ADMET Genetic Optimization Example Report

Run overview

This report was reconstructed from committed run artifacts. The build script does not rerun the genetic algorithm or ADMET-AI.

Input seeds: 12
Evaluated generation-0 seeds: 12
Invalid and deduplicated input counts: not captured
Recorded optimization generations: 4
Generated molecules per generation: 24
Total generation-log records: 108
Mutation records: 78
Crossover records: 18
Records passing filters: 83
Final candidates: 4
Random seed: not captured
Explicit stop reason: not captured; the artifacts end after generation 4
Exact dependency versions: not captured
Final candidate selection: generated molecules that pass a fresh config-based filter check, have no ADMET failure, and score strictly above their best ancestral seed; retain the highest-scoring molecule for each distinct ancestral-seed lineage

Configuration

Hard filters

Setting	Value
`mw_min`	120
`mw_max`	500
`logp_min`	-0.5
`logp_max`	5.0
`tpsa_min`	20
`tpsa_max`	140
`hbd_max`	5
`hba_max`	10
`rotb_max`	10
`sa_score_max`	5.0
`qed_min`	0.3
`risk_flags_max`	6

Scoring

Setting	Value
`qed_weight`	0.25
`admet_weight`	0.45
`sa_weight`	0.25
`property_weight`	0.05
`sa_transform`	clipped (10 - sa_score) / 9
`property_window_terms`	6

ADMET mapping

ADMET-AI requested: True
Risk threshold: 0.5
Positive endpoint keywords: hia, caco, bioavailability, solubility
Negative endpoint keywords: herg, ames, dili, cyp, pgp, p-gp, tox

Generation summary

Generation	Generated	Best score	Mean score	Passed	Population best
1	24	0.8803963	0.8216755	15	0.8822355
2	24	0.8819185	0.8466638	20	0.8822355
3	24	0.8767579	0.8277978	19	0.8822355
4	24	0.8841338	0.8495277	21	0.8841338

Final candidates

Molecule ID	Operation	Baseline seed(s)	QED	SA score	ADMET score	Total score	Delta total	Risks
`GA_g4_0085`	mutation	N-Propylbenzamide	0.7584	1.3220	0.8966	0.8841	+0.0019	high_CYP1A2_Veith;high_DILI;high_Skin_Reaction
`GA_g2_0033`	mutation	Ibuprofen	0.7675	1.4783	0.8963	0.8819	+0.0025	high_DILI
`GA_g1_0006`	mutation	Quinoline	0.5413	1.6061	0.9086	0.8274	+0.0152	high_CYP1A2_Veith;high_DILI
`GA_g1_0018`	mutation	Acetaminophen	0.6236	2.2939	0.9040	0.8267	+0.0116	high_DILI;low_PAMPA_NCATS

Interpretation

The best generation-0 score was 0.8822355 (N-Propylbenzamide). The best generated score was 0.8841338 (GA_g4_0085), a recorded increase of +0.0018983.

All 4 final candidates passed the configured filters, had successful ADMET evaluation, and improved total score relative to the best seed in their recorded ancestry. Filters were recalculated from the recorded properties and config.yaml rather than accepted solely from the log flag. The highest-scoring molecule represents each distinct ancestral-seed lineage. This deterministic rule does not claim an unrecorded fingerprint-diversity calculation.

Final-candidate risk flags were high_CYP1A2_Veith (2), high_DILI (4), high_Skin_Reaction (1), low_PAMPA_NCATS (1). These flags are model-derived triage signals, not observed toxicology outcomes.

Limitations and next steps

All ADMET values are model predictions and have not been experimentally validated.
The low-level mutation and crossover operators do not establish chemical feasibility.
SA score is a heuristic and does not guarantee a practical synthesis route.
The endpoint aggregation and keyword mapping are task-specific heuristics.
Reproduce the run with captured package versions and a fixed random seed.
Review top structures for medicinal-chemistry liabilities and scaffold diversity.
Validate prioritized endpoints with independent models and experimental assays.
Assess synthetic routes before advancing a candidate.