Claude Rewrites the Most Expensive Code in Biomolecular Models: 36 Optimization Kits, ~4x Speedups

Forum topic · 小凯 · 2026-09-18

Summary

Anthropic reports that Claude, supervised by two Anthropic researchers with no prior GPU kernel experience, rewrote the most compute-intensive components of 30+ open-source biomolecular models. The project targets triangle attention and triangle multiplication, the cubic-cost operations underlying models like AlphaFold3, OpenFold3, and Boltz-2. The result is FlashPairformer, a set of custom GPU kernels, released as 36 drop-in optimization kits in the GitHub repository anthropics/uplifting-biomolecular-modeling under Apache 2.0. On a single NVIDIA H100, Exact mode (bit-identical outputs) averages 1.6x speedup, Fast mode 4.1x, and Big mode 3.4x with reduced memory use. Big mode enables accurate folding of assemblies exceeding 10,000 tokens (TM-scores 0.92-0.997 for mitochondrial complex I, TRiC, proteasome, and 70S ribosome) and inference on systems over 70,000 tokens, though 31,000-70,000 residue viral capsid predictions collapsed (TM-score 0.08-0.14), attributed to poor generalization beyond training sizes. Benchmarking on FoldBench-Lite (1,925 pairs, 13 configs) showed no distinguishable pooled accuracy change (54.2-55.0% acceptable interfaces). Binder design costs reportedly fell from ~$10,000 to ~$150 per target, though designs lack wet-lab validation.

Anthropic's research report *How Claude is uplifting biomolecular modeling* (September 17, 2026) describes how Claude wrote custom GPU kernels targeting the two most expensive operations in structure-prediction models: triangle attention and triangle multiplication. These operations act on token triples, so cost grows cubically with system size (8x at double size, 27x at triple). The resulting kernel suite, FlashPairformer, ships as 36 plug-and-play optimization kits covering 30+ open-source models in the GitHub repository anthropics/uplifting-biomolecular-modeling (Apache 2.0).

Key points

What was optimized: Triangle attention and triangle multiplication, the cubic-cost bottleneck in AlphaFold3, OpenFold3, Boltz-2, Protenix, Chai-1, ColabFold, and others. Kits span co-folding/structure prediction (14), genomics (7), structure generation (6), hallucination-based design (3), inverse folding (3), and protein language models (3).
Three run modes per kit (original weights preserved):
Exact: bit-identical to original outputs, 1.6x average speedup on H100 (14 models tested)
Fast: minimal numerical deviation allowed, 4.1x average (13 models); per-model range 2.8x (AtlasFold) to 6.4x (Chai-1)
Big: reduced memory footprint, 3.4x average (13 models)
FlashPairformer internals: vs. NVIDIA's domain-standard kernels (cuEquivariance, BioNeMo Inference Runtime), triangle attention is 2.7x faster at pair width 128 and 2.9x at 256; triangle multiplication 1.7x and 3.2x. Gains reach ~4x at 2,048 tokens / width 128.
Accuracy: On FoldBench-Lite (1,925 model-target pairs, 13 configs), acceptable-interface rates were 54.8% (default), 55.0% (Exact), 54.5% (Fast), 54.2% (Big) — "no distinguishable pooled change." Caveat: per-model degradation could be masked by pooling; no per-model accuracy distribution is given.
Big mode scale: Accurately modeled assemblies >10,000 tokens — human mitochondrial complex I, TRiC chaperonin, proteasome, bacterial 70S ribosome, with TM-scores 0.92–0.997 ("to our knowledge" among the largest accurately folded structures). On 8x B300 (one node), 31,000–70,000-residue capsids and compartments ran seven times, but all collapsed into compact globules (TM-score 0.08–0.14 for scoreable cases). Anthropic attributes this to generalization failure — these assemblies are 40–90x larger than AlphaFold3's largest training slice — and frames the work as "extending what is computable, not what the model learned."
Binder design economics: Compared with an April 2026 round (~16,000-word prompts, sub-agents, human intervention, up to $10,000 and ~2,500 H100-hours per target), the new round used ~1,100-word prompts, no sub-agents, one H200 for 24 hours, at roughly $150 total (GPU + tokens) — a ~100x smaller compute budget. Median ipSAE: Opus 5 at 0.785, Mythos 5.1 at 0.781, Mythos 5 at 0.739 (vs. 0.749 prior round) — comparable, not uniformly better. All designs are unvalidated computational predictions; a joint Anthropic/Adaptyv Bio protein-design competition (5 targets including GPCRs, 5,000+ designs, up to $1M in Claude credits, $250K compute from Modal, DNA synthesis from Twist Bioscience) will provide wet-lab validation.
Human role: Code optimization was primarily done by Claude in the Claude Science environment, supervised by two Anthropic researchers familiar with biomolecular modeling but with no prior inference-optimization or GPU-kernel experience.

Reported limitations

Measurements only on H100 hardware
ColabFold baseline was v1.6.1, which has since shipped its own fused kernels
Exact and Fast modes use up to 3.2x more memory
No independent third-party benchmarking of FlashPairformer
The repository is a reference release: not maintained, contributions not accepted
Binder designs lack experimental validation

Why it matters

The broader significance may lie in AI coding rather than biology. Typical coding-agent benchmarks involve cheap verification (tests pass or fail). Optimizing scientific kernels requires expensive verification: knowing expected memory-access patterns, interpreting DockQ thresholds, reading pooled benchmarks. The report does not answer how close AI-written scientific code is to hardware limits, and closes with the caveat about computability versus learned capability. In the same week, AItonomy Foundation released ScienceIDE (sponsored by PhAI Labs and Alibaba Qwen), converting scientific codebases into verifiable agent learning environments and training PhAI-IDE-72B/9B/4B models (arXiv:2609.19134) — addressing whether such capability can be distilled and transferred. The gap between AlphaFold3's 768-token training slices and Big mode's 70,000-token (but collapsing) inference is precisely the distance between "can compute" and "computes correctly."

References

1. Anthropic, *How Claude is uplifting biomolecular modeling*, 2026-09-17. https://www.anthropic.com/research/claude-uplifts-biomolecular-modeling 2. Anthropic, *anthropics/uplifting-biomolecular-modeling* (GitHub, Apache 2.0), 2026-09-17. https://github.com/anthropics/uplifting-biomolecular-modeling 3. Unite.AI, *Anthropic Reports Claude Optimized 30+ Open-Source Biomolecular Models*, 2026-09-18. https://www.unite.ai/anthropic-reports-claude-optimized-30-plus-open-source-biomolecular-models 4. AI Insiders, *Anthropic says Claude rewrote 30 biology models to run 4x faster*, 2026-09-18. https://aiinsiders.net/article/anthropic-says-claude-rewrote-30-biology-models-to-run-4x 5. AItonomy Foundation, *ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments*, arXiv:2609.19134, 2026-09-16.