ABSTRACT General-purpose frontier language models are being increasingly utilized for protein-design work, yet their ability to understand and evaluate variant effects remains unclear. Here, we introduce PG-LLM, a benchmark comprising 276 protein-variant prioritization tasks: 217 from ProteinGym and a temporally held-out set of 59 from recently published studies. Each task follows the same format: a language model is asked to rank a list of variant sequences given only the wild-type protein sequence and an assay description with no access to tools, multiple-sequence alignments, or protein structures. We evaluate thirteen language models and 95 published protein predictors on the same variants with the same evaluation metric. Claude Opus 5 (Max) and GPT 5.6 Sol (Max) are the best performing LLMs with Spearman correlations of ρ = 0.406 and 0.402 respectively. Opus 5 outperforms 49 of 95 published protein predictors, including 41 of 46 sequence-only methods, and approaches ESM2-650M at ρ = 0.411, but remains below the leading predictor VenusREM at ρ = 0.523. We observe that variant-ranking performance scales with test-time compute across GPT, Claude, and Gemini models, but gains taper before closing the gap to specialist protein predictors. To address contamination risk, we create a held-out evaluation set with 59 DMS assays from 19 studies whose scores first became public after January 2026. On this set, we observe performance and test time compute scaling trends similar to those on the 217 tasks derived from ProteinGym. PG-LLM shows that tool-free language models capture substantial protein-variant signal, outperforming many sequence-based predictors while remaining below the strongest specialized models.

Machine learning (ML) has the potential to revolutionize antibody design and selection, but its success depends on access to well-curated datasets of antibody-antigen interactions. We developed a synthetic Fab yeast display library optimized for seamless integration with ML processes, focusing on sequence diversity within the complementary determining region heavy chain CDRH3 loop. The library incorporates key sequence features derived from human B cell repertoires captured in a compact antigen recognition module (ARM) format. Built with the VH1-69 heavy chain and four light chains, the library was evaluated against ten human and murine cell surface antigens, including programmed cell death ligand 1 (PD-L1), T cell immunoreceptor with immunoglobulin and immunoreceptor tyrosine-based inhibitory motif domains (TIGIT), and roundabout guidance receptor 1 (ROBO1). This approach yielded hundreds of antibodies with robust biophysical properties, some of which were validated by flow cytometry and immunohistochemistry. Furthermore, ML analysis identified additional antibodies for ROBO2 and PD-L2 from the aggregate sequencing data. The publicly available dataset establishes an ML-compatible framework designed to accelerate and streamline antibody discovery and development. A record of this paper’s transparent peer review process is included in the supplemental information.

Antibody folding and aggregation are major challenges in the development of relevant reagents and therapeutics. Antibodies face a biophysical trade-off; the immense diversity in complementarity-determining regions (CDRs), which is crucial for broad antigen recognition, comes at the cost of folding stability. How CDR sequences influence antibody folding remains poorly understood because of their sequence diversity and lack of large-scale data. Here we develop a high-throughput 'deep loop profiling' approach to quantify folding fitness across millions of diverse CDRs. Machine learning models trained on this dataset predict folding propensity directly from sequence and identify interpretable residue-level rules that reveal CDR1 and CDR2 as key folding determinants. Using these insights, we rescue two unstable nanobodies, including an aggregation-prone SARS-CoV-2 binder and a G-protein-coupled receptor-targeting intrabody, and build next-generation synthetic libraries enriched for biophysically optimized nanobodies. This approach provides a scalable framework for understanding and engineering folding competence in antibody-based scaffolds.

New SARS-CoV-2 variants have undergone repeated selective sweeps since the beginning of the COVID-19 pandemic, but the fitness advantages and mechanisms driving these sweeps are not fully understood. We developed a probabilistic modeling framework to analyze pandemic growth, infectivity, and immune escape, explicitly accounting for seven immune exposure histories in 5,732 experiments and estimating the effects of 835 mutations. We found infectivity was important for early variants, but as gains became zero-sum, growth became driven by consistent increase in immune escape conferred by a primarily additive effect of mutational accumulation. While phenotypic tradeoffs exist for individual mutations, successful viral strains boast assemblages of mutations that do not sacrifice infectivity for escape. Thus, during an apparent transition to endemicity, SARS-CoV-2 evolution ascended along an evolutionary ridge in the mutational space defined by infectivity and escape, with infectivity reaching an early peak and antigenicity continuing to evolve.

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.

