 250K compounds
Latent space: 196 dimensions
Valid SMILES generation rate: 68%
Application: Property optimization using penalty method (LogP, QED)

2.4.3 GAN (Generative Adversarial Network)

Two Networks:

Generator: Noise → Molecular SMILES
Discriminator: SMILES → Real or Fake?

Adversarial learning:
- Generator: Generate molecules to fool Discriminator
- Discriminator: Distinguish Real (training data) from Fake (generated)

Training Process:

1. Generator: Random noise → SMILES generation
2. Discriminator: Real/Fake classification
3. Generator: Update with Discriminator's gradient (make Fake "Real-like")
4. Discriminator: Improve classification accuracy
5. Repeat → Generator produces molecules similar to training data

Advantages: - ✅ Diverse molecule generation (with mode collapse countermeasures) - ✅ High-quality molecules (filtering effect by Discriminator)

Disadvantages: - ❌ Training instability (mode collapse, gradient vanishing) - ❌ Difficult to evaluate (quantifying generation quality)

Example: ORGANIC (Guimaraes et al., 2017)

Generator: LSTM (sequential SMILES generation)
Discriminator: CNN (SMILES string classification)
Application: Combined with reinforcement learning for property optimization

Insilico Medicine's Chemistry42: - GAN-based - Discovered IPF treatment candidate in 18 months - Technology: WGAN-GP (Wasserstein GAN with Gradient Penalty)

2.4.4 Transformer (GPT-like Models)

Principle: - Treat SMILES like natural language - Capture context with Attention mechanism - Large-scale pre-training (10^6-10^7 molecules)

Architecture:

SMILES: C C ( = O ) O [EOS]
↓
Tokenization: [C] [C] [(] [=] [O] [)] [O] [EOS]
↓
Transformer Encoder/Decoder
↓
Next token prediction: P(token_i+1 | token_1, ..., token_i)

Generation Method: 1. Start from start token ([SOS]) 2. Probabilistically sample next token 3. Repeat until [EOS] 4. Validate as SMILES

Advantages: - ✅ High valid SMILES generation rate (> 90%) - ✅ Benefits of large-scale pre-training (transfer learning) - ✅ Controllable generation (conditional generation)

Disadvantages: - ❌ Computational cost (large models) - ❌ Difficult to control novelty

Examples: 1. ChemBERTa (HuggingFace, 2020) - Pre-training: 10 million SMILES (PubChem) - Fine-tuning: High-accuracy QSAR with 100 samples

2. MolGPT (Bagal et al., 2021) - GPT-2 architecture - Conditional generation (specify property values)

3. SMILES-BERT (Wang et al., 2019) - Masked Language Model (MLM task) - High accuracy with small data via transfer learning

2.4.5 Reinforcement Learning (RL)

Setup:

Agent: Molecular generative model
Environment: Chemical space
Action: Select next token (SMILES generation)
State: Current SMILES prefix
Reward: Properties of generated molecule (QED, LogP, ADMET, etc.)

RL + Generative Model Flow:

Reward Function Design:

Reward = w1 * QED + w2 * LogP_score + w3 * SA_score + w4 * ADMET_score

QED: Quantitative Estimate of Druglikeness
LogP_score: Lipophilicity score (2-5 optimal)
SA_score: Synthetic Accessibility
ADMET_score: ADMET model prediction

w1, w2, w3, w4: Weights (multi-objective optimization)

Learning Algorithms: 1. Policy Gradient (REINFORCE) ``` ∇J(θ) = E[∇log π(a|s) * R]

π: Policy (generation probability) R: Cumulative reward ```

2. Proximal Policy Optimization (PPO) - More stable learning - Clip gradients to prevent large updates

Advantages: - ✅ Explicit optimization (control via reward function) - ✅ Multi-objective optimization possible - ✅ Balance exploration and exploitation

Disadvantages: - ❌ Difficult to design reward function - ❌ Time-consuming training

Example: ReLeaSE (Popova et al., 2018)

Technology: RL + LSTM
Objective: Optimize LogP, QED
Result: Generated molecules with properties superior to known drugs

2.4.6 Graph Generative Models

Principle: Directly generate molecular graphs

Generation Process:

1. Initial state: Empty graph
2. Add atom: Which atom type to add (C, N, O, etc.)
3. Add bond: Between which atoms to add bonds
4. Repeat: Until desired size is reached

Major Methods:

1. GraphRNN (You et al., 2018) - Sequential graph generation with RNN

2. Junction Tree VAE (Jin et al., 2018) - Generate molecular scaffold (Junction Tree) - 100% valid SMILES generation rate (theoretically)

3. MoFlow (Zang & Wang, 2020) - Based on Normalizing Flows - Exact probability density through invertible transformations

Advantages: - ✅ Generate only valid molecules (incorporate chemical constraints of graphs) - ✅ Can consider 3D structure

Disadvantages: - ❌ High computational cost - ❌ Complex implementation

2.5 Major Databases and Tools

2.5.1 ChEMBL

Overview: - Bioactivity database (European Bioinformatics Institute, EBI) - 2+ million compounds - 15+ million bioactivity data points - 14,000+ targets

Data Structure:

Compound: ChEMBL ID, SMILES, InChI, molecular weight, etc.
Assay: Assay type, target, cell line, etc.
Activity: IC50, Ki, EC50, Kd, etc.
Target: Protein, gene, cell, etc.

