#2 · ML / AI methods

Machine-learning-guided de novo design of protein and peptide binders

The dominant thread is computational design of binders in which generative or learning-based proposal is coupled to structure-based or biophysical scoring and then experimental validation. Nanobody workflows against a novel DSRCT target screened 288,000 designs and reported KD values of 0.66-305 nM by YSD and SPR, while SEPIA pairs about 26M AlphaSeq binding measurements with designed VHH-SEP pseudo-structures to supply training data for de novo antibody design. Active learning with AlphaFold screening recovers binders from peptide libraries for BET ET-domain targets, and an AlphaFold-Multimer screen yielded cell-confirmed nanobodies against GPCR MRGPRX2. New elements include ReGlyco, a glycosylation-aware filter for Nipah virus glycoprotein binders, and an MD-guided free-energy window linking selective membrane permeabilization to antiviral activity of amphipathic peptides. Tasks and modalities differ across reports, so the shared feature is methodological rather than a common benchmark.