raftbioworks/weaveflow_pdb_20260924

WeaveFlow: geometric moment memory

Completed private research artifacts for the September 24, 2026 campaign. WeaveFlow keeps a rigid frame for every residue and combines local invariant point attention with 32 global geometric moment summaries. It has no persistent residue-pair feature tensor. Exact fused dense attention is included as a control.

The systems result is strong; generation quality remains limited. All nine PDB-only variants scored 0/100 on their fixed designability panels. The online teacher-assisted moment student scored 14/100 joint refolding-and-geometry passes, the exact-attention student 12/100, and the pretrained FrameFlow teacher 32/100. Published FrameDiff scored 20/100 by refolding and 14/100 with geometry. These external baselines have different training histories. The cached-trajectory student scored 4/100 jointly at 80k and 6/100 at the separately declared 120k checkpoint. Every planned final sample is retained.

The campaign completed 830,000 new production updates across shared-parent lineages, retained 166 scheduled production checkpoints plus six diagnostic or deadline checkpoints, and evaluated 1,712 final backbones with two redesigns each. The elapsed Modal resource estimate is $165.34, or $174.72 including startup allowances, within the authorized $200 budget. This is resource accounting, not a provider invoice.

The 8.07M model completes a synthetic checkpointed AdamW update at 262,144 positions in 13.59 GiB of allocated memory. Exact fused attention also has linear activation memory; moment compression principally removes quadratic arithmetic. Original 27.05M online training used 7.05 GiB. A separate 100-update teacher microbatch-one profile used 3.35 GiB; no independent 80k quality result is assigned to that profile. Cached-label training used 3.34 GiB and had lower measured quality. These allocations exclude the CUDA context and allocator reserve.

• Research report, equations, results and reproduction commands
• W&B report run
• data/: source PDB-derived pack, prepared datasets, inherited splits and teacher cache.
• runs/: exact per-run source snapshots, configs, metrics, checkpoints and SHA256 sidecars; checkpoints include model, EMA, optimizer and RNG states.
• evaluations/: all raw backbones, ProteinMPNN designs, ESMFold predictions and results, including failures.
• benchmarks/: systems measurements and device metadata; final unified/capacity benchmarks pin H100.
• plans_and_tables/: timestamped protocols, audited scores, paired uncertainty, cost and lineage accounting.
• source/: implementation overlay, plots and dependency lockfile; snapshot.json identifies the base repository commit and source hashes.
• audit/: resource ledger, environment, tests, checkpoint index and immutable-revision archive verification.

The data is a selected PDB-derived pack, not the entire PDB. Computational designability is a ProteinMPNN/ESMFold self-consistency assay, not experimental validation. Synthetic long-sequence benchmarks do not demonstrate long-protein generation quality. A CA-based acceptance gate is not an all-atom clashscore; separate heavy-backbone clash counts are provided. Shared parents and one training seed per arm limit causal and statistical conclusions.

The archive retains per-run training source. Some early evaluator revisions lack complete source snapshots and four early generation panels lack contemporaneous checkpoint fingerprints; these gaps are listed in the protocol audit. Saved coordinates, sequences and predictions support independent score recomputation. The source overlay requires the base repository commit recorded in snapshot.json.

Code retains the repository MIT license in source/LICENSE, the vendored FrameFlow MIT notice and existing OpenFold notices. The frozen FrameFlow reference weights retain CC BY 4.0 attribution in source/model/frameflow/ORIGIN.md. Upstream datasets and other pretrained references retain their own licenses and provenance. ESMFold/ESM-2 weights are fingerprinted and referenced through their original distribution, not bundled in this archive.