A suite of open-source tools that bring modern AI/ML analysis to biophysics: protein design and stability, MD kinetics, cheminformatics/ADMET, cryo-EM QC, electrophysiology, and grounded literature search. Each is validated against real benchmarks and runs CPU-only, no GPU, no license. Many work right in your browser; the rest run on a free hosted backend (with a downloadable example when it's offline). Paste your sequence, molecule, recording, or micrograph and get an answer.
Name any researcher → it reads their actual papers (full-text where open-access), hybrid BM25+dense retrieval with LLM rerank, and answers ask / summarize / compare grounded with DOI citations.
MD trajectory → mechanism: tICA + Markov State Model (implied timescales, PCCA+ macrostates, MFPT), native-contacts, DSSP timeline, contact-map diffs. Chunked large-traj IO, 11 tests. Demo: alanine-dipeptide C7eq↔αR.
13 ADMET endpoints (solubility, hERG, CYPs, Ames, Caco-2, DILI…) with random + scaffold splits, PAINS/Brenk alerts, applicability domain + confidence. 18 tests. Flags Astemizole's hERG risk; calibrated on hard endpoints.
Per-residue biophysics from sequence: TM, aggregation, disorder, charge/pI, signal peptides, coiled-coil. ML layer beats raw scales (WALTZ-DB/DisProt/UniProt); validated vs UniProt annotations (disorder F1 0.86). 21 tests.
Cryo-EM micrograph QC: 2D astigmatic CTF fit (defocus + astigmatism + resolution), ice/drift detection, PASS/FLAG/REJECT. Known-defocus recovery mean error 5.1 Å; validated on EMPIAR-10025 & 10061. 14 tests.
Full intrinsic ephys feature set + RS/FS/IB classifier + QC. 100% cross-validated against eFEL, validated on real Allen human neurons (rheobase matches exactly). Real pyABF/NWB data, 31 tests.
ΔΔG prediction + stabilizing-mutation design. Benchmarked on full S669 vs FoldX/DDGun/ThermoMPNN (in-band), with SOTA antisymmetry (Ssym bias ≈0). 68 features, grouped-CV, 11 tests, metrics, GPU-free.
I mapped the whole field as a data scientist before picking what to build: 1,353 institutions, ~3,900 advisors, funding flows, research gaps, and 107 fundable software opportunities.
Every subfield has a great open model: Boltz-2, AlphaFold3, RFdiffusion, MACE, Cellpose. Almost no lab has a serving layer, agent orchestration, or analysis pipeline around it. These tools close that gap: model → usable answer, free and open, on hardware any lab already has. Each is open-source under MIT. Use it, self-host it, or read how it works.