Protein Language Models (pLMs) have significantly advanced computational biology. Yet their scale and reliance on redundant training data raise a fundamental question: do pLMs generalize the statistical grammar of proteins, or do they simply memorize their training data? To investigate this, we used pseudoperplexity as a probe for sequence-level memorization, comparing ProtT5′s pseudoperplexity on a pre-training proxy dataset against a post-training holdout of genuinely novel sequences. To ensure a valid comparison, we matched the datasets by sequence length, cluster size, and taxonomic family. As a statistical baseline, we trained n -gram language models; analysis of higher-order n -gram composition and a statistically significant divergence in perplexity confirmed that the post-training sequences were genuinely novel at the local sequence level. ProtT5 showed a statistically significant difference in pseudoperplexity between seen and unseen sequences, though further analysis revealed this memorization signal to be modest. These findings suggest that ProtT5 exhibits detectable but limited memorization of its training data as measured by a pseudoperplexity-based probe.

Proteins fold into their native three-dimensional (3D) structures by navigating complex energy landscapes shaped by the biophysical and biochemical properties of their sequence. Once folded, some sequence positions (dubbed residues) remain locally frustrated, reflecting functional constraints incompatible with optimal packing. This local energetic frustration provides important insights into protein function and dynamics, but its analysis typically relies on structure-based energy calculations and remains energetically costly at scale. Here, we introduce an ultra-fast sequence-based prediction of local energetic frustration directly from protein sequences using embeddings from protein language models (pLMs). Our method, coined FrustrAI-Seq, enables proteome-wide frustration profiling in minutes (17 minutes for the entire human proteome on a single Nvidia H100 GPU) while retaining biologically relevant performance as shown for the alpha-globin and beta-lactamase family. By eliminating the need for explicit structural or evolutionary information, this approach expands frustration analysis to protein regions and classes that were previously inaccessible, including intrinsically disordered regions and high-throughput de novo designed protein datasets. To support reproducibility and large-scale applications, we provide the largest freely available resource of precomputed local frustration scores to date (10^6 proteins), along with model weights and complete training and inference code at: github.com/leuschjanphilipp/FrustrAI-Seq. ### Competing Interest Statement The authors have declared no competing interest. Helmholtz Zentrum München, https://ror.org/00cfam450 Helmholtz Association’s Initiative Ramon y Cajal program MEGAFrustratEDS Barcelona Supercomputing Center, https://ror.org/05sd8tv96 NVIDIA Corporation European Social Fund Plus (ESF+) Spanish Ministry of Science, Innovation and Universities (MICIU)

Motivation: Proteins function through motion. Yet, most discoveries still commence with static representations of protein structures. Here, we investigated the feasibility of leveraging protein dynamics to improve homology detection. Results: We introduce ProtProfileMD, a sequence-to-3D-probability model that predicts, from an amino acid sequence, a profile of discrete structural representations capturing protein dynamics. We applied supervised parameter-efficient finetuning of the ProstT5 protein Language Model (pLM) to predict per-residue distributions over Foldseek's 3Di alphabet derived from motions observed in molecular dynamics. This original result reveals that the 3Di tokens, despite being coarse-grained descriptors of 3D structure, still offer sufficient resolution to capture aspects of conformational changes. This is evidenced by a correlation between fluctuations in the 3D protein structure over the course of a molecular dynamics trajectory and the entropy of 3Di states. Based on this insight, we introduce a proof-of-concept for making remote homology detection of proteins more sensitive by leveraging a protein's distinctive dynamic fingerprint captured by our model. Our method recovers flexibility signals with a fidelity that is biologically relevant, improving search and complementing protein structure predictions, for example, by flagging flexible, disordered, or other functionally relevant regions. Availability and Implementation: ProtProfileMD is available at github.com/finnlueth/ProtProfileMD. The associated training data and model weights are available at huggingface.co/datasets/finnlueth/ProtProfileMD and huggingface.co/finnlueth/ProtProfileMD. ### Competing Interest Statement The authors have declared no competing interest. BMBF: Bundesministerium für Bildung und Forschung); Deutsche Forschungsgemeinschaft, DFG-GZ: RO1320/4-1 National Research Foundation of Korea, RS-2023-00250470, RS-2020-NR049543, RS-2021-NR061659, RS-2021-NR056571, RS-2024-00396026 Novo Nordisk Foundation, NNF24SA0092560

The rise of protein Language Models (pLMs) is reshaping the landscape of protein prediction. Embeddings are powerful protein representations provided by pLMs, but they come at a cost: their generation requires expensive hardware, and leveraging models often requires expert knowledge. To some extent, these hurdles limit the ease of use and benefits of those methods both for experimental and computational biologists. Biocentral aims at providing a free and open embedding-based service, which addresses these challenges. We support standardized access to most pLMs currently in use, enabling researchers to generate embeddings, get embedding-based protein feature predictions, and train embedding-based models. Here, we showcase biocentral in a large-scale analysis of the BFVD virus database through biocentral’s predict module. We also show how readily biocentral’s training module reproduces an existing embedding-based prediction method. The server is accessible through a graphical user interface and a programmatic Application Programming Interface (API) at: https://biocentral.rostlab.org.

