Repellent/Insecticide Modeling Plan With Staged Quantum Descriptors

Summary

• Build the project in two stages. V1 is a chemically curated, probability-calibrated repellency model using standard cheminformatics features. V2 adds electronic-structure descriptors after V1 has identified the most informative compounds.
• Use the paper’s strategy as the template: feature-tokenized transformer for tabular molecular descriptors, frozen-backbone fine-tuning on small labeled data, calibrated uncertainty with cross-validation plus conformal prediction, and post-hoc interpretability with SHAP/attention.
• Do not implement a 3-class softmax model initially. non-repellent, repellent, and insecticidal are not cleanly exclusive, and the LifeChemicals insecticide file is currently an unlabeled screening pool, not supervised insecticide ground truth.
• Treat “effectiveness” as a calibrated probability of activity in V1. Only treat effectiveness as potency after real numeric labels such as % repellency, % mortality, LC50, or LD50 are added.

Implementation Changes

• Create a reproducible data-curation pipeline that reads the three SDFs, strips salts/counterions, canonicalizes structures, keeps one canonical record per InChIKey, preserves source metadata, and flags malformed/placeholder records in the decoy file for exclusion.
• Define the first supervised endpoint as repellent_active with 1 from the repellent SDF and 0 from curated non-repellent decoys. Use non-repellent as the complement of this probability, not as a separate head.
• Restrict the first production model to the dominant assay scope: Aedes aegypti repellency. Keep the other species as metadata for later multi-domain expansion, not mixed into the first supervised target.
• Build V1 with two model families trained on the same scaffold split: a strong baseline using Morgan fingerprints (radius=2, 2048 bits) plus RDKit 2D descriptors, and an FTTransformer over curated tabular features. Select the winner by scaffold-split PR-AUC, ROC-AUC, MCC, Brier score, and calibration.
• Pretrain the FTTransformer backbone on all curated molecules from the three SDFs using self-supervised masked-feature reconstruction. Freeze the backbone and fine-tune only the head on the labeled repellency task, mirroring the paper’s transfer-learning approach.
• Export model outputs in a single prediction table with: compound_id, canonical_smiles, scaffold_id, p_repellent, p_non_repellent=1-p_repellent, prediction_interval, ood_score, and split_id.
• Add conformal uncertainty exactly as a first-class output. Use scaffold-aware 5-fold cross-validation for out-of-fold predictions, then fit calibration plus conformal intervals on those predictions so each molecule gets both a probability and an uncertainty band.
• Create a quantum-descriptor stage for V2 that runs after V1 and only on a selected subset. The default subset is 300 labeled molecules chosen by class balance and scaffold diversity (150 repellent, 150 non-repellent) plus 200 unlabeled LifeChemicals molecules chosen from 100 highest p_repellent and 100 highest uncertainty.
• Use GFN2-xTB as the broad quantum layer. For each selected molecule, generate conformers with ETKDG, MMFF-optimize, keep the lowest-energy conformer, optimize with xTB, and extract at minimum: HOMO, LUMO, gap, dipole, total energy, partial charge mean/std/max, polarizability if available, and simple frontier-orbital-derived reactivity indices.
• Use DFT only as a validation/high-fidelity layer on a smaller subset of 100 molecules drawn from the xTB set with class balance and chemical diversity. Run single-point DFT on the xTB-optimized geometry to benchmark whether xTB descriptors are directionally reliable before expanding DFT usage.
• Train V2 as an ablation study, not a blind replacement. Compare three feature sets on the same labeled quantum subset: cheminformatics-only, quantum-only, and fused features. Promote the fused model only if it improves scaffold-split discrimination and calibration.
• Keep the LifeChemicals library unlabeled in training. Use it for self-supervised pretraining and later screening/ranking only. Do not train an insecticidal head until external insecticide assay labels are added from ChEMBL, PubChem, or literature.
• Reserve the later V3 interface now: once insecticide labels exist, extend the model to a multitask setup with independent outputs p_repellent and p_insecticidal, each with its own calibration and uncertainty, instead of collapsing them into one multiclass label.

Public Interfaces and Artifacts

• Standardize on four project artifacts: curated_molecules.parquet, split_manifest.json, model_predictions.parquet, and quantum_descriptors.parquet.
• curated_molecules.parquet should contain identifiers, canonical structure fields, source dataset, label fields, assay/species metadata, scaffold, and a qc_status column for excluded/problematic molecules.
• quantum_descriptors.parquet should key by compound_id and store conformer provenance, method (xtb or dft), charge/multiplicity, convergence status, and the extracted electronic descriptors.
• model_predictions.parquet should expose only calibrated probabilities and uncertainty outputs that downstream ranking or simulation selection will consume.

Test Plan

• Verify data curation is deterministic: repeated runs produce identical canonical SMILES, label counts, and exclusion counts.
• Verify no leakage: no Bemis-Murcko scaffold appears in both train and validation folds.
• Benchmark V1 against random and majority baselines; require materially better PR-AUC, MCC, and Brier score.
• Check probability calibration with reliability plots and expected calibration error; reject any model with good AUC but poor calibration.
• For V2, evaluate gain only on the same labeled subset with quantum descriptors; require improvement in at least one ranking metric and one calibration metric before adopting fused features.
• For the DFT benchmark subset, require that xTB-derived HOMO/LUMO gap rankings correlate strongly enough with DFT to justify continued xTB use as the bulk descriptor engine.
• Run interpretability on the final selected model and confirm important features are chemically plausible rather than assay/source artifacts.

Assumptions and Defaults

• The LifeChemicals insecticide SDF is treated as an unlabeled candidate library until real insecticide activity labels are added.
• The first supervised endpoint is Aedes aegypti repellency because it is the dominant labeled target and gives the cleanest initial task.
• “Effectiveness” in the first release means calibrated probability of repellency, not potency. Potency modeling is deferred until numeric assay labels exist.
• The decoy SDF contains some malformed or placeholder structures; those will be excluded during curation rather than force-fit into training.
• The default quantum workflow is xTB first, DFT second, and quantum descriptors are used to improve a screened subset before any attempt to scale them to the full library.