https://pages.htgaa.org/2026a/fiona-connolly/-homework/index.htmlHugoenhttps://pages.htgaa.org/2026a/fiona-connolly/-homework/week-05-hw-protein-design-part-ii/index.htmlMon, 01 Jan 0001 00:00:00 +0000https://pages.htgaa.org/2026a/fiona-connolly/-homework/week-05-hw-protein-design-part-ii/index.htmlWeek 5 — Protein Design II AI-driven peptide and protein engineering, worked end-to-end on two targets. TL;DR Tool stack for peptide design: PepMLM (generate) → AlphaFold3 (validate) → PeptiVerse (triage) → moPPIt (re-target). Each tool catches a failure the others miss. Target 1: SOD1-A4V (ALS). PepMLM alone produces mode-collapsed peptides that all dock at the wrong AF3 default surface. moPPIt with motif guidance produces target-aware chemistry. Advance: B3 PAEKWFVFWHPT (sub-µM predicted Kd, dimer-interface targeted). Target 2: MS2 L-protein. ESM-style saturation scan vs random vs experiment-led picks. Big finding: language-model preference and experimental lysis function have r = +0.007 correlation. The model’s top picks would have destroyed function. Meta-lesson: Unsupervised protein language models predict sequence plausibility, not function. On under-represented protein families they can be actively misleading. Course: HTGAA Spring 2026 · Lecture (Mar 3): Gabriele Corso, Pranam Chatterjee — Protein Design Part II · Author: Fiona C (Committed Listener BioPunk Node)