Crane Bioworks AI Monthly / 2026.07

WRAPs, Science, July 2026

Membrane proteins make up nearly a third of the proteome and are targeted by over half of clinically approved drugs, but their hydrophobic surfaces complicate production and structural analysis. The Baker lab’s answer is WRAPs: RFdiffusion-designed amphipathic proteins that wrap a target’s hydrophobic, lipid-facing surface and render it thermostable and water-soluble without any detergent.

The final paper reports WRAPs for monomeric and oligomeric beta-barrel proteins and for helical multipass membrane proteins, including targets without an experimental structure. The designs preserved native sequence, fold, active-site properties, and ligand binding. A 2.95 Å cryo-EM structure of a WRAPed mycobacterial porin demonstrated that the approach can support membrane-protein structure determination in solution.

Same problem shape Crane campaigns are built around: a protein that works but can’t survive the conditions you need to produce, purify, or formulate it in. WRAPs turns “can’t get it soluble” into a tractable design problem.

Baker lab et al., Science 393, eadr3817 (2026). doi:10.1126/science.adr3817

Decoding Bispecific Antibody Developability, bioRxiv, June 2026

Bispecific and multi-domain formats compound developability risk. Two binding arms mean two sets of liabilities, plus whatever the new interface introduces. This study built the first large, uniform dataset to actually quantify that: 160 bispecific antibodies and their 65 parental antibodies on a matched knobs-into-holes CrossMab IgG1 scaffold, profiled across 10 biophysical assays on the PROPHET-Ab high-throughput platform.

The result splits developability properties into three classes. Hydrophobicity and surface charge inherit cleanly from parent to bispecific (Spearman ρ ≈ 0.85 to 0.95), so screening the parental arms predicts the bispecific’s fate on those axes. Self-association and polyreactivity inherit only partially (ρ ≈ 0.60 to 0.88), with emergent outliers driven by Fv-Fv charge complementarity between the two arms. Thermostability is poorly predicted from the parents (ρ < 0.4), so it requires testing on the assembled bispecific.

The practical rule: some liabilities triage at the single-arm stage, others only show up once the format is assembled. Skipping format-specific testing on the assumption that “the parents were clean” is where bispecific programs get burned late.

bioRxiv, posted 2026-06-19 (preprint). bioRxiv 2026.06.15

Anti-VQIVYK nanobody design, bioRxiv, April 2026

Aggregation-targeted design has a validation problem: most designed binders are only ever tested against clean, synthetic protein. This study aimed higher. Starting from biophysical modelling of low-energy conformations of VQIVYK, the hexapeptide motif that forms the structural core of tau neurofibrillary tangles, the authors generated 145 candidate nanobodies through backbone sampling and neural-network-guided CDR3 design, then prioritized computationally before testing.

Two candidates showed robust ELISA binding to synthetic full-length tau (binding indices of 148.9% and 140% relative to reference controls). One of those two also showed strong reactivity against post-mortem Alzheimer’s disease human brain tissue, with a binding index of 236.1%, exceeding the positive control (222.9%). That’s a materially higher bar than a synthetic-protein ELISA because the aggregate architecture in real diseased tissue is not guaranteed to match the design target.

bioRxiv, posted 2026-04-05 (preprint). bioRxiv 2026.04.01

Four predictors have shipped since our last knowledgebase check. None of these carry prospective wet-lab validation yet, so treat them as benchmark-stage until proven on a real campaign.

• AggrescanAI: a sequence-only aggregation-prone-region predictor using ProtT5 embeddings instead of structure. On the external AmyPro benchmark it reported AUROC 0.69, ahead of Aggrescan, Aggrescan3D, ANuPP, and AggreProt. J. Mol. Biol., 2026. doi:10.1016/j.jmb.2026.169643
• ProMelt / TmProt: a 45,441-protein T_m training set plus a LoRA-fine-tuned ESM-2 predictor, built to expose a real problem: predictors trained on proteomics-scale T_m data generalize poorly to classical biophysical T_m assays, the kind that actually matters for a stability campaign. TmProt reaches AUC 0.75 to 0.77 on biophysics-based validation sets where proteomics-trained baselines degrade. bioRxiv, posted 2026-05-11. bioRxiv 2026.05.07
• GraphESMStable: a ΔΔG predictor fusing a frozen sequence PLM with a trainable GNN via residue-level cross-attention, plus a dedicated “epistasis decoder” for multi-point mutations, a known weak spot for single-mutant tools like ThermoMPNN. On the Megascale benchmark (28 proteins, 28,312 mutations): median Spearman 0.84, versus 0.83 for SPURS and 0.77 for ThermoMPNN. bioRxiv, posted 2026-01-09. bioRxiv 2026.01.08
• ThermoFusion: the newest entrant, fusing ThermoMPNN structural embeddings with ESM2 sequence embeddings for single-point enzyme stability prediction, with an explicit focus on out-of-distribution generalization. Benchmarked on 105 enzymes and 3,144 mutation sites. bioRxiv, posted 2026-07-07. bioRxiv 2026.07.04

Three practical takeaways from July’s roundup:

• Membrane proteins have a promising new design route. Membrane proteins have been the “too hard to touch” category for developability work. WRAPs treats insolubility as a design problem without requiring an experimental structure, the same philosophy behind every Crane stability campaign, applied to the worst case.
• Format complexity needs format-specific testing. The bispecific developability map shows some liabilities transfer cleanly from single-domain screening and some don’t. Assume nothing about a multi-domain construct until it’s tested as assembled.
• The T_m/ΔΔG predictor field is crowded, and the honest gap is generalization. Four new tools shipped this cycle alone, and ProMelt/TmProt is explicit about why leaderboard wins don’t always survive contact with a real biophysical assay. That’s the case for running multiple models and checking them against wet-lab data, not trusting a single benchmark score.