The Critical Assessment of Functional Annotation (CAFA) is a long-standing community effort to independently assess computational methods for protein function prediction, to highlight well-performing methodologies, to identify bottlenecks in the field, and to provide a forum for the dissemination of results and exchange of ideas. In its fifth round (CAFA5) of triennial challenges, a partnership with Kaggle Inc. facilitated participation from a large community of data scientists and computational biologists through a competitive prospective challenge on the crowdsourcing platform. In this work, we present an in-depth analysis of the submitted predictions and report improvements in accuracy over all methods from the previous CAFA challenges. We further introduce a new evaluation setting for proteins with pre-existing (incomplete) annotations and identify the need for methods that better leverage existing annotations to predict those that will be discovered later. Finally, we characterize the prospective evaluation framework by examining performance on a strict set of unpublished annotations and across intermediate database releases. Our results indicate that recent developments in the field, such as the availability of protein language models and accurately predicted 3D structures, as well as the growth of experimental annotations through biocuration, have all contributed to performance improvements. 1.

Understanding the relationships among amino acid sequences, structures and functions in proteins and how they evolve, remains a central challenge in molecular biology. It is still unclear which sequence elements differentially contribute to structural integrity or molecular function. Even more, there are ongoing debates on whether protein folds emerge as a result of evolution or as a consequence of physical laws. The energy landscapes theory states that proteins are minimally frustrated systems, i.e. they fold by minimising their energetic conflicts. However, some local frustration, believed to be selected for functional reasons, remains in the native state of proteins. Here, we combine reverse folding and structure prediction methods with sequence and local frustration analysis to address the aforementioned ideas. We found that reverse folding techniques are unable to erase evolutionary conserved frustration from certain residues, even when detrimental for structural integrity. We propose that certain frustration hotspots behave like architectural spandrels, not directly shaped by selection but emerging from physical constraints in protein folds which evolution can later co-opt for function. Our results provide a new perspective revealing how sequence variation and functional specificity could evolve from evolutionary, structural and biophysical constraints. ### Competing Interest Statement The authors have declared no competing interest.

Motivation Infectious diseases continue to be a leading cause of mortality and pose a significant global health threat. Thus, the development of tools for surveillance and early detection of emerging pathogens is needed.Results We introduce PathogenFinder2, a novel, alignment-free, taxonomy-agnostic model for predicting bacterial pathogenic capacity in humans using protein language models. It outperforms previous methods, particularly for novel taxa, and provides interpretable outputs by highlighting proteins most relevant to pathogenic potential. These insights aid the identification of virulence factors, vaccine targets, and infection-related metabolic pathways. Furthermore, we introduce the Bacterial Pathogenic Capacity Landscape, which reveals patterns linked to host condition, infection site, microbial antagonism, and environmental origin.Availability The model is freely available online at https://genepi.dk/pathogenfinder2, or as a standalone program (https://github.com/genomicepidemiology/PathogenFinder2).

Abstract Recent progress in modeling techniques and high-throughput screening has significantly enhanced the accessibility of protein engineering. Nevertheless, further progress gets hindered by the lack of robust benchmarks that capture the practical challenges for real-world protein engineering. Here, we introduced DeePEn, a De pth-s e nsitive benchmark for P rotein En gineering that quantifies a model’s generalization capabilities when predicting protein fitness at increasing mutational distance from the wildtype or training data. We defined distance as the number of simultaneous point mutations, i.e., single amino acid variants (SAVs), moving from wild-type to mutant ( edit distance in computer science jargon). Specifically selecting four deep mutational scanning (DMS) datasets with sufficient multi-mutation data points from ProteinGym, we assessed recent predictive models, including general and biophysics-informed protein Language Models (pLMs), and a non-transformer neural network. Our results highlight how the performance of all models deteriorates with increasing mutational distance and that no single metric sufficiently captures the diverse requirements of protein engineering. To overcome these shortcomings, DeePEn provides a readily available resource for multi-metric benchmarking that focuses on the prediction of distant variants.

