GenMol: XAI-Guided Fragment-Based Molecule Generation

A computational pipeline for generating novel antimicrobial molecules through systematic recombination of XAI-extracted pharmacophoric fragments. This repository validates explainable AI (XAI) attributions by demonstrating that fragments identified as important by RGCN models produce active molecules when recombined.

Overview

This work is part of a thesis series on explainable AI for antimicrobial drug discovery:

Paper	Focus	Key Contribution
Paper 1	XAI Evaluation Framework	4-tier framework; RGCN-SME identified as deployment-ready
Paper 2	Fragment Extraction	12,993 fragments from 68,736 compounds; SELECT design rules
Paper 3 (This Work)	Fragment Recombination	700 molecules via acyl transfer (amide/urea); Tier 3 validation

Key Results

Model	Tier 3 Actives	Hit Rate	Top Library	Top Hit Rate
SA	68	39.8%	SA_CA	72.2%
EC	26	17.4%	EC	70.0%
CA	8	5.2%	CA	9.1%

Hit Rate Definition:

• Hit Rate = Overall hit rate across all 7 libraries combined (Total Tier 3 actives / Total Tier 3 compounds)
• Top Hit Rate = Best-performing individual library for that pathogen model

Example: For SA model, 68 of 171 total Tier 3 compounds were predicted active (39.8% overall), but the SA_CA library achieved 72.2% (26/36), demonstrating library-specific optimization.

---

• Novelty: 99.6% (697/700) of generated molecules are novel (not in training data)
• Linkage Diversity: 600 amides (85.7%) + 100 ureas (14.3%) from oxazolidinone fragments
• Library-Specific Validation: SA library achieves 58% SA hit rate vs 0% EC hit rate
• XAI Attribution: 95% positive attributions for Tier 3 actives
• SA Score Significance: Tier 3 actives cluster at SA scores 3-4 (p<0.0001 for SA model)

Repository Structure

GenMol/
├── README.md                           <- This file
├── requirements.txt                    <- Python dependencies
├── LICENSE                             <- MIT license
│
├── data/
│   ├── source_fragments/               <- 14 source fragment CSVs
│   │   ├── SA_positive_scaffolds.csv
│   │   ├── SA_positive_substituents.csv
│   │   ├── EC_positive_scaffolds.csv
│   │   ├── EC_positive_substituents.csv
│   │   ├── CA_positive_scaffolds.csv
│   │   ├── CA_positive_substituents.csv
│   │   ├── SA_EC_positive_scaffolds.csv
│   │   ├── SA_EC_positive_substituents.csv
│   │   ├── SA_CA_positive_scaffolds.csv
│   │   ├── SA_CA_positive_substituents.csv
│   │   ├── CA_EC_positive_scaffolds.csv
│   │   ├── CA_EC_positive_substituents.csv
│   │   ├── TRIPLE_positive_scaffolds.csv
│   │   └── TRIPLE_positive_substituents.csv
│   │
│   ├── genmol_input.csv                <- 700 generated molecules with fragment metadata
│   │
│   ├── predictions/                    <- RGCN model predictions & XAI attributions
│   │   ├── genmol_SA_predictions.csv   <- 700 compounds evaluated against SA model
│   │   ├── genmol_EC_predictions.csv   <- 700 compounds evaluated against EC model
│   │   └── genmol_CA_predictions.csv   <- 700 compounds evaluated against CA model
│   │
│   └── novelty_duplicates.csv          <- 3 compounds found in training data
│
├── libraries/                          <- Built safe fragment libraries
│   ├── SA_library/
│   │   ├── safe_library_SA.json
│   │   └── SA_builder_diagnostics.txt
│   ├── EC_library/
│   ├── CA_library/
│   ├── SA_EC_library/
│   ├── SA_CA_library/
│   ├── CA_EC_library/
│   └── TRIPLE_library/
│
├── scripts/
│   ├── safe_library_builder.py         <- Build libraries from source_fragments/
│   ├── amide_molecule_generator.py     <- Generate molecules via acyl transfer (amide/urea)
│   ├── tier3_prediction_analysis.py    <- Analyze Tier 3 predictions
│   ├── verify_novelty.py               <- Verify molecules not in training data
│   └── utils/
│       ├── valency_completion.py
│       └── smiles_cleaner.py
│
├── outputs/                            <- Generated molecules
│   └── generated_molecules.csv
│
└── results/                            <- Analysis results
    └── tier3_reports/

