Preprints, patents, and platforms shipped during a decade and a half of protein engineering and applied AI.
Maximising the yield of recombinantly expressed proteins is a critical part of any protein engineering pipeline. ExpressUrself captures spatial characteristics of the sequence around the start codon to predict expression to a high degree of accuracy on previously unseen transcripts.
Improvements to Geometric Vector Perceptrons for sampling sequences from a known backbone. Treats the trained classifier as an Energy-Based Model and improves median identity from 40.2% to 44.7%; AlphaFold-predicted structures of sampled sequences resemble originals (avg TM-score 0.84).
ConvTOX is a CNN that classifies protein toxins across the domains of life: >80% on animal/plant toxins, >50% on bacterial. Generalizes to toxin types (neuro vs. myo) and identifies structural similarity between toxins.
A recurrent-neural-network design system produces de novo variants of an antifungal peptide at <50% identity to wild-type. Molecular dynamics and in vitro assays confirm chitin binding and antifungal activity equal to or exceeding the WT peptide.