Crane Bioworks AI Monthly / 2026.08

Combining Stability-Centered Atomistic Design with Machine Learning for Targeted Enzyme Optimization, bioRxiv, July 2026

FuncLib and its high-throughput successor htFuncLib, the Fleishman lab’s stability-centered design method, build enzyme libraries where every member is guaranteed to fold. What they don’t encode is which fold is good at your specific reaction. MLEE (machine learning-assisted enzyme engineering) closes that gap: it screens and sequences an initial round from the stability-guaranteed pool, trains a model on the result, then uses that model to pick a small, targeted follow-up set.

The follow-up set was just 12 variants. Against the htFuncLib baseline, it raised the hit rate for the global top 0.05% of variants by 5 to 12 times. The best variant increased turnover on 4-hydroxy-β-damascone by 11.8-fold, with >99% regioisomeric excess. The stability step isn’t a separate campaign here. It’s the search space the activity model gets to work with.

Wan, Bagherpoor Helabad, Fraedrich, Fleishman, Weissenborn, bioRxiv, posted 2026-07-20 (preprint). bioRxiv 2026.07.20

Experimentally validated deep learning control of protein aggregation, Communications Chemistry, April 2026

The Damborsky and Bednar labs at Masaryk University ran the kind of check aggregation predictors rarely get: they took AggreProt’s residue-level calls, ordered the mutations it recommended, and measured what happened at the bench. The test set was 34 hexapeptide fragments from the model enzyme LinB plus seven full-length proteins drawn from the AmyPro amyloid database. Predictions agreed with experiment in 79% of the 41 cases.

That’s not a benchmark score against a held-out dataset. It’s a prospective check on whether the tool’s top-ranked aggregation-prone regions are the ones that actually cause trouble, and whether suppressing them in silico holds up when the protein is expressed. For a solubility-limited candidate, that’s the difference between a predictor you screen with and one you design around.

Cima, Kunka, Planas-Iglesias et al., Commun. Chem. 9, 121 (2026). doi:10.1038/s42004-026-02007-5

Protein structure-informed deep learning enables species-specific codon optimization, bioRxiv, April 2026

Codon optimization tools usually work from sequence and host codon tables alone, which is why two genes with identical protein output can express at wildly different soluble yields. PISCO adds the missing input: a geometric graph model reads the folded structure alongside host codon usage, on the premise that translation speed across a folding domain, not just overall codon frequency, decides how much of the product ends up soluble.

Against structure-blind baselines, PISCO improved codon recovery by 6% and cut divergence from natural coding sequences by 42 to 50% across three similarity metrics. The authors then expressed genes optimized both ways and report higher soluble, functional yield from the structure-informed designs. No fold-change is published yet for that comparison, so treat the direction as validated and the magnitude as open.

bioRxiv, posted 2026-04-21 (preprint). bioRxiv 2026.04.21

Basecamp Research × Procter & Gamble

The partnership behind P&G’s cold-water laundry enzymes is a plain statement of the stability problem this whole field exists to solve: a protease that cleans well at 40°C usually loses activity at 15°C, and reformulating around that gap is exactly the kind of single-cycle stability project biotechs keep asking Crane about. Basecamp mines sequences from organisms adapted to cold environments and uses its models to propose variants tuned for low-temperature activity, aiming to cut the hot water a wash cycle needs.

Neither side has published quantitative wash-performance results since the partnership was announced, so this belongs in the roundup as a live industrial bet, not a validated win. It’s still the clearest large-scale case of the “better proteins, not capability” thesis: the goal was never a new detergent category, just enzymes that hold their activity where the old ones didn’t.

P&G Research & Development, ongoing partnership. pg.co.uk → Basecamp Research partnership

IFUM: most ΔΔG predictors learn a single number per mutation and never model what the unfolded state looks like. IFUM jointly estimates ΔG and the folded/unfolded equilibrium ensemble, represented as residue-pair distance distributions, and reports that joint training improves accuracy over predicting ΔG alone, including on insertions and deletions that break most point-mutation tools. Nature Communications, January 2026. doi:10.1038/s41467-026-68637-4

Two things worth carrying into next month:

• The stability step keeps paying for itself downstream. MLEE didn’t need a bigger model, it needed a library where every member already folds, so the follow-up round could spend its entire budget on activity. That’s the same logic behind stabilizing before directed evolution: fix the foundation once, then every subsequent search gets cheaper.
• Prospective checks are rarer than they should be, and worth more when they show up. AggreProt’s 79% agreement across 41 cases (34 hexapeptide fragments and seven full-length proteins) is a smaller number than most benchmark papers report, and more useful, because it comes from actually expressing the redesigned proteins rather than scoring a held-out test set. That’s the standard a stability claim should be held to before it reaches a wet-lab decision.