Abstract Carbonic anhydrases are among the fastest known biocatalysts, reversibly facilitating the hydration of CO 2 to HCO 3 - at rates up to 10 7 s - 1 , which warrants their investigation for industrial carbon capture technologies. However, engineering carbonic anhydrases to maintain stability under harsh industrial process conditions remains a key challenge, and sequence-to-function datasets compatible with machine learning to inform forward engineering are lacking. Here, we developed a high-throughput platform that couples cell-free gene expression with a gaseous CO 2 colorimetric assay to map the fitness landscapes of carbonic anhydrases. From 96 diverse natural homologs, we identified a robust variant from the Aquificota phylum and conducted an exhaustive mutational scan and functional assessment of this enzyme at 70°C and 90°C, covering >99% of all single-amino acid substitutions (totaling 4,365 mutations assayed in 39,285 reactions). This biochemical landscape was used to benchmark 22 zero-shot protein fitness models and identify critical mutations that improved enzyme stability at 90°C by more than three-fold. We then used both zero-shot protein language models and supervised learning to filter 419 model-generated variants from a ProteinMPNN library of 100,000 sequences, leading to a best-in-class enzyme that retained activity after incubation at 95°C. This work demonstrates that integrating cell-free enzyme engineering with machine learning enables opportunities for high-throughput experimental measurements to benchmark and improve protein language models, accelerate design loops, and expand functional exploration within protein families where experimental information is limited.

Abstract Generative models are increasingly used for protein design, but the lack of standardized evaluation frameworks limits comparison across model classes and hinders translation to experimental success. Here, we introduce a unified sampling and benchmarking framework that enables controlled sequence generation across alignment, protein language, and structure-based models, and apply it to Tobacco etch virus (TEV) protease. Across hundreds of thousands of designed sequences, different models explore distinct regions of sequence space with no clear computational selection metrics to assess enzymatic function. Experimental evaluation reveals large differences in functional outcomes, ranging from non-functional variants to sequences with 9-fold higher activity than wildtype. Machine learning-designed libraries achieve a 39.32% hit rate (percentage of variants matching or exceeding wildtype activity) compared to 6.06% for an error-prone PCR baseline. Structure-based models perform best overall, with hit rates of 74.4% and 66.8% for ESM-IF1 and ProteinMPNN, respectively. Commonly used selection metrics do not strongly correlate with experimental activity, highlighting a gap between in silico evaluation and enzyme function. Together, these results establish a generalizable framework for benchmarking generative protein models and demonstrate the necessity of experimental validation for guiding model development and sequence prioritization.

All folded proteins continuously fluctuate between their low-energy native structures and higher energy conformations that can be partially or fully unfolded. These rare states influence protein function, interactions, aggregation, and immunogenicity, yet they remain far less understood than protein native states. Although native protein structures are now often predictable with impressive accuracy, conformational fluctuations and their energies remain largely invisible and unpredictable, and experimental challenges have prevented large-scale measurements that could improve machine learning and physics-based modeling. Here, we introduce a multiplexed experimental approach to analyze the energies of conformational fluctuations for hundreds of protein domains in parallel using intact protein hydrogen-deuterium exchange mass spectrometry. We analyzed 5,778 domains 28-64 amino acids in length, revealing hidden variation in conformational fluctuations even between sequences sharing the same fold and global folding stability. Site-resolved hydrogen exchange NMR analysis of 13 domains showed that these fluctuations often involve entire secondary structural elements with lower stability than the overall fold. Computational modeling of our domains identified structural features that correlated with the experimentally observed fluctuations, enabling us to design mutations that stabilized low-stability structural segments. Our dataset enables new machine learning-based analysis of protein energy landscapes, and our experimental approach promises to reveal these landscapes at unprecedented scale.

Abstract Biological AI learns primarily from existing observations, but many questions cannot be answered from available data alone. Here we show that AI can instead acquire knowledge by acting directly on biological systems and learning from the consequences. We developed a closed-loop framework in which autonomous agents design protein variants, construct and characterize them in a robotic laboratory, learn from the resulting experimental feedback, and decide what experiments to perform next. We then allowed the system to operate continuously and without human intervention for approximately one month, during which multiple agents independently explored protein sequence space while learning from shared experimental experience. Applied to glycoside hydrolases, the agents discovered enzymes with substantially altered substrate specificity toward non-native sugars and progressively learned the structure of the underlying sequence–function landscape. The resulting experimental experience also revealed determinants of substrate specificity and protein expression that were not specified as learning objectives. These results demonstrate that AI can autonomously interact with biology over extended periods to acquire knowledge through experience, establishing a framework for biological discovery driven by continuous experimental interaction.

