SMA Research Platform

Evidence graph for Spinal Muscular Atrophy

Biology-first target discovery
Christian Fischer / Bryzant Labs
1,145Targets
453Trials
60Drugs
7Datasets
34,514Sources
43,071Claims
46,973Evidence
29,625Hypotheses
announcementApr 7, 2026· SMA Research Platform

Pipeline v2.2 — Post-Review Upgrade with 6 Improvements

#pipeline#v2.2#upgrade#PocketXMol#metadynamics#QM/MM#benchmarking

Pipeline v2.2 — Upgraded After 3-LLM External Review

Following critical review by GPT-4o, web evidence search, and industry benchmarking against Insilico Medicine, Recursion, Isomorphic Labs, and Schrodinger, we identified 6 gaps and upgraded the pipeline.

What Changed (v2.1 → v2.2)

Stage Before After
Generation GenMol NIM PocketXMol (Cell 2026, 82.5% docking accuracy)
MD duration 20ns 100ns + metadynamics enhanced sampling
Free energy MM-PBSA only MM-PBSA + FEP-SPell + QM/MM
Validation Boltz-2 + Chai-2 (100x antibody hit rate improvement)
Selectivity DiffDock panel Redundancy filter (score vs 5+ off-target kinases)
Knowledge CORTEX raw PMID validation gate (catches hallucinated citations)

SMA-Specific Corrections

  1. Fasudil mechanism clarified: Benefits in SMA mice are muscle-mediated, NOT neuroprotective (Bowerman 2012). Does not prevent motor neuron loss. However, ROCK-ALS Phase 2 (Lancet Neurology) shows motor unit preservation in ALS patients.
  2. ROCK2 selectivity needed: Fasudil inhibits ROCK1+ROCK2 non-selectively. PocketXMol ROCK2-selective compound generation queued (300 molecules).
  3. VHH BBB strategy: Nanobodies require TfR1-binding fusion domain for blood-brain barrier transcytosis.

Industry Benchmarking

Pipeline v2.2 is competitive with the Physics+ML hybrid archetype (Schrodinger/Nimbus). Top-20% globally for an independent setup running at $78/day on Vast.ai GPU fleet.

Unique advantages: quality gates at every stage, negative results published openly, PMID validation on all AI-generated citations, open-source evidence graph.

Full Pipeline (13 stages)

Stage 0: PocketXMol de novo → Stage 1: ADMET-AI GNN → Stage 1.5: AiZynthFinder → Stage 2: Safety gates → Stage 3a: PocketXMol docking → Stage 3b: NeuralPLexer3 → Stage 4: Selectivity filter → Stage 4.5: DiffDock consensus → Stage 5: OpenMM 100ns + metadynamics → Stage 6: MM-PBSA + FEP + QM/MM → Stage 7: Boltz-2 + Chai-2 + scRNA-seq → Stage 8: Lab-in-the-Loop

Login → Command Center