Crane Bioworks AI Monthly / 2026.06

Riff-Diff: computational enzyme design by catalytic motif scaffolding: Nature, 2026

Riff-Diff (Rotamer Inverted Fragment Finder-Diffusion) places catalytic residue arrays into de novo protein backbones using fragment-based rotamer matching combined with diffusion-based backbone generation. The problem it addresses is not just stability or just function in isolation: it is generating proteins that have both simultaneously from scratch, without a natural enzyme as a starting point.

Braun et al. tested 35 sequences across two mechanistically distinct reactions: a retro-aldol condensation and a Morita-Baylis-Hillman reaction. Active enzymes were confirmed for both. Nearly all fold correctly and remain active at 90 °C or above, thermal tolerance that most natural enzymes require directed evolution to approach. Catalytic rates in the validated variants are in the range of laboratory-evolved enzymes.

The implication: de novo enzyme design historically traded activity for novelty. Riff-Diff shows that industrial-grade thermostability can be an output of the design step itself, not a follow-up stabilisation campaign.

Braun, Tripp, Chakatok et al., Nature 649, 237-245 (2026). doi:10.1038/s41586-025-09747-9

MGnify Stability Dataset + SaProtΔG / ESM3ΔG, Rocklin and Ovchinnikov labs: bioRxiv, May 2026

The Rocklin lab at Northwestern and the Ovchinnikov lab at MIT released the largest open protein stability dataset to date: experimental folding stability measurements for 1.8 million diverse protein domains, from over 200,000 sequence families in the MGnify metagenomic database, all 60-80 amino acids in length. Two models fine-tuned on this resource, SaProtΔG and ESM3ΔG, achieve Spearman rank correlations of 0.88 and 0.87 respectively against held-out measurements, with RMSE of 0.8 kcal/mol across a 6 kcal/mol dynamic range.

The generalization envelope is what distinguishes these models from previous ΔΔG tools. They predict effects of insertions and deletions, not only point substitutions. They recover stability trends in thermophilic organisms without explicit temperature supervision. SaProtΔG generalises to nanobody aggregation temperature (T_agg) without any nanobody training data, a zero-shot cross-domain transfer that smaller curated training sets have not enabled.

Current limitation: the 60-80 AA domain scope means full-length multi-domain proteins require transfer learning or supplemental data. Dataset and model weights are openly released.

Tsuboyama, Rocklin, Ovchinnikov et al., bioRxiv, May 2026. bioRxiv 2026.05.19

BYU team: expert-guided deep learning for luciferase thermostability: ACS Catalysis, 2026

NanoLuc luciferase is the dominant bioluminescent reporter in high-throughput biology. Its practical ceiling: wild type denatures above roughly 40 °C, limiting use in elevated-temperature assays and expression systems that run warm.

Researchers at Brigham Young University applied an expert-guided deep learning workflow: deep learning proposes candidate mutations, structural domain knowledge filters for modifications to termini and loops distal from the catalytic centre that preserve allosterically coupled networks, and wet-lab assay confirms outcomes. Validated results: variant B.07 gained +7.2 °C T_m and variant B.09 gained +5.1 °C T_m relative to wild type, with sustained luminescence at elevated temperatures.

The design logic (stabilise the periphery, preserve the coupled core) is a generalizable principle for any enzyme engineering campaign where active-site disruption is the primary failure mode.

BYU team, ACS Catal. (2026). doi:10.1021/acscatal.5c08789

JanusDDG: Communications Biology, 2026

Predicting ΔΔG from sequence alone, without structural input, is useful for large-scale scanning where running structure prediction for each candidate adds prohibitive latency. JanusDDG achieves this with explicit physical constraints: a bidirectional cross-attention mechanism compares wild-type and mutant protein language model embeddings, and the architecture enforces thermodynamic antisymmetry (ΔΔG of A→B equals −ΔΔG of B→A) and transitivity (ΔΔG values are internally consistent across intermediate states).

These constraints are known failure modes of most supervised ΔΔG predictors, including ThermoMPNN and StaB-ddG. They learn approximate antisymmetry from data but violate it on out-of-distribution inputs. JanusDDG enforces them structurally. Result: sequence-only accuracy matching or exceeding structure-based methods on single and multi-point mutation benchmarks. For stability scanning at scale, this materially changes the precision-latency tradeoff.

JanusDDG team, Commun. Biol. (2026). doi:10.1038/s42003-026-09632-9

DETANGO: disentangled protein language models: bioRxiv, February 2026

Protein language models learn two things simultaneously: which sequences fold stably, and which positions are functionally conserved. These signals overlap. A mutation that lowers PLM log-likelihood may be destabilising, or may be hitting a critical catalytic or binding residue, so the model output is ambiguous. This is a documented failure mode of zero-shot PLM fitness predictors, and it limits their hit rates for activity-first engineering campaigns.