Disease-associated genetic variants occur extensively in noncoding regions like promoters, but current methods focus primarily on single nucleotide variants (SNVs) that typically have small regulatory effect sizes. Expanding beyond single nucleotide events is essential with insertions and deletions (indels) representing the logical next step as they are readily identifiable in population data and more likely to disrupt regulatory elements. However, existing methods struggle with indel prediction, and clinical interpretation often requires assessing complete promoter haplotypes rather than individual variants. We present LOL-EVE (Language Of Life for Evolutionary Variant Effects), a conditional autoregressive transformer trained on 13.6 million mammalian promoter sequences that enables both zero-shot indel prediction and complete promoter sequence scoring. We introduce three benchmarks for promoter indel prediction: ultra rare variant prioritization, causal eQTL identification, and transcription factor binding site disruption analysis. LOL-EVE's superior performance demonstrates that evolutionary patterns learned from indels enable accurate assessment of broader promoter function. Application to Genomics England clinical data shows that LOL-EVE can prioritize promoter haplotypes in known developmental disorder genes, suggesting potential utility for clinical variant assessment. LOL-EVE bridges individual variant prediction with haplotype-level analysis, demonstrating how evolution-based genomic language models may assist in evaluating regulatory variants in complex genetic cases.

Understanding RNA structure and predicting the functional consequences of mutations are fundamental challenges in computational biology with broad implications for therapeutic development and synthetic biology. Current evaluation of machine learning-based RNA models suffers from disparate experimental datasets and inconsistent performance assessments across different RNA families. To address these challenges, we introduce RNAGym, a large-scale benchmarking framework specifically designed for three core tasks–RNA fitness, secondary structure, and tertiary structure prediction. The framework integrates extensive datasets, including 70 standardized deep mutational scanning assays covering over a million mutations across diverse RNA types; 901k chemical-mapping reactivity profiles for secondary structure; and 215 diverse tertiary structures curated from the PDB. RNAGym is designed to facilitate a systematic comparison of RNA models, offering an essential resource to enhance the understanding and development of these models. ### Competing Interest Statement The authors have declared no competing interest.

In drug discovery, mapping interactions between genes within cellular systems is a crucial early step. This helps formulate hypotheses regarding molecular mechanisms that could potentially be targeted by future medicines. The CausalBench Challenge was an initiative to invite the machine learning community to advance the state of the art in constructing gene-gene interaction networks. These networks, derived from large-scale, real-world datasets of single cells under various perturbations, are crucial for understanding the causal mechanisms underlying disease biology. Using the framework provided by the CausalBench benchmark, participants were tasked with enhancing the capacity of the state of the art methods to leverage large-scale genetic perturbation data. This report provides an analysis and summary of the methods submitted during the challenge to give a partial image of the state of the art at the time of the challenge. The winning solutions significantly improved performance compared to previous baselines, establishing a new state of the art for this critical task in biology and medicine.

Retrieving homologous protein sequences is essential for a broad range of protein modeling tasks such as fitness prediction, protein design, structure modeling, and protein-protein interactions. Traditional workflows have relied on a two-step process: first retrieving homologs via Multiple Sequence Alignments (MSA), then training mod- els on one or more of these alignments. However, MSA-based retrieval is computationally expensive, struggles with highly divergent sequences or complex insertions & deletions patterns, and operates independently of the downstream modeling objective. We introduce Protriever, an end-to-end differentiable framework that learns to retrieve relevant homologs while simultaneously training for the target task. When applied to protein fitness prediction, Protriever achieves state-of-the-art performance compared to sequence-based models that rely on MSA-based homolog retrieval, while being two orders of magnitude faster through efficient vector search. Protriever is both architecture and task-agnostic, and can flexibly adapt to different retrieval strategies and protein databases at inference time – offering a scalable alternative to alignment-centric approaches.

Recurrent waves of viral infection necessitate vaccines and therapeutics that remain effective against emerging viruses. Our ability to evaluate interventions is currently limited to assessments against past or circulating variants, which likely differ in their immune escape potential compared with future variants. To address this, we developed EVE-Vax, a computational method for designing antigens that foreshadow immune escape observed in future viral variants. We designed 83 SARS-CoV-2 spike proteins that transduced ACE2-positive cells and displayed neutralization resistance comparable to variants that emerged up to 12 months later in the COVID-19 pandemic. Designed spikes foretold antibody escape from B.1-BA.4/5 bivalent booster sera seen in later variants. The designed constructs also highlighted the increased neutralization breadth elicited by nanoparticle-based, compared with mRNA-based, boosters in non-human primates. Our approach offers targeted panels of synthetic proteins that map the immune landscape for early vaccine and therapeutic evaluation against future viral strains.

