admet_genetic

ADMET Genetic Optimization

Use this skill to build and run a molecular optimization loop from seed SMILES. The target artifact is a ranked set of optimized candidate molecules with auditable lineage, scores, and report artifacts.

The sidecar deliberately does not provide a fixed GA engine. The agent must assemble and tune mutation, crossover, evaluation, filtering, and selection for the user's objective. kernel.py provides reusable molecule normalization, ADMET aggregation, lineage validation, and result visualization.

Special Reminder

In this skill, when see references/<file_name>.md is suggested, use host call to retrieval the complementary material.

host.skills.read("admet_genetic", "references/<file_name>.md")

Prerequisites

conda create -n admet-sa-ga python=3.11 -y
conda activate admet-sa-ga
python -m pip install pandas pyyaml matplotlib rdkit
python -m pip install admet-ai  # depends on torch; installation/import may take time

After creating the environment, select it with host.env.use("admet-sa-ga") before importing this skill's sidecar. Switching environments restarts the session kernel, so switch before constructing the pipeline. An in-kernel pipeline is fine while you are still exploring; see Formal runs below for what a run has to leave behind.

See references/admet.md for ADMET-AI installation details, endpoint behavior, runtime notes, and troubleshooting.

Data Contracts

For molecular representation, expected fields, candidate recording and lineage logging, see references/data_contracts.md. Must view these contracts before running the main pipeline.

Core Workflow

1. Collect user-provided seed molecules or uploaded files and normalize them into a CSV input. The CSV should contain smiles; include molecule_id when stable user-facing IDs are available, otherwise synthesize deterministic IDs.
2. Standardize each input using standardize_smiles(...), then use canonicalize_smiles(...) from kernel.py where a strict canonical string is needed. Molecule ID and canonical SMILES must be one-to-one for all logged records.
3. Design a genetic algorithm that includes molecular mutation and crossover. Match population size, generation count, operators, filters, and scoring weights to the user’s problem scale and constraints. For a starter design and implementation choices, see references/ga.md.
4. Evaluate each valid molecule with RDKit descriptors, QED, SA-Score, and ADMET predictions. Aggregate ADMET endpoints into admet_score and admet_risk_flags; preserve raw endpoint outputs. See references/data_contracts.md for required evaluation fields.
5. Apply hard filters, compute total score, select diverse candidates by Morgan fingerprint similarity, and update the population. See references/ga.md for starter designs.
6. Assess whether the final candidates improve on the seeds and satisfy the user’s requirements. If they do not, adjust GA parameters, mutation/crossover operators, filters, or scoring weights, then rerun the internal GA workflow before finalizing output.
7. Output final candidates, logs, report, visualization dashboard, and any other produced artifacts. See Artifacts for log schema and lineage rules.