Crane Bioworks AI Monthly / 2026.09

AI-redesigned starting points and outcomes enhance protein evolution, Nature, July 2026

David Liu’s lab at the Broad Institute tested a specific hypothesis: if ProteinMPNN redesigns a protein’s sequence while keeping its fold and function, the redesign should be more mutationally robust than the wild type, and a more robust starting point should evolve further under the same selection pressure. They redesigned three distinct botulinum neurotoxin (BoNT) proteases with ProteinMPNN, then ran phage-assisted evolution on the redesigns and on wild type in parallel, selecting for a new specificity: cleaving ataxin-2 instead of the natural substrate.

Evolving from the redesigned starting point won every time. The three selected designs showed improved expression and thermal stability going in, with catalytic efficiencies 1.7 to 2.8 times higher than wild-type BoNT/E even before evolution started. After evolution, the best redesign-derived protease, carrying 46 ProteinMPNN substitutions plus 18 evolution-derived mutations (16% sequence divergence from natural BoNT/E), reached more than 79-fold greater selected specificity for ataxin-2 than the best protease evolved from wild type in the same selection.

This is the clearest evidence yet for a thesis this roundup keeps returning to: stabilizing a protein before you optimize it doesn’t just protect what you have, it changes how far the next round of search can go.

Krasnow, Xu, Zhang, Mahadeshwar, Tao, McCreary, Hemez, Brown, Jiang, Liu, Nature 657, 525-538 (2026), published 2026-07-22. doi:10.1038/s41586-026-10820-0

A machine learning model predicts protein stability of annotated and alternate protein isoforms, bioRxiv, July 2026

A gene doesn’t always translate from a single start codon. Alternate translation initiation produces isoforms with different N-termini from the same mRNA, and the N-terminus is exactly where the cell’s degradation machinery reads its stability cues (the N-end rule and related N-degron pathways). This study trained a model to predict stability directly from N-terminal sequence, then checked its calls against measured stability using a global protein stability (GPS) reporter assay.

The validation confirmed known destabilizing motifs and surfaced novel ones the N-end rule alone doesn’t explain. Applied to annotated proteins against their alternate-start isoforms, the model found pairs differing by only a handful of N-terminal residues with drastically different predicted stability. For anyone chasing expression yield, that’s a lever most codon-optimization work never touches: the same transcript can be quietly producing two populations of a protein with different half-lives, before a single downstream mutation enters the picture.

bioRxiv, posted 2026-07-16 (preprint). bioRxiv 2026.07.16

Discriminator-Guided Inverse Folding for Multi-Property Protein Design, Advanced Science, June 2026

Multi-property training data is the standing obstacle in this category: almost nobody has a dataset labeled for thermostability and solubility on the same variants. Zheng and colleagues at Zhejiang University worked around it. DGIF adds a discriminator module that steers ESM-IF1’s internal decoding states, built from predictors trained independently on single-property datasets. The result is three variants: DG-Thermo, DG-Sol, and DG-Dual, the last optimizing both properties in the same generation pass.

DG-Dual outperformed ESM-IF1 in top-K recall on the Megascale test set and SoluProtMutDB. The reported Pareto-front shift uses predicted properties. Megascale also supplied training data, so its test split should not be described as an independent dataset. The paper also reports wet-lab validation on an alcohol dehydrogenase. That supports the approach on one target; it does not establish general performance across protein families.

Zheng, Liu, Liu et al., Zhejiang University, Adv. Sci. (2026), online 2026-06-09. doi:10.1002/advs.75988

EvoMax: the EVOLVEpro/ALDE lineage of active-learning directed evolution gets a new entrant. EvoMax combines Gaussian process regression, protein language models, and inverse folding to prioritize the next round under sparse experimental data. Applied to Fanzor2, a family of compact eukaryotic nucleases, the best variant (FanzMAX v3-hLa) reached up to 97% editing efficiency at its best-performing locus and a mean of about 33% across 19 endogenous loci, outperforming the established compact editors enNlovFz2 and enCnCas12f1 by more than 2.6-fold. Not a stability result, and the target here is genome editors rather than an industrial enzyme, but it’s the same sparse-data active-learning stack a stability campaign leans on, proven on a harder search problem. Nature Biotechnology, 2026. doi:10.1038/s41587-026-03272-4

Amylo-Pipe: aggregation-prone-region predictors tell you where a protein might aggregate. They rarely tell you how fast. Amylo-Pipe pairs both, adding a kinetics layer including a light-chain-specific submodule (VLAmY-Pred) for antibody developability work. In a worked example on two light chain variants with measured aggregation rates, AL12 (−1.91) and AL103 (−4.40), the tool correctly ranked AL12 as the faster aggregator. That’s a small case, not a validation cohort, but the kinetics layer is the part most aggregation predictors skip. bioRxiv, posted 2026-06-09. bioRxiv 2026.06.09

Two things worth carrying forward:

• Stability keeps showing up as the precondition, not the payoff. The Broad Institute result makes this explicit: a redesigned, more stable, better-expressed protein wasn’t just easier to work with, it evolved further under the exact same selection pressure than wild type did. The isoform paper makes the same point from a different angle: stability differences that already exist between two forms of the “same” protein, before anyone introduces a mutation, decide how much usable product a construct actually yields.
• Multi-property claims are getting more honest about their evidence. DGIF combines benchmark evaluation with wet-lab validation on one enzyme. Keep predicted properties, held-out test data, and experimental measurements distinct when assessing that evidence.