CellSim

Open-source, non-AI, physics-first drug-discovery triage platform.

CellSim is the in-silico front of a wet-lab shortlist. Drop in a receptor PDB and a list of SMILES, walk away for coffee, come back to a ranked CSV with ΔG ± CI, pocket-fit flags, ADMET descriptors, and one-page drug profile dashboards for your top hits.

Every method is physics-grounded (classical force fields, semi- empirical QM, alchemical free energy, flux-balance analysis) and every rate constant cites a PMID or a cached physics calculation. No neural scoring, no learned potentials, no black-box surrogates — see MISSION.md for the discipline and GOAL for the 5-campaign roadmap.

Status: Campaign 1 (Atomic → Molecular Foundation) is in progress. The pre-restart HeLa / p53 / cisplatin cell prototype is frozen under OLD/ as a regression snapshot (still builds; still passes its 8 headless validators).

Quickstart — first docking run in 5 minutes

# 1. Install the environment (one-time).
conda env create -f environment.yml      # or mamba
conda activate cellsim

# 2. Run the end-to-end biologist workflow on a bundled cocrystal.
python -m src.dock.batch \
    --smi benchmarks/dock/1stp_batch_5.smi \
    --receptor benchmarks/dock/1stp.pdb \
    --out-csv /tmp/run/report.csv \
    --mc 4 --profile-top-k 3 \
    --crystal-pdb benchmarks/dock/1stp.pdb \
    --crystal-resname BTN

Output:

RANK  NAME                     TRIAGE         ΔG(kcal)   K_d        POCKET  STRAIN       Ro5  QED   logS
   1  biotin_TRUE_BINDER       follow_up         -7.44    3.5 µM    ✓       acceptable  ✓    0.49  -1.53
   2  ibuprofen_negative       deprioritise      -7.36    4.1 µM    ?       good        ✓    0.82  -3.09
   3  aspirin_negative         drop              -6.66   13.0 µM    ✓       acceptable  ✓    0.55  -1.99
   4  acetaminophen_negative   drop              -6.31   23.8 µM    ?       acceptable  ✓    0.59  -1.97
   5  caffeine_negative        drop              -5.48   95.7 µM    ?       good        ✓    0.54  -0.87

The TRIAGE column (follow_up / review / deprioritise / drop) is the one-decision column: CellSim synthesises ΔG + pose-strain + PoseBusters + ADMET into one verdict with a paste-ready reason string. Wet-lab users read one column, not five booleans.

Plus profile_01_biotin_TRUE_BINDER.png, profile_02_...png, profile_03_...png — one-page dashboards showing 3D + charges, predicted CYP3A4 sites-of-metabolism, HOMO/LUMO, and the full Lipinski / QED / logS datasheet.

If you do not know the binding-site coordinates of your target, omit --center / --box and CellSim auto-detects pockets via fpocket (the canonical non-ML geometric pocket finder).

What CellSim can do today

Layer	What it does	Status
1.1 Chem	SMILES → OpenFF-parametrised system (AM1-BCC charges)	✅ 9/10 full tier, 10/10 RDKit tier
1.2 MD	Classical Langevin MD, solvated protein loader (AMBER14 + TIP3P)	✅ 1 ps ubiquitin Cα RMSD 0.74 Å
1.3 Docking	Vina + Meeko + PoseBusters + fpocket auto-site	✅ mini-bench 2/3 canonical gate
1.3 FEP	Alchemical ΔG_hyd + ΔG_bind (DDM); cellsim fep-binding {dg,ddg,bench,validate} + fep-report + bench-all	✅ pipeline complete end-to-end; 50+ smoke tests; Milestone A + B numbers pending GPU
1.4 Quantum	xTB GFN2 single-point + CYP3A4 SoM predictor (BDE)	✅ 10/10 sane + 3/3 SoM smoke
1.5 Coarse-grained	Martini 3 membrane / bilayer MD	⏳ scaffold only
1.6 UQ	Monte-Carlo / Sobol / split-conformal for ΔG bounds	✅ triad shipped
1.7 Blind harness	PDBBind scale gate + red-team slot	⏳ 3-cocrystal mini-bench shipped; PDBBind scale pending
x-cut cache	SQLite physics-prior memoisation (AM1-BCC, Vina, xTB)	✅ shipped; wired into Layers 1.1 / 1.3 / 1.4

Cross-cutting UX:

• src/dock/batch.py — one-command ranked screen with MC error bars and optional --profile-top-k auto-dashboards.
• src/chem/profile.py — six-panel per-compound profile combining 3D charges, SoM predictions, HOMO/LUMO, BDE chart, and Ro5 / QED / logS callouts.
• src/chem/admet.py — Lipinski / TPSA / QED / ESOL solubility (all published formulae; no ML).
• src/dock/pocket_detect.py — auto-binding-site detection so any receptor PDB works out-of-box.
• src/uq/dock_mc.py — honest ΔG ± CI from N-seed Monte-Carlo.
• src/cache/ — content-addressed SQLite cache for every expensive physics call; pass --cache /path/to/c.sqlite to cellsim dock and repeat runs short-circuit identical (ligand, receptor, box, seed) tuples (~1000× per-compound Vina speedup, ~19× end-to-end on a small screen).