Accurately predicting protein fitness with minimal experimental data is a persistent challenge in protein engineering. We introduce PRIMO (PRotein In-context Mutation Oracle), a transformer-based framework that leverages in-context learning and test-time training to adapt rapidly to new proteins and assays without large task-specific datasets. By encoding sequence information, auxiliary zero-shot predictions, and sparse experimental labels from many assays as a unified token set in a pre-training masked-language modeling paradigm, PRIMO learns to prioritize promising variants through a preference-based loss function. Across diverse protein families and properties-including both substitution and indel mutations-PRIMO outperforms zero-shot and fully supervised baselines. This work underscores the power of combining large-scale pre-training with efficient test-time adaptation to tackle challenging protein design tasks where data collection is expensive and label availability is limited.

A generative artificial-intelligence tool has designed a synthetic CRISPR system that successfully edits human DNA and sharply reduces off-target effects. A generative artificial-intelligence tool has designed a synthetic CRISPR system that successfully edits human DNA and sharply reduces off-target effects.

Quantitatively mapping enzyme sequence-catalysis landscapes remains a critical challenge in understanding enzyme function, evolution, and design. In this study, we leveraged emerging microfluidic technology to measure catalytic constants-kcat and KM-for hundreds of diverse orthologs and mutants of adenylate kinase (ADK). We dissected this sequence-catalysis landscape's topology, navigability, and mechanistic underpinnings, revealing catalytically heterogeneous neighborhoods organized by domain architecture. These results challenge long-standing hypotheses in enzyme adaptation, demonstrating that thermophilic enzymes are not universally slower than their mesophilic counterparts. Semisupervised models that combine our data with the rich sequence representations from large protein language models predict orthologous ADK-sequence catalytic parameters better than existing approaches. Our work demonstrates a promising strategy for dissecting sequence-catalysis landscapes across enzymatic evolution, opening previously unexplored avenues for enzyme engineering and functional prediction.

Giant tortoises exhibit exceptional longevity, often exceeding the human lifespan. To understand the genomic and epigenomic basis of their longevity, we analyzed the DNA sequence and methylome of Jonathan, an Aldabra giant tortoise (Aldabrachelys gigantea), estimated to be 192 years old. Relative to other giant tortoises (Aldabrachelys gigantea and Chelonoidis abingdonii), we found Jonathan has gene variants in pathways associated with aging, including DNA repair and telomere regulation. Consistent with his advanced age, Jonathan has significant age-related changes in DNA methylation and methylation entropy, compared with a 5-year-old Aldabra individual. Notably, we found that low entropy regions in Jonathan's methylome were enriched for genes involved in the electron transport chain. This suggests that high-fidelity transcription of these genes may be crucial for extreme longevity. With this data, we propose a model for aging, that links efficient mitochondrial energy production with nuclear maintenance of low methylation entropy. ### Competing Interest Statement The Regents of the University of California are the sole owner of patents and patent applications directed at epigenetic biomarkers for which Steve Horvath is a named inventor; SH is a founder and paid consultant of the non-profit Epigenetic Clock Development Foundation that licenses these patents. SH is a Principal Investigator at the Altos Labs, Cambridge Institute of Science. The other authors declare no competing interests.

Protein language models have emerged as powerful tools for learning rich protein representations, improving performance in tasks like structure prediction, mutation effect estimation, and homology detection. Their ability to model complex sequence distributions also holds promise for designing novel, functional proteins with broad applications in therapeutics, materials, and sustainability. However, due to the vastness of sequence space, efficient exploration methods are essential for protein engineering. Despite this, most existing protein design approaches using protein language models rely on single-mutant sampling strategies borrowed from Natural Language Processing, which fail to capture critical epistatic interactions between amino acid positions that are essential for protein function. In this work, we develop a comprehensive in silico protein design evaluation framework to systematically compare different sampling methods. After a thorough review of existing sampling strategies for language models, we introduce several approaches specifically tailored for protein design. We demonstrate that sampling strategies that consider multiple mutations simultaneously significantly outperform single-mutant approaches by better capturing epistatic effects between residue pairs. We evaluated these strategies using our framework, investigating the effects of key hyperparameters and providing practical guidance on the relative strengths of each method depending on design objectives. ### Competing Interest Statement The authors have declared no competing interest.