API Access:

from chembl_webresource_client.new_client import new_client

# Target search
target = new_client.target
kinases = target.filter(target_type='PROTEIN KINASE')

# Get compound activity data
activity = new_client.activity
egfr_data = activity.filter(target_chembl_id='CHEMBL203', pchembl_value__gte=6)
# pchembl_value ≥ 6 → IC50 ≤ 1 μM

Use Cases: - ChEMBL: Bioactivity data, QSAR training - PubChem: Chemical structures, literature search - DrugBank: Approved drugs, clinical information - BindingDB: Protein-ligand binding affinity

2.5.2 PubChem

Overview: - Chemical information database (NIH, USA) - 100+ million compounds - Bioactivity assay data (PubChem BioAssay)

Features: - Free access - REST API, FTP provided - 2D/3D structure data - Literature links (PubMed)

Applications: - Building compound libraries - Structure search (similarity, substructure) - Property data collection

2.5.3 DrugBank

Overview: - Approved and clinical trial drug database (Canada) - 14,000+ drugs - Detailed pharmacokinetic, pharmacodynamic data

Information: - ADMET properties - Drug-drug interactions - Target information - Clinical trial status

Applications: - Drug repurposing (finding new indications for approved drugs) - ADMET training data - Benchmarking

2.5.4 BindingDB

Overview: - Protein-ligand binding affinity database - 2.5+ million binding data points - 9,000+ proteins

Data Types: - Ki (inhibition constant) - Kd (dissociation constant) - IC50 (50% inhibitory concentration) - EC50 (50% effective concentration)

Applications: - Docking validation - Binding affinity prediction model training

2.5.5 RDKit

Overview: - Open-source cheminformatics library - Python, C++, Java support - Standard tool for drug discovery MI

Main Functions: 1. Molecular I/O: Read/write SMILES, MOL, SDF 2. Descriptor calculation: 200+ physicochemical properties 3. Molecular fingerprints: ECFP, MACCS, RDKit FP 4. Substructure search: SMARTS, MCS (maximum common substructure) 5. 2D drawing: Molecular structure visualization 6. 3D structure generation: ETKDG (distance geometry method)

Usage Example:

from rdkit import Chem
from rdkit.Chem import Descriptors, AllChem

# Read SMILES
mol = Chem.MolFromSmiles('CC(=O)OC1=CC=CC=C1C(=O)O')  # Aspirin

# Calculate descriptors
mw = Descriptors.MolWt(mol)  # 180.16
logp = Descriptors.MolLogP(mol)  # 1.19

# Molecular fingerprint
fp = AllChem.GetMorganFingerprintAsBitVect(mol, radius=2, nBits=2048)

# Similarity calculation
mol2 = Chem.MolFromSmiles('CC(C)Cc1ccc(cc1)C(C)C(=O)O')  # Ibuprofen
similarity = DataStructs.TanimotoSimilarity(fp, fp2)  # 0.31

2.6 Overall Drug Discovery MI Workflow

Time and Cost for Each Stage:

Cost Reduction: - Virtual Screening: $500M → $50M (90% reduction) - ADMET failure reduction: $200M → $50M (75% reduction) - Experimental number reduction: 1,000 experiments → 100 experiments (90% reduction)

Learning Objectives Check

Upon completing this chapter, you will be able to explain:

Basic Understanding (Remember & Understand)

• ✅ Explain 4 types of molecular representations (SMILES, molecular fingerprints, 3D descriptors, Graph)
• ✅ Understand QSAR principles and 5-step workflow
• ✅ Explain ADMET 5 items specifically
• ✅ Understand differences among 4 types of molecular generative models (VAE, GAN, Transformer, RL)
• ✅ Grasp features of major databases (ChEMBL, PubChem, DrugBank, BindingDB)

Practical Skills (Apply & Analyze)

• ✅ Calculate Lipinski's Rule of Five and evaluate druglikeness
• ✅ Calculate molecular similarity using Tanimoto coefficient
• ✅ Interpret QSAR model performance metrics (R², ROC-AUC)
• ✅ Judge development feasibility from ADMET prediction results
• ✅ Appropriately select machine learning methods (RF, SVM, NN, GNN)

Application Ability (Evaluate & Create)

• ✅ Design workflow suitable for new drug discovery projects
• ✅ Evaluate QSAR model Applicability Domain
• ✅ Judge application scenarios for molecular generative models
• ✅ Build datasets using ChEMBL API
• ✅ Optimize overall drug discovery MI flow

Exercises

Easy (Basic Confirmation)

Q1: What compound does the following SMILES string represent?

CCO

a) Methanol b) Ethanol c) Propanol d) Butanol

Q2: In Lipinski's Rule of Five, what is the upper limit for molecular weight?

