Crane Bioworks AI Monthly / 2026.05

Deep learning-guided dual-fitness T7 RNAP engineering: Nucleic Acids Research, April 2026

T7 RNA polymerase is the workhorse of mRNA manufacturing. It has two practical problems: it denatures above ~40 °C, and it generates immunogenic double-stranded RNA (dsRNA) by-products that contaminate therapeutic mRNA preparations. These problems are mechanistically linked: dsRNA by-products form preferentially at elevated temperatures, exactly where high-yield transcription runs. Fixing one without the other is not useful.

Jiang et al. applied independent deep learning models to each fitness axis, then co-selected variants that improved all three simultaneously. The validated outcome: final engineered variants showed a >10 °C increase in T_m over wild type, a 60-fold increase in transcriptional activity above 48 °C, and about 70% reduction in immunogenic dsRNA by-product content, all three axes improved in the same variants, confirmed in cell transfection assays.

This is a near-perfect illustration of why multi-property engineering matters: the stability gain is only commercially useful if activity and purity move with it. The 70% dsRNA reduction translates directly to downstream purification cost savings in mRNA vaccine and therapeutic manufacturing.

Jiang, Kang et al., Nucleic Acids Res. 54, gkag259 (2026). doi:10.1093/nar/gkag259

iCASE: Nature Communications, January 2025

Most stability campaigns assume you’ll sacrifice activity. The iCASE strategy challenges that assumption at the physics level. The method identifies two measurable molecular quantities: isothermal compressibility (β_T), which negatively correlates with melting temperature, and a dynamic squeezing index (DSI), which positively correlates with catalytic activity. A structure-based supervised ML model trained on these features predicts residue positions where mutations can push both quantities in the desired direction at once.

Validated across six structurally distinct enzyme classes: a monomeric glutaminase (PG), a TIM-barrel xylanase (XY), a hexameric decarboxylase (GADA), a transglutaminase (MTGase), a laccase, and the PET hydrolase PES-H1. Thermostability and activity were synergistically improved in each case. Because the method targets physical properties of the chain rather than fold-specific residues, it transferred across all six fold types without retraining.

Ke et al., Nat. Commun. 16, 600 (2025). doi:10.1038/s41467-025-55944-5

PRIME: Science Advances, November 2024

PRIME is a temperature-guided protein language model trained on 96 million bacterial sequences annotated with host-organism optimal growth temperatures (OGTs). The premise: amino acid distributions that survive high temperatures in nature encode stability patterns that generalise to engineered proteins. Applied to five structurally unrelated targets without any prior mutagenesis data for each target, the model proposed 30-45 single-site mutations per protein.

Validated outcomes, protein by protein: LbCas12a reached +6.5 °C T_m from an 8-site combination variant identified in fewer than 100 total mutants tested across three iterative AI-experiment cycles. T7 RNAP reached +12.8 °C T_m in a 12-site combination variant screened from 95 single-site candidates. A VHH nanobody showed improved antigen-binding affinity while maintaining thermostability. Across all five proteins, >30% of recommended single-site mutants showed measurable improvement in the target property.

The multi-property framing is important: PRIME optimised for thermostability, binding affinity, polymerase activity, and alkaline resilience in different targets using the same model. Temperature is treated as a proxy for general fitness, not a single-axis endpoint.

Tang et al., Sci. Adv. 10, eadr2641 (2024). PMC11601203

SPURS: Nature Communications, 2025

ThermoMPNN has been the standard ΔΔG predictor since 2023. SPURS improves on it by wiring ProteinMPNN’s per-residue structural representations directly into ESM2’s attention layers, giving the language model access to both evolutionary signal and local geometry before scoring each mutation. On 12 benchmark datasets, SPURS achieves a median Spearman correlation of 0.83 versus 0.77 for ThermoMPNN, with statistically significant improvement on 24 of 28 proteins (Mann-Whitney U, P < 0.05). On 141 deep mutational scanning datasets, SPURS-augmented predictions outperformed the baseline in 82% of cases, with a +15% median improvement in Spearman correlation.

The most practically useful gap is directional: SPURS is specifically better at identifying stabilising mutations. Most published datasets are enriched for destabilising variants (they’re easier to produce and measure) so most models encode that imbalance. SPURS partially corrects for it. For a stability engineering campaign, the stabilising direction is the only direction that matters.

Li & Luo, Nat. Commun. (2025). doi:10.1038/s41467-025-67609-4 Code: github.com/luo-group/SPURS

Baker lab: tP4H case study: Angew. Chem. Int. Ed., 2025

King et al. used ProteinMPNN to redesign a non-heme iron enzyme (tP4H, from the Fe(II)/αKG superfamily) for greater thermostability, solubility, and expression without sacrificing native catalytic activity. The stabilised scaffold was then subjected to site-saturation mutagenesis targeting a non-native remote C(sp³)-H hydroxylation reaction.

Variants from the stabilised scaffold showed much larger gains in non-native activity than equivalent variants from the wild-type background under identical screening. The mechanism is indirect: a higher-stability scaffold tolerates more structurally disruptive active-site mutations without unfolding, expanding the fitness landscape that directed evolution can access. Stability isn’t the end goal here; it’s the margin that makes harder functional evolution tractable. Computational stabilisation before directed evolution is an investment in evolvability, not conservatism.

King, Sumida et al., Angew. Chem. Int. Ed. 64, e202414705 (2025). PMC11290999

ORI: Nature Communications, March 2026

Ontology Reinforcement Iteration (ORI) encodes structured biological knowledge, what a “thermostable chitinase” or “high-activity lysozyme” means, as semantic constraints that guide a generative model, then iteratively updates the model from real experimental feedback. The closed loop is: generate variants → measure in the lab → update the model → generate again.

