Skip to content

for-researchers.md

Latest commit

 

History

History
95 lines (75 loc) · 4.68 KB

for-researchers.md

File metadata and controls

95 lines (75 loc) · 4.68 KB
Raw

For researchers

Quick orientation for cheminformaticians, immunologists, mast-cell biologists, and ML / AI folks who want to scrutinize, fork, or extend this work.

What this repo is

A reproducible, MIT-licensed in silico pipeline for hypothesis generation in MCAS / MCAD. Five standardized experiments are published as of the latest commit; each carries its own pre-registration-style report:

ID Method Result
EXP-001 RDKit BRICS + bioisosteric enumeration + warhead grafting from 7 natural ITC seeds 113 ranked SFN-class candidate analogs
EXP-002 Tanimoto (Morgan r=2, 2048 bits) vs curated reference ligands per target 157 compounds × 8 MCAS targets
EXP-003 13 cysteine-reactive SMARTS + KEAP1 pharmacophore filter 79 compounds carry warhead + pass filter
EXP-004 RandomForest on Morgan FP, trained on PyTDC hERG / AMES / BBB Valid AUC 0.89–0.91
EXP-005 Composite (evidence + target + warhead + safety + drug-likeness) Ranked rescue / maintenance / remission CSVs

Every result CSV is in outputs/ and rebuilds from a single seeds file + 5 scripts.

Reproduce the whole pipeline in 5 minutes

git clone https://github.com/mrdulasolutions/MCAS.Opensource.git
cd MCAS.Opensource
python -m venv .venv && source .venv/bin/activate
pip install -e . PyTDC scikit-learn 'setuptools<81'

python scripts/build_compound_library.py    # PubChem fetch + RDKit validation
python scripts/validate_smiles.py
python scripts/generate_sfn_analogs.py      # SFN-class analog generation
python scripts/score_warheads.py            # covalent warhead detection
python scripts/score_against_targets.py     # ligand-based screening (8 targets)
python scripts/run_qsar.py                  # PyTDC ADMET QSAR
python scripts/rank_hypotheses.py           # composite ranking + hypothesis-doc updates

Wall-clock: ~3 minutes on a CPU laptop after the one-time PyTDC download.

Things we'd love independent eyes on

  1. Composite weighting in EXP-005 — author-chosen. Run a sensitivity analysis (±50% per weight) and tell us where the ranking is fragile.
  2. Reference ligand sets in scripts/score_against_targets.py — we missed compounds. Add what we missed.
  3. QSAR baselines — RandomForest on Morgan FP is a strong baseline, not state-of-the-art. ChemProp / DeepChem GraphConv / 3D models likely add 1–3 AUC points. Train a better one and we'll merge it.
  4. Covalent reactivity ranking — EXP-003's SMARTS detector is binary. xTB / DFT electrophilicity indices would let us rank reactivity, not just detect presence. We don't have the compute; you might.
  5. KEAP1 docking — Kelch-domain crystal structures (PDB 4L7B, 5FNQ, 6T7V) exist. Vina/smina or CovDock against Cys-151 for the top-50 ranked compounds would convert ligand similarity into a structure-based estimate.
  6. Mast-cell-specific bioactivity data — pull every ChEMBL assay tagged "β-hexosaminidase release," "LAD2," "histamine release," or "tryptase release" and train a true mast-cell-stabilizer predictor.
  7. Selectivity / polypharmacology score — promiscuous binders should be penalized. We don't have a kinome / NHR / GPCR-panel proxy in here.

Datasets you can cite directly

Citation

See CITATION.cff. Cite the repo + the commit hash you used; we re-tag with a Zenodo DOI quarterly.

How to contribute back

  1. Fork.
  2. Make your change — ideally as a new experiment under experiments/EXP-NNN-*.md so other researchers can replicate.
  3. Open a PR with an explicit hypothesis + falsifiability statement.
  4. CI runs SMILES validation on data/ changes.

If you want to set up a formal academic collaboration (data sharing agreement, MTA logistics, joint authorship), see for-academia.md.

What we will refuse

  • Patent-encumbered contributions.
  • Closed-data validation (we will publish our predictions; you publish yours, even if negative).
  • Predictions claimed as clinical recommendations.