Understanding how molecular changes caused by genetic variation drive disease risk is crucial for deciphering disease mechanisms. However, interpreting genome sequences is challenging because of the vast size of the human genome, and because its consequences manifest across a wide range of cells, tissues and scales – spanning from molecular to whole organism level. Here, we present Phenformer, a multi-scale genetic language model that learns to generate mechanistic hypotheses as to how differences in genome sequence lead to disease-relevant changes in expression across cell types and tissues directly from DNA sequences of up to 88 million base pairs. Using whole genome sequencing data from more than 150 000 individuals, we show that Phenformer generates mechanistic hypotheses about disease-relevant cell and tissue types that match literature better than existing state-of-the-art methods, while using only sequence data. Furthermore, disease risk predictors enriched by Phenformer show improved prediction performance and generalisation to diverse populations. Accurate multi-megabase scale interpretation of whole genomes without additional experimental data enables both a deeper understanding of molecular mechanisms involved in disease and improved disease risk prediction at the level of individuals.

Recent breakthroughs in AI coupled with the rapid accumulation of protein sequence and structure data have radically transformed computational protein design. New methods promise to escape the constraints of natural and laboratory evolution, accelerating the generation of proteins for applications in biotechnology and medicine. To make sense of the exploding diversity of machine learning approaches, we introduce a unifying framework that classifies models on the basis of their use of three core data modalities: sequences, structures and functional labels. We discuss the new capabilities and outstanding challenges for the practical design of enzymes, antibodies, vaccines, nanomachines and more. We then highlight trends shaping the future of this field, from large-scale assays to more robust benchmarks, multimodal foundation models, enhanced sampling strategies and laboratory automation.

Designing novel functional proteins crucially depends on accurately modeling their fitness landscape. Given the limited availability of functional annotations from wet-lab experiments, previous methods have primarily relied on self-supervised models trained on vast, unlabeled protein sequence or structure datasets. While initial protein representation learning studies solely focused on either sequence or structural features, recent hybrid architectures have sought to merge these modalities to harness their respective strengths. However, these sequence-structure models have so far achieved only incremental improvements when compared to the leading sequence-only approaches, highlighting unresolved challenges effectively leveraging these modalities together. Moreover, the function of certain proteins is highly dependent on the granular aspects of their surface topology, which have been overlooked by prior models. To address these limitations, we introduce the Sequence-Structure-Surface Fitness (**S3F**) model — a novel multimodal representation learning framework that integrates protein features across several scales. Our approach combines sequence representations from a protein language model with Geometric Vector Perceptron networks encoding protein backbone and detailed surface topology. The proposed method achieves state-of-the-art fitness prediction on the ProteinGym benchmark encompassing 217 substitution deep mutational scanning assays, and provides insights into the determinants of protein function. Our code is at https://github.com/DeepGraphLearning/S3F.

Genetic studies reveal extensive disease-associated variation across the human genome, predominantly in noncoding regions, such as promoters. Quantifying the impact of these variants on disease risk is crucial to our understanding of the underlying disease mechanisms and advancing personalized medicine. However, current computational methods struggle to capture variant effects, particularly those of insertions and deletions (indels), which can significantly disrupt gene expression. To address this challenge, we present LOL-EVE (Language Of Life across EVolutionary Effects), a conditional autoregressive transformer model trained on 14.6 million diverse mammalian promoter sequences. Leveraging evolutionary information and proximal genetic context, LOL-EVE predicts indel variant effects in human promoter regions. We introduce three new benchmarks for indel variant effect prediction in promoter regions, comprising the identification of causal eQTLs, prioritization of rare variants in the human population, and understanding disruptions of transcription factor binding sites. We find that LOL-EVE achieves state-of-the-art performance on these tasks, demonstrating the potential of region-specific large genomic language models and offering a powerful tool for prioritizing potentially causal non-coding variants in disease studies. ### Competing Interest Statement The authors have declared no competing interest.

Quantitatively mapping enzyme sequence-catalysis landscapes remains a critical challenge in understanding enzyme function, evolution, and design. Here, we expand an emerging microfluidic platform to measure catalytic constants-k cat and K M-for hundreds of diverse naturally occurring sequences and mutants of the model enzyme Adenylate Kinase (ADK). This enables us to dissect the sequence-catalysis landscape's topology, navigability, and mechanistic underpinnings, revealing distinct catalytic peaks organized by structural motifs. These results challenge long-standing hypotheses in enzyme adaptation, demonstrating that thermophilic enzymes are not slower than their mesophilic counterparts. Combining the rich representations of protein sequences provided by deep-learning models with our custom high-throughput kinetic data yields semi-supervised models that significantly outperform existing models at predicting catalytic parameters of naturally occurring ADK sequences. Our work demonstrates a promising strategy for dissecting sequence-catalysis landscapes across enzymatic evolution and building family-specific models capable of accurately predicting catalytic constants, opening new avenues for enzyme engineering and functional prediction.