#22 · Structural bio & protein design

Auditable predictions for protein yield, structure search, and proteome recoding

The dominant thread is that computational protein analyses are reported with checkable outputs. ProtoPure uses an LLM-based workflow to extract 12 standardized purification parameters from literature across UniProt proteomes, producing a database. Struct-Caliper provides query-specific p-values for structure search through representation-aware randomization, improving homolog recovery. A pooled E. coli assay measures monomeric yield for more than 39,000 proteins, benchmarks inverse folding, and trains the ESMC-MonoYield predictor. The glutamine-removal study ranks residues by replaceability using ProteinMPNN and MSA conservation, with proteome-wide recoding scope across 4,148 E. coli proteins. Not every item involves language models; the shared element is auditable output. What is new is the emphasis on uncertainty-aware, checkable results across LLM-mediated extraction and predictive modeling of experimental and published data.