Scaling ESM-C 600M to 6B gave no zero-shot gain over ESM-2 650M on 201 ProteinGym assays: a non-significant regression (grouped Spearman rho 0.419 vs 0.431).
Add NeuroAutomata's protein tools to Claude or ChatGPT with one MCP connector URL. Which tools are free, which need Pro, and how the connection is secured.
An independent check corrected one of our own live ESM-2 claims about disordered regions — and overturned an internal assessment too. Why we don't self-certify.
Before building a per-position confidence view for our variant scorer, we pre-registered a pass/fail bar — witnessed, no escape clause. It failed. We didn't ship it.
ESM-2 ranks germline TP53 DNA-binding-domain variants reliably (Spearman rho 0.46–0.68) but carries no usable signal in the disordered regions (0 to −0.26).
Across 13,927 INSR ectodomain variants, ESM-2 ranks at ρ=0.594 on the L1 leucine-rich repeat and ρ=−0.088 at the αCT helix — same protein, same model, same assay. The contrast is structural.
Across the 5 canonical TPMT clinical alleles, ESM-2 ranks every one deleterious from sequence alone. VAMP-seq catches *2, *3B, *3C. ESM-2 covers *5 and *7.
ESM-2 vs 5,949 BRCA1 and PTEN variants — validated against SGE and VAMP-seq. Where it works, where it's blind, and why the combination matters.
ESM-2 predictions vs 6,142 CYP2C9 variants from the largest pharmacogenomic DMS dataset. What we found — including where it fails.
We built NeuroAutomata to make ESM-2 protein variant scoring accessible without setup. Validation results, including the one protein where it failed.
How I built a multi-agent system for natural language queries for Human Protein Atlas data, from naive RAG to AI verification architecture
Detailed validation methodology, reproducibility protocols, and AI agent architecture for the HPA natural language query system