EloKriti

Chemistry-aware active learning for the discovery of BPA-free dental resin backbones using molecular-dynamics-derived network properties.

This repository contains the active-learning and multi-objective Bayesian-optimization code used in:

Active-Learning Discovery of BPA-Free Dental Resin Candidates through Backbone Screening and Network Validation

EloKriti uses Gaussian-process surrogate models to prioritize unevaluated candidate backbones for explicit virtual-curing and molecular-dynamics evaluation.

The three optimization objectives are:

• maximize elastic modulus;
• minimize volumetric shrinkage;
• minimize fractional free volume.

The surrogate models are used only for candidate selection. Final property values and Pareto assignments are based on explicit molecular-dynamics evaluations.

EloKriti/
├── polymers_bo/
│   ├── data.py
│   ├── surrogate.py
│   ├── bo.py
│   └── suggest.py
├── scripts/
│   ├── suggest_next.py
│   ├── analyze_hypervolume.py
│   └── make_al_progress.py
├── notebooks/
│   └── candidate_pool_construction.ipynb
├── images/
│   └── AL_Workflow_Final.png
├── environment.yml
├── requirements.txt
├── README.md
└── LICENSE

The Python packages required for the active-learning, candidate-pool construction, and analysis workflows are:

numpy==1.26.4
pandas==2.3.3
scipy==1.17.0
scikit-learn==1.8.0
matplotlib==3.10.8
rdkit==2024.03.5
tqdm==4.67.3

Install the required Python packages using:

pip install -r requirements.txt

A complete Conda environment specification used for the molecular-simulation and analysis workflow is provided in environment.yml.

The principal software versions used for the molecular-simulation workflow were:

Python       3.11.14
HTPolyNet    1.0.9
GROMACS      2025.4
AmberTools   24.8
GAFF         1.81

EloKriti uses two primary CSV files:

1. a candidate-pool file containing candidates available for active-learning selection;
2. an evaluated-candidate file containing candidates for which MD-derived objective values are available.

The candidate-pool CSV must contain:

reaction_id
reactant_1
molecular_weight_Boltzmann_average
logP_Boltzmann_average
TPSA_Boltzmann_average
normalized_monomer_phi_Boltzmann_average
normalized_backbone_phi_Boltzmann_average

The evaluated-candidate CSV contains the same candidate information together with:

elastic_modulus
volumetric_shrinkage
ffv
status
iteration

Here:

• elastic_modulus is the MD-derived elastic modulus;
• volumetric_shrinkage is the MD-derived volumetric shrinkage;
• ffv is the fractional free volume;
• status records the outcome of the candidate evaluation;
• iteration records the active-learning iteration associated with the evaluation.

Only successfully evaluated candidates with complete values for all three optimization objectives are used to train the surrogate models.

The study-specific input CSV files used by these scripts are provided in the associated Zenodo archive. After downloading the archive, place the required CSV files in a local data/ directory at the repository root.

The construction of the 1000-candidate EloKriti screening pool is documented in:

notebooks/candidate_pool_construction.ipynb

The notebook reproduces the structural filtering, canonical-structure deduplication, and diversity-selection procedure used to construct the candidate pool. The workflow includes polymerization-mechanism and molecular-weight filtering, exclusion of BPA/BPF/BPS/BPAF structural motifs, canonical-SMILES deduplication, and Morgan-fingerprint-based Butina clustering followed by greedy MaxMin selection.

The GAFF-compatible intermediate candidate table (gaff_ok.csv) and the study-specific candidate tables required to reproduce this workflow are provided with the associated Zenodo research-data archive.

From the repository root, run:

python scripts/suggest_next.py \
  --candidate-csv candidate_pool.csv \
  --evaluated-csv evaluated_clean_no_expansion.csv \
  --output-csv next_candidates.csv \
  --top-k 10 \
  --seed 42 \
  --scalarizations 64

The script ranks the unevaluated candidates and writes the selected batch to next_candidates.csv.

After the selected candidates are evaluated using the molecular-dynamics workflow, the successful results are added to the evaluated-candidate CSV and the selection procedure can be repeated for the next acquisition round.

The hypervolume and Pareto-front analyses reported in the paper were performed retrospectively after the molecular-dynamics evaluations were completed.