Non-AI discipline

CellSim is strictly physics-first and non-AI by design. This is load-bearing, not a preference. Every prediction must trace to a physics calculation or a literature-cited empirical formula. See MISSION.md §"No black-box / no AI surrogates" for the five ground rules and the narrow "ML as accelerator only" exception clause.

Explicitly excluded:

• ML potentials (MACE / NequIP / OrbNet / Allegro) as the force path.
• GNINA CNN-scored docking as the primary evidence (Vina only; GNINA may ship as an explicitly labeled fast-guess alongside).
• "Deep ensembles" for UQ. Sobol + Monte-Carlo + MAPIE conformal (post-hoc, non-parametric) only.

Validation that runs on every PR

.github/workflows/smoke.yml provisions the cellsim conda env and runs 34 gates in ~15 min, grouped by layer:

• Layer 1.1 chem — 10-drug parametrise, ADMET, AM1-BCC cache.
• Layer 1.2 MD — 3-drug vacuum MD, 1UBQ solvate + 1 ps MD.
• Layer 1.3 docking + triage — Meeko prep, 1STP re-dock, fpocket detect, 3-cocrystal mini-bench, refine, Vina cache, pose SDF/PDB export, off-target selectivity, strain diagnostic (UFF-ensemble), strain-gate top-pose promotion, triage rule table, shortlist filter, triage-PNG dashboard, kinase-receptor heads-up, CYP3A4 DDI-risk strain-downgrade.
• Layer 1.4 quantum — 10-drug xTB, 3-drug CYP3A4 BDE SoM, heme-accessibility SoM, PySCF DFT single-point.
• Layer 1.6 UQ — MC-dock, conformal, streptavidin + EGFR calibrations.
• Cross-cut — cache round-trip.

A regression in any gate blocks merge. See BENCHMARKS.md for the numeric targets each gate enforces.

Layout

src/
  chem/      Layer 1.1  parametrise, ADMET, profile dashboard
  md/        Layer 1.2  OpenMM MD driver, PDBFixer protein loader
  dock/      Layer 1.3  Vina + Meeko + PoseBusters + fpocket + batch
  quantum/   Layer 1.4  xTB GFN2, CYP3A4 SoM predictor
  cg/        Layer 1.5  Martini 3 (scaffold; not populated yet)
  cache/     cross-cut  SQLite + HDF5 physics-prior cache (scaffold)
  uq/        Layer 1.6  Monte-Carlo dock, Sobol (pending)
  bridge/    cross-cut  Layer-1 → Layer-2 rate-law emitter (future)
  core/      physics-neutral RNG, telemetry, constants
benchmarks/
  chembl/    10 canonical drugs
  md/        1UBQ ubiquitin
  dock/      1STP / 1M17 / 3PTB cocrystals + mini_bench.yaml
tests/       one folder per src/ module, same smoke pattern
OLD/         frozen pre-restart Campaign-2 prototype (builds; passes
             its own 8 headless validators as regression snapshot)
docs/        strategic plan, professor debriefs, campaign scope
scripts/     CLI utilities (fetch_pdb, fetch_chembl_sample, ...)

How good is it?

See BENCHMARKS.md for the full scorecard — every current number (pose-recovery %, Spearman on two receptor families, cache speed-ups, CI gate list) with honest caveats and per-row reproducers.

Headline numbers today:

• 3-cocrystal mini-bench: 2/3 = 67 % pose recovery at canonical top-3 < 2 Å gate.
• Streptavidin calibration: Spearman ρ = 1.00 across 14 orders of magnitude of K_d (but Pearson r = 0.98 hides the fact that Vina's absolute ΔG saturates on tight binders — MAE 4.99 kcal/mol).
• Trypsin calibration: MAE = 0.91 kcal/mol on benzamidine analogs (Vina's absolute is accurate here), but Spearman ρ = 0.09 within the narrow 4 kcal/mol window (noise floor).
• End-to-end cache speed-up: 19× on a 5-compound batch rerun.

How to cite

Every prediction in CellSim carries its method provenance (tool version, force field, seed, search parameters). To cite a run, include the commit SHA and the provenance block from the relevant result envelope in your publication's methods.

License

MIT. See LICENSE.