Antibodies are central mediators of the adaptive immune response, and they are powerful research tools and therapeutics. Antibody discovery requires substantial experimental effort, such as immunization campaigns or in vitro library screening. Predicting antibody-antigen binding a priori remains challenging. However, recent machine learning methods raise the possibility of in silico antibody discovery, bypassing or reducing initial experimental bottlenecks. Here, we report a virtual screen using AlphaFold-Multimer (AF-M) that prospectively identified nanobody binders to MRGPRX2, a G protein-coupled receptor (GPCR) and therapeutic target for the treatment of pseudoallergic inflammation and itch. Using previously reported nanobody-GPCR structures, we identified a set of AF-M outputs that effectively discriminate between interacting and non-interacting nanobody-GPCR pairs. We used these outputs to perform a prospective in silico screen, identified nanobodies that bind MRGPRX2 with high affinity, and confirmed activity in signaling and functional cellular assays. Our results provide a proof of concept for fully computational antibody discovery pipelines that can circumvent laboratory experiments.

Many proteins useful in modern medicine or bioengineering are challenging to make in the lab, fuse with other proteins in cells, or deliver to tissues in the body because their sequences are too long. Shortening these sequences typically involves costly, time-consuming experimental campaigns. Ideally, we could instead use modern models of massive databases of sequences from nature to learn how to propose shrunken proteins that resemble sequences found in nature. Unfortunately, these models struggle to efficiently search the combinatorial space of all deletions, and are not trained with inductive biases to learn how to delete. To address this gap, we propose SCISOR, a novel discrete diffusion model that deletes letters from sequences to generate protein samples that resemble those found in nature. To do so, SCISOR trains a de-noiser to reverse a forward noising process that adds random insertions to natural sequences. As a generative model, SCISOR fits evolutionary sequence data competitively with previous large models. In evaluation, SCISOR achieves state-of-the-art predictions of the functional effects of deletions on ProteinGym. Finally, we use the SCISOR de-noiser to shrink long protein sequences, and show that its suggested deletions result in significantly more realistic proteins and more often preserve functional motifs than previous models of evolutionary sequences.

Many proteins useful in modern medicine or bioengineering are challenging to make in the lab, fuse with other proteins in cells, or deliver to tissues in the body, because their sequences are too long. Shortening these sequences typically involves costly, time-consuming experimental campaigns. Ideally, we could instead use modern models of massive databases of sequences from nature to learn how to propose shrunken proteins that resemble sequences found in nature. Unfortunately, these models struggle to efficiently search the combinatorial space of all deletions, and are not trained with inductive biases to learn how to delete. To address this gap, we propose SCISOR, a novel discrete diffusion model that deletes letters from sequences to generate protein samples that resemble those found in nature. To do so, SCISOR trains a de-noiser to reverse a forward noising process that adds random insertions to natural sequences. As a generative model, SCISOR fits evolutionary sequence data competitively with previous large models. In evaluation, SCISOR achieves state-of-the-art predictions of the functional effects of deletions on ProteinGym. Finally, we use the SCISOR de-noiser to shrink long protein sequences, and show that its suggested deletions result in significantly more realistic proteins and more often preserve functional motifs than previous models of evolutionary sequences.

Viruses are a major threat to global health due to their rapid evolution, extensive diversity, and frequent cross-species transmission. Although advances in machine learning and the expanding availability of sequence and structural data have accelerated large-scale mutation effect prediction, viral proteins, and particularly fast-evolving antigenic proteins, pose unique biological and data-related challenges that may limit model performance. We introduce EVEREST, a curated dataset for evaluating model performance on (i) forecasting real-world viral evolution (31 clades across 4 viruses) and (ii) concordance with lab-based deep mutational scanning assays (45 proteins, $>$340,000 variants). Using EVEREST, we show that state-of-the-art protein language models trained across the protein universe substantially under-perform on viral proteins relative to alignment-based models trained on homologous proteins. This under-performance persists even in low-sequence regimes, as is the case during a novel viral outbreak. We develop calibrated reliability metrics to quantify confidence in model predictions where no evaluation datasets exist. For more than half of the WHO-prioritized pandemic-threat viruses, current models fail to produce reliable predictions, highlighting the urgent need for more data or new modeling approaches. Together, these findings reveal key factors driving model under-performance and provide actionable recommendations for improving viral mutation effect prediction in preparation for current and future outbreaks.

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.

The relationship between pH and enzyme catalytic activity, especially the optimal pH (pHopt) at which enzymes function, is critical for biotechnological applications. Hence, computational methods to predict pHopt will enhance enzyme discovery and design by facilitating accurate identification of enzymes that function optimally at specific pH levels, and by elucidating sequence–function relationships. Here we proposed and evaluated various machine learning methods for predicting pHopt, conducting extensive hyperparameter optimization and training over 11,000 model instances. Our results demonstrate that models utilizing language model embeddings markedly outperform other methods in predicting pHopt. We present EpHod, the best-performing model, to predict pHopt, making it publicly available to researchers. From sequence data, EpHod directly learns structural and biophysical features that relate to pHopt, including proximity of residues to the catalytic centre and the accessibility of solvent molecules. Overall, EpHod presents a promising advancement in pHopt prediction and will potentially speed up the development of enzyme technologies. Accurately predicting the optimal pH level for enzyme activity is challenging due to the complex relationship between enzyme structure and function. Gado and colleagues show that a language model can effectively learn the structural and biophysical features to predict the optimal pH for enzyme activity.

