chemaudit-structure-filter

name	chemaudit-structure-filter
description	Filter generative-chemistry outputs through ChemAudit's 6-stage multi-stage funnel with a composite 0-1 scorer and REINVENT 4-compatible scoring endpoint. Use when user says "filter molecules", "drug-like filter", "structure filter", "filter generative output", "REINVENT scoring", "REINVENT Component", "lead-like filtering", "fragment filter", "funnel these SMILES", "novelty filter", or "score generative batch". Four built-in presets (drug_like, lead_like, fragment_like, permissive). Sync for ≤1000 SMILES, async for larger.
license	MIT
allowed-tools	Bash(curl:*) Bash(python:*)
compatibility	Requires a running ChemAudit instance. Async batch mode requires Celery workers.
metadata	{"author":"Kohulan Rajan","version":"1.0.0","mcp-server":"chemaudit","category":"cheminformatics","tags":["generative-chemistry","filtering","drug-likeness","REINVENT","virtual-screening"]}

ChemAudit Structure Filter

Overview

Multi-stage funnel for generative-chemistry outputs and virtual screening hits. Six sequential stages, each of which can reject a molecule independently:

1. parse — RDKit parse + sanitization.
2. valence — valence validity check.
3. alerts — structural alert catalogs (PAINS, Brenk, Kazius, NIBR — configurable).
4. property — MW, LogP, TPSA, rotatable bonds, rings within bounds.
5. sa_score — synthetic accessibility score ≤ threshold.
6. dedup — InChIKey-based deduplication.

Optional 7th stage: 7. novelty — Tanimoto similarity vs ChEMBL (gated by enable_novelty).

Counts how many molecules enter and leave each stage (funnel diagram data).

Paired with a composite 0-1 scorer that blends validity / QED / alert-free / SA into a single reward, useful for generative-model reinforcement learning loops. Invalid SMILES score null, never 0.0 (D-14 / Pitfall 4) so generative agents can distinguish "unparseable" from "parseable but terrible".

REINVENT 4 Component API is supported out of the box.

Four built-in presets

See references/preset-configs.md for full threshold details.

Preset	MW	LogP	TPSA	RotB	Rings	SA	Alerts
drug_like	200–500	-1 to 5	≤140	≤10	—	≤5.0	PAINS + Brenk + Kazius
lead_like	200–350	-1 to 3.5	≤140	≤7	—	≤4.0	PAINS + Brenk + Kazius + NIBR
fragment_like	100–300	-1 to 3	≤100	≤3	≤3	≤3.0	PAINS only
permissive	100–800	-5 to 8	≤200	≤15	—	≤7.0	none

Weight vectors for the composite score differ per preset (D-15):

Preset	validity	qed	alert_free	sa
drug_like	0.3	0.3	0.2	0.2
lead_like	0.2	0.4	0.2	0.2
fragment_like	0.2	0.2	0.3	0.3
permissive	0.4	0.3	0.1	0.2

Sync vs async split (D-22)

• POST /structure-filter/filter with ≤ 1000 SMILES → sync FilterResponse.
• POST /structure-filter/filter with > 1000 SMILES → returns StructureFilterBatchUploadResponse with job_id, task runs on Celery.
• POST /structure-filter/batch/upload (file upload) → always async.

Workflow

1. Filter with a preset (sync, ≤1000 SMILES)

curl -sS -X POST http://localhost:8000/api/v1/structure-filter/filter \
  -H 'Content-Type: application/json' \
  -d '{
    "smiles_list": ["CCO", "c1ccccc1", "CC(=O)Oc1ccccc1C(=O)O", "[Ni]"],
    "preset": "drug_like"
  }'

Response (synchronous):

{
  "input_count": 4,
  "output_count": 1,
  "stages": [
    {"stage_name": "parse", "stage_index": 1, "input_count": 4, "passed_count": 4, "rejected_count": 0, "enabled": true},
    {"stage_name": "valence", "stage_index": 2, "input_count": 4, "passed_count": 4, "rejected_count": 0, "enabled": true},
    {"stage_name": "alerts", "stage_index": 3, "input_count": 4, "passed_count": 3, "rejected_count": 1, "enabled": true},
    ...
  ],
  "molecules": [
    {"smiles": "CCO", "status": "rejected", "failed_at": "property", "rejection_reason": "MW 46.07 below minimum 200"},
    {"smiles": "CC(=O)Oc1ccccc1C(=O)O", "status": "passed", "failed_at": null, "rejection_reason": null},
    ...
  ]
}

2. Filter with explicit config (overrides preset)

curl -sS -X POST http://localhost:8000/api/v1/structure-filter/filter \
  -H 'Content-Type: application/json' \
  -d '{
    "smiles_list": ["CCO", "c1ccccc1C(=O)O"],
    "config": {
      "min_mw": 100, "max_mw": 400,
      "min_logp": -1, "max_logp": 4,
      "max_tpsa": 120, "max_rot_bonds": 8,
      "max_rings": null, "max_sa_score": 4.5,
      "use_pains": true, "use_brenk": true, "use_kazius": false, "use_nibr": false,
      "enable_novelty": false, "novelty_threshold": 0.85,
      "weight_validity": 0.25, "weight_qed": 0.35,
      "weight_alert_free": 0.2, "weight_sa": 0.2
    }
  }'

If both preset and config are given, preset wins.

3. Composite score per molecule (0-1)

curl -sS -X POST http://localhost:8000/api/v1/structure-filter/score \
  -H 'Content-Type: application/json' \
  -d '{
    "smiles_list": ["CCO", "c1ccccc1C(=O)O", "invalid_smiles"],
    "preset": "drug_like"
  }'

Response: {"scores": [0.412, 0.871, null]}. null means unparseable — preserve that in downstream ML pipelines rather than coercing to 0.