Validated results: a lysozyme variant with 100-fold higher activity, a chitinase stable at 85 °C where natural comparator enzymes showed little or no activity, and dual-function enzyme variants, all produced in real experimental settings. The 85 °C chitinase is directly relevant for industrial processes that run at 60-80 °C. The framework handles stability and function as independent or joint objectives in the same system.

He et al., Nat. Commun. 17 (2026). doi:10.1038/s41467-026-69855-6

AIEEP platform, Zhao lab (UIUC): Nature Communications, July 2025

The Zhao lab at UIUC published the most complete AI-robot closed loop to date. Their platform integrates ESM-2 and epistasis models with an LLM orchestration layer and robotic lab execution. Input: a protein sequence and a quantifiable fitness measurement. No human judgment is required between design cycles.

Benchmarked on two enzymes (AtHMT methyltransferase and YmPhytase), four autonomous engineering cycles produced variants with up to 26-fold increased activity and 90-fold greater substrate specificity versus the starting sequence, characterising fewer than 500 total variants in four weeks. Conventional directed evolution for comparable gains typically requires more than 2,000 variants over six months. What the platform does not yet handle: stability or solubility as co-objectives. The fitness loop closes on one property at a time.

Zhao et al., Nat. Commun. 16, 5648 (2025). doi:10.1038/s41467-025-61209-y

FLAb2: bioRxiv (preprint), December 2025

Chungyoun & Gray (Johns Hopkins) benchmarked 30 protein AI and biophysical models against the largest public therapeutic antibody developability dataset assembled to date: 4 million antibodies, 32 studies, seven properties (thermostability, expression, aggregation, binding, pharmacokinetics, polyreactivity, immunogenicity). The result: current AI models produce statistically significant correlations on only about 20% of developability datasets. For 80%, no model achieves significance.

The diagnosis is specific: germline edit distance accounts for about 40% of the apparent predictive power of protein language models: models that look good on leaderboards are partly measuring how different an antibody is from its germline template, not predicting developability. Training data composition matters more than architecture. Thermostability and expression are learnable from sequence. Aggregation, pharmacokinetics, and immunogenicity are not yet reliably predictable. Several models (IgLM, ProGen2, Chai-1, ESM2) show statistically significant correlation specifically for thermostability and expression.

The practical implication: computational developability flags are useful for prioritisation, not shortlisting. Use them to reduce the candidate pool, then confirm experimentally.

Chungyoun & Gray, bioRxiv, December 2025 (preprint). bioRxiv 2025.12.27

EnzyArena benchmark + wet-lab validation: bioRxiv, March 2026

The question this paper actually answers: does an ML-designed mutation library outperform random mutagenesis in practice? Ten representative models (zero-shot fitness predictors and supervised kinetic-parameter predictors) were benchmarked on the EnzyArena dataset, then a 150-mutant wet-lab campaign compared ML-prioritised versus randomly selected variants head-to-head.

Result: an ML-guided library achieved a 39.32% hit rate versus 6.06% for error-prone PCR random baseline: a 6.5× enrichment. ESM-1v (zero-shot PLM) achieved the highest hit rate among individual models. Supervised kinetic predictors underperformed the random baseline on independent datasets despite strong benchmark scores. The diagnosis: supervised predictors overfit to training data and lose their advantage on novel proteins. Consensus of multiple zero-shot models consistently outperformed any single model and is the most robust strategy for mutation library design on a new target.

Multiple institutions, bioRxiv, March 2026 (preprint). bioRxiv 2026.03.29

Quara Supreme, Novonesis, November 2025

Novonesis launched Quara Supreme, a thermostable phospholipase C engineered for vegetable oil water degumming (soybean, rapeseed, canola). The enzyme operates at elevated process temperatures without the free-fatty-acid side reactions that complicate earlier-generation phospholipases, eliminating caustic dosing steps and simplifying heater configuration.

The commercial claim: up to 2% improvement in oil yield via enzymatic splitting of phospholipids to release diglycerides into the oil phase. At commodity processing volumes, 2% yield recovery is meaningful revenue. Novonesis does not publish the molecular engineering details behind commercial products, but this launch is a clean illustration of the value architecture: thermostability improvement → wider process temperature window → quantifiable yield gain. Stability, in this context, is not a specification; it is the product.

Novonesis press release, November 2025. novonesis.com → Quara Supreme

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

• Stabilise first, evolve harder. The Baker lab tP4H result establishes the mechanism: stability headroom expands the fitness landscape available to directed evolution. This is now a validated strategy, not a hypothesis. Computational stabilisation before directed evolution is leverage, not conservatism.
• The stability/activity tradeoff is not a law of nature. iCASE, ORI, and the T7 RNAP dual-fitness paper all show simultaneous improvement in both dimensions across structurally diverse proteins. The tradeoff is real in parts of sequence space. Physical-property-guided search finds the regions where it breaks down.
• Zero-shot consensus beats single supervised models on new proteins. EnzyArena makes this case with wet-lab data: a 39% hit rate versus 6% for random mutagenesis, driven by zero-shot PLM consensus. Supervised kinetic predictors overfit to training targets and underperform random on novel sequences. The implication for library design is concrete: run multiple zero-shot scorers, combine their rankings, do not rely on a single model.
• For antibody developability, computational flags are for prioritisation, not decisions. FLAb2 shows that 80% of developability datasets yield no statistically significant correlation from current models. Thermostability and expression are learnable; aggregation and PK are not yet reliably predictable. Use computational screens to reduce the candidate pool, then confirm experimentally before committing to a development track.