Data Files

Generated Molecules (data/genmol_input.csv)

The main dataset containing all 700 generated molecules with fragment metadata:

Column	Description
COMPOUND_ID	Unique identifier (CMPD_001 to CMPD_700)
SMILES	Product structure with atom mapping
source_library	Origin library (SA, EC, CA, SA_EC, SA_CA, CA_EC, TRIPLE)
route_class	Synthetic route (A, B, or C)
product_linkage_type	amide or urea - the bond formed
acid_fragment_smiles	SMILES of acid-bearing fragment
amine_fragment_smiles	SMILES of amine-bearing fragment
acid_handle_origin	Type of acid handle (native_acid, latent_ester, etc.)
MW, LogP, TPSA, etc.	Physicochemical properties
SA_score	Synthetic accessibility score (Ertl)

Prediction Files (data/predictions/)

Each prediction CSV contains all 700 generated compounds evaluated against one RGCN model:

File	Model	Contents
genmol_SA_predictions.csv	S. aureus	Predictions + XAI attributions
genmol_EC_predictions.csv	E. coli	Predictions + XAI attributions
genmol_CA_predictions.csv	C. albicans	Predictions + XAI attributions

Key columns in prediction files:

Column	Description
COMPOUND_ID	Unique identifier (CMPD_001 to CMPD_700)
SMILES	Molecule structure with atom mapping
ensemble_prediction	Mean probability from 5-fold ensemble
prediction	Binary classification (0/1)
decision_scenario	Tier classification (A, B, C, or D)
murcko_substructure_N_smiles	SMILES of Nth Murcko scaffold
murcko_substructure_N_attribution	XAI attribution score for Nth scaffold

Decision Scenarios:

• A (Tier 3): High agreement + High reliability - most trustworthy
• B: High agreement + Low reliability
• C: Low agreement + High reliability
• D: Low agreement + Low reliability

Seven-Library Architecture

Library	Description	Source Fragments	Purpose
SA	S. aureus-specific	Scaffolds + Substituents	Gram-positive only
EC	E. coli-specific	Scaffolds + Substituents	Gram-negative only
CA	C. albicans-specific	Scaffolds + Substituents	Antifungal only
SA_EC	Dual SA+EC active	Scaffolds + Substituents	Antibacterial broad-spectrum
SA_CA	Dual SA+CA active	Scaffolds + Substituents	Cross-kingdom
CA_EC	Dual CA+EC active	Scaffolds + Substituents	Cross-kingdom
TRIPLE	Active against all 3	Scaffolds + Substituents	Pan-active broad-spectrum

Installation

Install dependencies

pip install -r requirements.txt

### Requirements

- Python 3.8+
- RDKit 2023.03+
- pandas
- numpy
- scipy (optional, for statistical tests)