4. REINVENT 4 Component API

The endpoint accepts the REINVENT 4 input contract verbatim (raw list, not wrapped in a model):

curl -sS -X POST "http://localhost:8000/api/v1/structure-filter/reinvent-score?preset=drug_like" \
  -H 'Content-Type: application/json' \
  -d '[
    {"input_string": "CCO", "query_id": "q1"},
    {"input_string": "c1ccccc1C(=O)O", "query_id": "q2"},
    {"input_string": "invalid", "query_id": "q3"}
  ]'

Response:

{
  "output": {
    "successes_list": [
      {"query_id": "q1", "output_value": 0.412},
      {"query_id": "q2", "output_value": 0.871}
    ]
  }
}

Critical invariant: invalid SMILES are omitted, not scored 0.0 (REINVENT expects this — see Pitfall 4 / D-14). Don't insert zeros yourself.

REINVENT 4 config.toml snippet:

[[parameters.component]]
type = "ExternalProcess"
command = "curl -X POST http://chemaudit:8000/api/v1/structure-filter/reinvent-score?preset=drug_like -H 'Content-Type: application/json' -d @-"

5. Async batch (file upload, always async)

curl -sS -X POST http://localhost:8000/api/v1/structure-filter/batch/upload \
  -F "file=@generated_library.csv" \
  -F "preset=lead_like"

Or with explicit JSON config:

curl -sS -X POST http://localhost:8000/api/v1/structure-filter/batch/upload \
  -F "file=@library.sdf" \
  -F 'config={"min_mw":200,"max_mw":450,...}'

6. Track async progress

WebSocket:

const ws = new WebSocket(`ws://localhost:8000/ws/structure-filter/${job_id}`);

Poll:

curl -sS http://localhost:8000/api/v1/structure-filter/batch/<job_id>/status

Response: {job_id, status, progress, current_stage}.

7. Async results

curl -sS http://localhost:8000/api/v1/structure-filter/batch/<job_id>/results

Returns the same FilterResponse shape wrapped in {job_id, status, result}.

8. Download

# Plain text — one passed SMILES per line
curl -sS http://localhost:8000/api/v1/structure-filter/batch/<job_id>/download/passed_txt \
  -o passed.txt

# Full CSV — every molecule with status, failed_at, rejection_reason
curl -sS http://localhost:8000/api/v1/structure-filter/batch/<job_id>/download/full_csv \
  -o results.csv

Examples

Example 1 — "Filter 10K generated SMILES and save the passing ones"

1. POST /structure-filter/batch/upload with the file and preset=drug_like → get job_id.
2. Track via WS or poll /status.
3. GET /structure-filter/batch/<job_id>/download/passed_txt → plain-text file of passed SMILES.

Example 2 — "Integrate with REINVENT"

Wire the /reinvent-score endpoint as an ExternalProcess component. Choose the preset via query string. REINVENT's query_id round-trip is preserved; agents skip invalid SMILES naturally.

Example 3 — "Fragment-like filter on a small library"

1. POST /structure-filter/filter with smiles_list (≤1000) and preset="fragment_like".
2. Read molecules[], keep those with status="passed".
3. Render the funnel: stages[] has input_count / passed_count / rejected_count per stage.

Example 4 — "Turn on novelty filtering vs ChEMBL"

Set enable_novelty: true and novelty_threshold: 0.85 in an explicit config (no preset enables novelty by default):

curl -sS -X POST http://localhost:8000/api/v1/structure-filter/filter \
  -H 'Content-Type: application/json' \
  -d '{"smiles_list": [...], "config": {"min_mw": 200, "max_mw": 500, ..., "enable_novelty": true, "novelty_threshold": 0.85}}'

Molecules with Tanimoto > 0.85 to any ChEMBL compound are rejected at the novelty stage.

Rate limits

• /structure-filter/filter: 20/min.
• /structure-filter/score, /structure-filter/reinvent-score: 30/min.
• /structure-filter/batch/upload: 3/min.
• /structure-filter/batch/<job_id>/status: 60/min.
• /structure-filter/batch/<job_id>/results: 30/min.
• /structure-filter/batch/<job_id>/download/<format>: 10/min.

Troubleshooting

400 "Unknown preset ''"

Valid presets: drug_like, lead_like, fragment_like, permissive. Case-sensitive.

400 "No valid SMILES found in input"

File parsed but every row errored. For CSV, ensure the column is named exactly SMILES. For SDF, ensure molecules have valid MOL blocks.

Score 0.0 vs null — which is it?

null = parse failure or SMILES rejected before scoring. 0.0 = parsed and scored but every component was 0. Critical distinction for generative models; never collapse them.

REINVENT returns fewer items than I sent

Expected — invalid SMILES are omitted from successes_list, not zero-scored. Match by query_id, not list position.

Async job_id found in /status but no results

Celery worker crashed. Check docker compose logs worker and re-upload. Redis TTL is 1 hour on metadata.

Funnel shows 0 passed but all stages have 100% input

All molecules pass individual stages but none make it through dedup. Means the input contains only duplicates (identical InChIKeys). Deduplicate before filtering, or disable dedup by adjusting the pipeline.

High rejection count at alerts stage for a fragment-screen

fragment_like preset has only PAINS enabled. If you want Brenk filtering for fragments, use an explicit config rather than the preset.

Further reading

• references/preset-configs.md — exhaustive threshold tables and weight vectors.
• chemaudit-qsar-ready — different goal: ML-ready curation with provenance, not pass/fail filtering.
• chemaudit-molecule-validation — single-molecule scoring when you need per-check detail.