Every result below is a computational hypothesis — a prediction from simulation and structure modeling, not a wet-lab-validated finding, and not a statement of clinical efficacy. We publish negative and contradicted results alongside promising ones, because rigorous self-correction is the point.

Between June 18–20 the platform's autonomous compute fleet reached several milestones on canonical SMA priority targets. This post documents what ran, what it means, and — just as importantly — what it does not yet mean.

8×H100 molecular dynamics went live

An NVIDIA Innovation Lab 8×H100 node is now saturated running 100 ns OpenMM molecular-dynamics simulations on priority targets (the p38 family MAPK11/MAPK12/MAPK14, KEAP1, HDAC2, and others). Each run produces a stable trajectory landed directly to canonical storage. These simulations test whether a docked pose is dynamically stable — a much stronger signal than a static docking score alone.

MM-GBSA binding-energy triage — explicitly preliminary

From those trajectories we computed end-state MM-GBSA binding-energy estimates for 34 compound/target pairs, now browsable on the new MD Leads page. These values rank-order candidates within a single target — they are not absolute binding affinities. They use a single trajectory, 100 frames, and no entropy term.

We are deliberately conservative here. A multi-model review consortium (three independent large language models cross-examining the candidate set) flagged the whole MM-GBSA list PREMATURE: the decoy and positive-control benchmark that would demonstrate the method can actually separate real binders from look-alikes has not yet completed. Two candidates are flagged and de-ranked outright — one MAPK12 value that is an artifact of molecular size, and one MICU1 candidate carrying an aromatic-nitro toxicophore. No row on the Leads page is a wet-lab-actionable lead. It is a ranked hypothesis set awaiting its own validation gate.

Multi-method PPI consensus + NMJ binder design

The new PPI Consensus page surfaces protein–protein interaction pairs predicted strong (interface ipTM ≥ 0.70) by two or more independent co-folding methods (AlphaFold3, Boltz-2, OpenFold3, Chai-1), with on-axis SMA-mechanism pairs highlighted (e.g. KEAP1–CUL3, MAPK11–MAPKAPK2, STMN2–tubulin). Alongside, a first set of de-novo single-domain (VHH) binders was designed against the neuromuscular-junction target MuSK Ig1; these are single-model Chai-1 predictions and are marked as awaiting orthogonal AlphaFold3 confirmation before any consensus claim.

The fleet correcting itself

A compute-waste sentinel autonomously detected and de-rostered off-axis bulk docking work (the ERK kinases MAPK1/MAPK3) that no longer reflects platform priorities, reallocating that capacity toward canonical targets — a reversible, auditable action. We treat the ability to stop doing low-value computation as a first-class feature.

What this is, and is not

This is preclinical, in-silico hypothesis generation. None of it establishes that any compound binds its target in a wet lab, modifies disease, or is safe. The value is in surfacing the strongest computational hypotheses transparently — with their caveats attached — so they can be prioritized for the experiments that would actually test them.