## Quick Start

### 1. Build Safe Libraries

Build fragment libraries from source CSVs:

```bash
# Build a single library
python scripts/safe_library_builder.py --library SA --base-dir .

# Build all 7 libraries
python scripts/safe_library_builder.py --all --base-dir .

2. Generate Molecules

Generate novel molecules via acyl transfer:

python scripts/amide_molecule_generator.py \
    --library SA \
    --safe-library libraries/SA_library/safe_library_SA.json \
    --max-products 100 \
    --seed 42

3. Analyze Tier 3 Predictions

After running RGCN predictions, analyze results:

# Analyze a single model
python scripts/tier3_prediction_analysis.py --model SA --base-dir .

# Analyze all 3 models
python scripts/tier3_prediction_analysis.py --all --base-dir .

4. Verify Novelty

Verify that generated molecules are not present in the training data:

python scripts/verify_novelty.py \
    --base-dir . \
    --training-dir /path/to/training/data

The training directory should contain:

• S_aureus_input.csv
• E_coli_input.csv
• C_albicans_input.csv

Pre-computed results: 697/700 (99.6%) molecules confirmed novel. The 3 duplicates are simple molecules from the TRIPLE library listed in data/novelty_duplicates.csv.

Methodology

Fragment Recombination via Acyl Transfer

Molecules are generated by coupling acid-bearing fragments with amine-bearing fragments using SMIRKS:

[C:1](=[O:2])[O:3][H].[N:5;H2,H1;!$(N=[!#6])]>>[C:1](=[O:2])[N:5]

Product Linkage Types:

• Amide (N-C(=O)-C): 600 molecules (85.7%) - from esters, lactones, anhydrides
• Urea (N-C(=O)-N): 100 molecules (14.3%) - from oxazolidinones (cyclic carbamates)

Route Classification:

• Route A (Native): Anhydrides, carboxylic acids - mildest conditions (rt-40°C)
• Route B (Easy Latent): Methyl/ethyl esters - mild conditions (60-80°C)
• Route C (Hard Latent): Lactones, oxazolidinones (urea formation), tert-butyl/benzyl esters - harsh conditions (100°C)

Why Acyl Transfer?

Acyl transfer reactions (amide and urea formation) were chosen as the fragment coupling strategy because:

1. Synthetic accessibility: Both amide and urea bonds are among the easiest to form in medicinal chemistry, requiring simple reagents and mild conditions
2. Pharmaceutical relevance: Amide bonds are the most common linkage in FDA-approved drugs; ureas are prevalent in kinase inhibitors and antimicrobials
3. Robust methodology: High-yielding reactions with well-established protocols suitable for parallel synthesis
4. Validation focus: Provides a reliable baseline for validating XAI-derived fragments before exploring more complex chemistries

Future Directions: Advanced coupling strategies such as click chemistry (CuAAC), Suzuki-Miyaura cross-coupling, or C-H activation could expand the chemical space accessible from XAI-identified fragments.

SMILES Atom Mapping Notation

The generated molecule SMILES contain atom mapping numbers (e.g., [C:1], [N:5]) that track atoms through the coupling reaction:

Product SMILES:  O=[C:1](NCCN(O)Cc1ccccc1)[NH:5]Cc1nccs1
                     ↑                      ↑
                 Atom map 1             Atom map 5
              (from acid fragment)   (from amine fragment)

Atom Map	Origin	Chemical Role
[C:1]	Acid fragment	Carbonyl carbon of the newly formed amide/urea
[N:5]	Amine fragment	Nitrogen of the newly formed amide/urea

These atom maps are useful for:

• Attribution mapping: Tracing which XAI attributions belong to which parent fragment
• Synthetic clarity: Showing the retrosynthetic disconnection point
• Fragment tracking: Understanding how the molecule was assembled

Note: The atom mapping does not affect molecular structure visualization in RDKit or other tools. To obtain clean SMILES without atom maps:

from rdkit import Chem
mol = Chem.MolFromSmiles(mapped_smiles)
for atom in mol.GetAtoms():
    atom.SetAtomMapNum(0)
clean_smiles = Chem.MolToSmiles(mol)

Tier 3 Internal Consistency

Tier 3 (Scenario A) represents the highest confidence predictions:

1. Prediction Agreement: All 5 ensemble models agree on classification
2. Explanation Reliability: Fragment attributions show >70% consistency with magnitude >0.1

Safety Filtering

All generated molecules are filtered for:

• PAINS A/B/C substructures
• Michael acceptor precursors
• Alkyl halides and azides

Key Findings

Novelty Verification

99.6% (697/700) of generated molecules are confirmed novel - not present in any of the three model training datasets (68,725 unique structures). The 3 duplicates are simple molecules (acetohydrazide derivatives) from the TRIPLE library.

Attribution Predicts Transferability

High training attribution (>0.2) predicts cross-pathogen transfer better than high training activity (>90%).

Route B Superiority

Route B (mild aminolysis) dominates at 57.1% of generated compounds, with significantly better synthetic accessibility (SA_score 3.22) compared to Route C (SA_score 3.66, p<0.0001).

SA Score Goldilocks Zone

Tier 3 actives cluster at SA scores 3-4 (p<0.0001), indicating optimal synthetic accessibility.

Citation

@article{xai_fbdd_antimicrobial,
  title={TBD},
  author={Onawole, Abdulmujeeb T; Blaskovich, Mark A.T.; Zuegg, Johannes},
  journal={TBD},
  year={2025},
  note={Paper 3: Fragment recombination validation}
}

License

MIT License - see LICENSE for details.

Contact

For questions about the methodology or scripts, please open an issue.