To reproduce the hypervolume analysis, run:

python scripts/analyze_hypervolume.py \
  --input evaluated_clean_no_expansion.csv \
  --output-prefix al_hypervolume \
  --mc-samples 300000 \
  --seed 42

This script:

• transforms the objectives to a common maximize-all convention;
• computes the cumulative Pareto front at each active-learning iteration;
• normalizes the objectives using fixed bounds from the complete evaluated set;
• estimates dominated hypervolume using 300,000 Monte Carlo samples;
• uses random seed 42 and the same Monte Carlo sample set for all iterations.

To regenerate the active-learning progress figure, run:

python scripts/make_al_progress.py

The active-learning progress script uses evaluated_clean_no_expansion.csv as the evaluated-candidate dataset.

The hypervolume analysis is used only as a retrospective progress metric and does not contribute to candidate acquisition.

Additional analyses performed during peer review are provided in the scripts/ directory.

The following scripts are included:

scripts/pi1070_selection_benchmark.py scripts/representation_sensitivity.py scripts/scalarization_batch_benchmark.py

This analysis compares the EloKriti acquisition strategy with uniform-random selection and greedy maximin space-filling selection on the fully evaluated PI1070 polymer dataset.

Run with:

python scripts/pi1070_selection_benchmark.py \
  --dataset data/PI1070.csv \
  --seeds 50

The default evaluation budgets are:

10, 19, 27, 35, 41, 50, 57, 66, 71

The analysis reports Pareto recall and normalized hypervolume recovery.

This analysis evaluates the sensitivity of surrogate performance to the molecular representation using chronological pre-acquisition validation.

Run with:

python scripts/representation_sensitivity.py \
  --evaluated data/evaluated_clean_no_expansion.csv \
  --repeats 10

The tested representations include the original TF-IDF/SVD representation, case-preserving SMILES representations, continuous molecular descriptors, Morgan fingerprints combined with descriptors, and canonicalized SMILES representations.

This analysis evaluates sensitivity to the number of random objective scalarizations and to the batch-selection strategy using the PI1070 benchmark.

Run with:

python scripts/scalarization_batch_benchmark.py \
  --data data/PI1070.csv \
  --seeds 20

The analysis compares 16, 32, 64, 128, and 256 random linear scalarizations and additionally evaluates a diversity-aware batch-selection strategy using 64 scalarizations.

The EloKriti acquisition strategy is a ParEGO-inspired random-linear-scalarization expected-improvement procedure and is not an implementation of standard ParEGO.

Numerical outputs used in the revised manuscript and Supporting Information are archived in the associated Zenodo research-data record.

Out-of-sample surrogate performance was evaluated using chronological pre-acquisition validation. For each active-learning round, the surrogate was trained only on candidates whose MD labels were available before that round and then evaluated on candidates acquired in that round.

Run with:

python scripts/preacquisition_surrogate_diagnostics.py \
  --evaluated data/evaluated_clean_no_expansion.csv

The original Gaussian-process surrogate was additionally compared with a simple linear-regression baseline using the five continuous molecular descriptors. The same chronological pre-acquisition train/test protocol is used for both models.

python scripts/surrogate_baseline_comparison.py \
  --evaluated data/evaluated_clean_no_expansion.csv

To assess possible structural overlap across the chronological training and validation boundary, each post-initial candidate is compared only with candidates whose labels were available before its acquisition round.

python scripts/structure_similarity_audit.py \
  --evaluated data/evaluated_clean_no_expansion.csv

The repository contains the executable active-learning, Gaussian-process-surrogate, candidate-selection, Pareto-analysis, hypervolume-analysis, and candidate-pool-construction workflows associated with EloKriti.

The Python dependencies are provided in requirements.txt, while the complete Conda environment used for the molecular-simulation and analysis workflow is provided in environment.yml.

Large study-specific molecular-dynamics files and simulation artifacts are kept separately from the executable source code.

This GitHub repository contains the executable active-learning and Bayesian-optimization code associated with EloKriti.

The accompanying research-data archive will contain the study-specific evaluated-candidate tables, molecular structures, molecular-dynamics inputs and outputs, simulation topologies, construction diagnostics, and processed property data used in the revised manuscript.

These data will be deposited in a versioned Zenodo research-data record. The Zenodo DOI and corresponding versioned GitHub release will be added here when the research-data record is finalized.

If you use this code, please cite:

Active-Learning Discovery of BPA-Free Dental Resin Candidates through Backbone Screening and Network Validation

The final journal citation and DOI will be added after publication.

This project is distributed under the MIT License. See LICENSE.