DETANGO explicitly removes the stability-correlated component from PLM mutation-effect predictions to isolate function-specific signal. In benchmark tests, methods using only stability or only conservation both underperformed DETANGO at identifying ligand-binding and cryptic functional residues.

The practical framing: once a scaffold is computationally stabilised (via ProteinMPNN redesign, PROSS, or a Riff-Diff de novo structure), applying DETANGO upstream of the functional screening step should improve hit rates by preventing stability noise from masking the functional signal.

DETANGO team, bioRxiv, February 2026. bioRxiv 2026.02.03

Therapeutic Nanobody Profiler (TNP), Oxford Protein Informatics Group: Communications Biology, 2026

The Therapeutic Antibody Profiler (TAP) has been the standard for early-stage mAb developability screening since 2019. Nanobodies need a different tool. Without a light chain, CDR3 loops fold back over the FR2 region, and exposed FR2 tetrad residues confer solubility in the absence of a VL partner. Antibody metrics calibrated on IgG architecture don’t transfer cleanly to VHH. That point was made concrete by last month’s FLAb2 benchmark, which showed that current AI developability tools produce no statistically significant correlation on 80% of antibody datasets and are worse on nanobody-specific ones.

TNP, from the Oxford Protein Informatics Group, defines six computationally derived metrics covering non-specific binding, polyspecificity, hydrophobicity, self-association, and aggregation, all calibrated on 36 clinical-stage nanobody sequences and 72 more nanobodies with in vitro developability data. CDR3 and the FR2 tetrad are treated explicitly as VHH-specific risk hotspots.

Practical use: before committing to SEC, DLS, or HIC characterisation of a candidate panel, run TNP to flag high-risk sequences. The calibration set is small, but it is the right molecule class and it is the largest public clinical-nanobody benchmark currently available.

Gordon, Gervasio, Souders et al., Commun. Biol. 9, 344 (2026). doi:10.1038/s42003-026-09594-y Web tool: opig.stats.ox.ac.uk/webapps/tnp

CRADLE-1, Cradle Bio, deployed at Novonesis and 6 of top-25 pharma: bioRxiv, March 2026

Cradle Bio published the engineering workflow and benchmark results behind CRADLE-1, their automated multi-property protein lead optimisation platform. Core efficiency claim: 4-7× fewer wet-lab rounds than rational design to reach the same optimisation endpoint. The system fine-tunes foundation protein language models on evolutionary neighbourhoods and lab-in-the-loop experimental data, then runs multi-model workflows that simultaneously optimise binding, activity, and thermostability across VHHs, scFvs, IgGs, peptides, and enzymes.

Benchmark context: across all four challenges in the Align to Innovate competition (30 teams from pharma and academia), CRADLE-1 won or tied first place with zero human intervention. The platform is deployed at Novonesis and six of the top-25 global pharmaceutical companies. Novonesis’s adoption specifically covers stability engineering campaigns, providing external validation of the round-reduction claim in a production context where iteration time maps directly to manufacturing cost.

Cradle Bio team, bioRxiv, March 8, 2026. bioRxiv 2026.03.06

Reading across these seven papers and one industry launch, four conclusions hold:

• De novo design now delivers thermal stability by default. Riff-Diff shows that the generative scaffold can impose 90 °C tolerance. Stabilisation is no longer a separate campaign after design. Combining this with last month’s Baker lab tP4H finding (stability headroom expands functional evolvability), the case for stabilising early and stabilising structurally has moved from hypothesis to proven strategy.
• The open training data gap is closing fast. The 1.8M domain Rocklin/Ovchinnikov dataset is the largest public experimental resource for folding stability to date. Expect fine-tuned models for aggregation, pH tolerance, and expression yield to emerge over the next two quarters as groups build on this training commons.
• Stability and function can now be separated computationally. DETANGO addresses the core ambiguity that limits zero-shot PLM screening: whether a mutation is bad because it destabilises the fold or because it disrupts function. Disentangling these signals reshapes how mutation scanning results should be interpreted before wet-lab commitment.
• Nanobody developability now has dedicated infrastructure. TNP closes the gap that FLAb2 identified: antibody tools were not designed for VHH and don’t perform on VHH. A clinical-nanobody-calibrated profiler now exists and is openly available. For any programme generating therapeutic nanobody candidates, TNP belongs at the front of the characterisation funnel.