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.

Applying machine learning to biological sequences - DNA, RNA and protein - has enormous potential to advance human health, environmental sustainability, and fundamental biological understanding. However, many existing machine learning methods are ineffective or unreliable in this problem domain. We study these challenges theoretically, through the lens of kernels. Methods based on kernels are ubiquitous: they are used to predict molecular phenotypes, design novel proteins, compare sequence distributions, and more. Many methods that do not use kernels explicitly still rely on them implicitly, including a wide variety of both deep learning and physics-based techniques. While kernels for other types of data are well-studied theoretically, the structure of biological sequence space (discrete, variable length sequences), as well as biological notions of sequence similarity, present unique mathematical challenges. We formally analyze how well kernels for biological sequences can approximate arbitrary functions on sequence space and how well they can distinguish different sequence distributions. In particular, we establish conditions under which biological sequence kernels are universal, characteristic and metrize the space of distributions. We show that a large number of existing kernel-based machine learning methods for biological sequences fail to meet our conditions and can as a consequence fail severely. We develop straightforward and computationally tractable ways of modifying existing kernels to satisfy our conditions, imbuing them with strong guarantees on accuracy and reliability. Our proof techniques build on and extend the theory of kernels with discrete masses. We illustrate our theoretical results in simulation and on real biological data sets.

Computational methods for assessing the likely impacts of mutations, known as variant effect predictors (VEPs), are widely used in the assessment and interpretation of human genetic variation, as well as in other applications like protein engineering. Many different VEPs have been released, and there is tremendous variability in their underlying algorithms, outputs, and the ways in which the methodologies and predictions are shared. This leads to considerable difficulties for users trying to navigate the selection and application of VEPs. Here, to address these issues, we provide guidelines and recommendations for the release of novel VEPs.

Predicting and quantifying phenotypic consequences of genetic variants in rare disorders is a major challenge, particularly pertinent for ‘actionable’ genes such as thyroid hormone transporter MCT8 (encoded by the X-linked SLC16A2 gene), where loss-of-function (LoF) variants cause a rare neurodevelopmental and (treatable) metabolic disorder in males. The combination of deep phenotyping data with functional and computational tests and with outcomes in population cohorts, enabled us to: (i) identify the genetic aetiology of divergent clinical phenotypes of MCT8 deficiency with genotype-phenotype relationships present across survival and 24 out of 32 disease features; (ii) demonstrate a mild phenocopy in ~400,000 individuals with common genetic variants in MCT8; (iii) assess therapeutic effectiveness, which did not differ among LoF-categories; (iv) advance structural insights in normal and mutated MCT8 by delineating seven critical functional domains; (v) create a pathogenicity-severity MCT8 variant classifier that accurately predicted pathogenicity (AUC:0.91) and severity (AUC:0.86) for 8151 variants. Our information-dense mapping provides a generalizable approach to advance multiple dimensions of rare genetic disorders.

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.

Multiplexed assays of variant effect (MAVEs) are a critical tool for researchers and clinicians to understand genetic variants. Here we describe the 2024 update to MaveDB ( https://www.mavedb.org/ ) with four key improvements to the MAVE community's database of record: more available data including over 7 million variant effect measurements, an improved data model supporting assays such as saturation genome editing, new built-in exploration and visualization tools, and powerful APIs for data federation and streamlined submission and access. Together these changes support MaveDB's role as a hub for the analysis and dissemination of MAVEs now and into the future.

Pancreatic ductal adenocarcinoma (PDAC) is a rare, aggressive cancer often diagnosed late with low survival rates, due to the lack of population-wide screening programs and the high cost of early detection methods. To enable early detection of high-risk individuals, we develop a transformer-based model trained on longitudinal Veterans Affairs electronic health record (EHR) with 19,426 PDAC cases and ∼15.9 million controls. Our model combines diagnostic and medication trajectories to predict PDAC risk within a 6-, 12-, and 36-month assessment window. Incorporating medication significantly improved performance; among the top 1,000-5,000 highest-risk patients in a cohort of 1 million patients, 3-year PDAC incidence is 115-70 times higher than a reference estimate based on age and sex alone. Furthermore, analysis of most predictive features highlights the role of events such as chronic inflammatory conditions and specific medications on overall PDAC risk. Our work provides an AI-driven identification of high-risk individuals, with a potential to improve early detection, enhance patient care, and reduce healthcare costs.