Q3: A compound has an hERG inhibition IC50 of 0.5 μM. How is this classified in risk assessment? a) Low risk b) Medium risk c) High risk

Medium (Application)

Q4: When comparing ECFP4 fingerprints (2048 bits) of two molecules, the following results were obtained. Calculate the Tanimoto coefficient.

Molecule A: Number of positions with bit 1 = 250
Molecule B: Number of positions with bit 1 = 280
Number of positions with bit 1 in both = 120

Tanimoto = (A ∩ B) / (A ∪ B)

A ∩ B = 120 (number of bits that are 1 in both)
A ∪ B = 250 + 280 - 120 = 410

Tanimoto = 120 / 410 = 0.293

Q5: A QSAR model was built and the following performance was obtained. Is this model practical?

Training data: R² = 0.92, MAE = 0.3
Test data: R² = 0.58, MAE = 1.2

Q6: A compound's Caco-2 permeability is Papp = 5 × 10^-7 cm/s. Evaluate this compound's oral absorption.

Hard (Advanced)

Q7: When generating 10,000 novel molecules with ChemVAE (latent space 196 dimensions), 6,800 were valid SMILES. Of the generated molecules, 40% had a Tanimoto coefficient > 0.95 with the most similar molecule in the training data (ZINC 250K). How do you evaluate this result? What improvement strategies are available?

Q8: For the following drug discovery project, which machine learning method should be selected? Explain with reasoning.

Project: Development of EGFR (epidermal growth factor receptor) kinase inhibitor
Data: EGFR activity data from ChEMBL, 15,000 compounds
Task: IC50 prediction (regression)
Goal: R² > 0.80, prediction time < 1 sec/compound
Additional requirement: Want model interpretability (which structures contribute to activity)

# Requirements:
# - Python 3.9+
# - shap>=0.42.0

"""
Example: Q8: For the following drug discovery project, which machine 

Purpose: Demonstrate data visualization techniques
Target: Advanced
Execution time: 1-5 minutes
Dependencies: None
"""

from sklearn.ensemble import RandomForestRegressor
from sklearn.model_selection import cross_val_score
import shap

# Model training
rf = RandomForestRegressor(
    n_estimators=500,  # Number of trees (more = more stable)
    max_depth=20,      # Depth limit (prevent overfitting)
    min_samples_leaf=5,
    n_jobs=-1          # Parallelization
)

rf.fit(X_train, y_train)

# Performance evaluation
cv_scores = cross_val_score(rf, X_train, y_train, cv=5, scoring='r2')
print(f"Cross-validation R²: {cv_scores.mean():.3f} ± {cv_scores.std():.3f}")

# Feature importance
importances = pd.DataFrame({
    'feature': feature_names,
    'importance': rf.feature_importances_
}).sort_values('importance', ascending=False)

print("Top 10 important features:")
print(importances.head(10))

# SHAP interpretation
explainer = shap.TreeExplainer(rf)
shap_values = explainer.shap_values(X_test[:100])  # Sample 100 compounds

shap.summary_plot(shap_values, X_test[:100], feature_names=feature_names)
# → Visualize which features contribute to high/low activity

Next Steps

In Chapter 2, you understood drug discovery-specific MI methods. In Chapter 3, you'll implement these methods in Python and actually run them. Through 30 code examples using RDKit and ChEMBL, you'll acquire practical skills.

Chapter 3: Implementing Drug Discovery MI in Python - RDKit & ChEMBL Practice →

References

1. Gómez-Bombarelli, R., et al. (2018). "Automatic chemical design using a data-driven continuous representation of molecules." ACS Central Science, 4(2), 268-276.

2. Guimaraes, G. L., et al. (2017). "Objective-Reinforced Generative Adversarial Networks (ORGAN) for Sequence Generation Models." arXiv:1705.10843.

3. Jin, W., Barzilay, R., & Jaakkola, T. (2018). "Junction tree variational autoencoder for molecular graph generation." ICML 2018.

4. Lipinski, C. A., et al. (2001). "Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings." Advanced Drug Delivery Reviews, 46(1-3), 3-26.

5. Popova, M., et al. (2018). "Deep reinforcement learning for de novo drug design." Science Advances, 4(7), eaap7885.

6. Rogers, D., & Hahn, M. (2010). "Extended-connectivity fingerprints." Journal of Chemical Information and Modeling, 50(5), 742-754.

7. Wang, S., et al. (2019). "SMILES-BERT: Large scale unsupervised pre-training for molecular property prediction." BCB 2019.

8. Zhavoronkov, A., et al. (2019). "Deep learning enables rapid identification of potent DDR1 kinase inhibitors." Nature Biotechnology, 37(9), 1038-1040.

9. Gaulton, A., et al. (2017). "The ChEMBL database in 2017." Nucleic Acids Research, 45(D1), D945-D954.

10. Landrum, G. (2023). RDKit: Open-source cheminformatics. https://www.rdkit.org

Return to Series Table of Contents | Return to Chapter 1 | Proceed to Chapter 3 →