Background:The Critical Assessment of Functional Annotation (CAFA) is a community effort held to understand the field of computational protein function prediction. Every three years, since 2010, the organizers initiate an experiment to collect function predictions on a large set of proteins and then evaluate the performance of predicting methods on a subset of proteins that have accumulated experimental annotations between the submission deadline and the evaluation time. CAFA provides an independent and rigorous assessment of the current state of the art, thus leveling the playing field, highlighting successes, revealing bottlenecks, and offering a forum for the exchange of ideas in protein science. Here, we report the results of the fourth CAFA experiment (CAFA4). Results:CAFA4 featured the participation of 148 methods from 70 research groups on a total of 46,205 unique proteins over a 5-year annotation accumulation phase, the longest in any CAFA. In a comparison across CAFA2-CAFA4 methods, the prediction of Gene Ontology (GO) terms has clearly improved across all three GO aspects and traditional evaluation settings. While not achieving the first rank, several CAFA2 and CAFA3 methods featured in the top ten methods in many evaluations, suggesting that earlier methods still hold relevance. The performance is weaker in the newly introduced "partial knowledge" evaluation category (proteins with experimental annotations before submission deadline that gained additional annotations in the same GO aspect during the annotation accumulation phase), highlighting the need for a new class of methods. The rankings of the methods were stable over the years in traditional evaluation settings, but less so in the new partial knowledge evaluation. Overall, the field continues to progress with some influx of new participants. Sustained efforts will be necessary to substantially advance it.

A bstract Protein Language Models (pLMs) generate per-protein embeddings that encode functional, structural, and evolutionary information, yet the relationships captured in these representations remain difficult to explore systematically. ProtSpace ( https://protspace.app ) is a web application for interactive visualization of pLM embedding spaces, enabling hypothesis generation directly in the browser without installation. Unlike traditional network-based tools that exclusively visualize amino acid sequence similarity, ProtSpace explores embedding spaces, revealing relationships often not captured by traditional comparisons. Users provide protein sequences or pre-computed embeddings through a Google Colab notebook or the Python CLI; the pipeline applies dimensionality reduction, retrieves 38 annotation types spanning UniProt, InterPro, NCBI Taxonomy, TED structural domains, and sequence-based predictors served via Biocentral, and produces a portable binary file for the browser-based viewer. WebGL-accelerated rendering supports interactive exploration of over 570,000 proteins. Distinctive features include per-point pie charts for multi-label annotations and integrated 3D structure viewing through AlphaFold2 predictions. All computation happens on the user’s machine, ensuring data privacy. We demonstrate the utility of ProtSpace through a progressive zoom-in across biological scales: from global proteome organization of Swiss-Prot, through cross-species comparison revealing conserved and lineage-specific families, to functional hypothesis generation within the beta-lactamase superfamily. ProtSpace is freely available at https://protspace.app under the Apache 2.0 license. K ey points ProtSpace is a free, open-source web application that visualizes protein Language Model (pLM) embeddings as interactive maps, scaling to 570,000 proteins entirely client-side. A zero-installation Google Colab notebook and a Python CLI prepare visualization-ready bundles from FASTA files, UniProt queries, or pre-computed HDF5 embeddings, automatically retrieving 38 annotation types from five sources (UniProt, InterPro, NCBI Taxonomy, TED structural domains, and Biocentral sequence predictors) alongside custom CSV metadata. Application examples demonstrate that embedding visualizations generate testable biological hypotheses at multiple scales, from proteome-wide organization through species-level comparison to family-level functional discovery, and that these are complementary to traditional sequence-based analyses.

Motivation: Regions in intrinsic disordered proteins (IDPs) constitute important continuous aspects of protein function. While their existence on a structural continuum is widely accepted, most computational predictions have, nevertheless, focused on binary classifications. Existing datasets are severely limited in size and experimental evidence for continuous disorder. Results: Building on recently released datasets of continuous protein disorder and flexibility, we introduce UdonPred, a lightweight neural network exclusively inputting embeddings from the protein Language Model (pLM) ProstT5 to predict per-residue protein disorder from sequence alone. Training and evaluating UdonPred on seven datasets with divergent definitions of disorder and flexibility suggests that not model capacity, but agreement and nuance of disorder annotations, remains the main driver of performance. Binary disorder annotations can be reliably predicted from a multitude of different disorder and flexibility datasets, but there is still room for improvement in predicting continuous disorder. ### Competing Interest Statement The authors have declared no competing interest.

Protein language models have become essential tools for engineering novel functional proteins. The emerging paradigm of family-based language models makes use of homologous sequences to steer protein design and enhance zero-shot fitness prediction, by imbuing models with an ability to explicitly reason over evolutionary context. To provide an open foundation for this modelling approach, we introduce ProFam-1 , a 251M-parameter autoregressive protein family language model (pfLM) trained with next-token prediction on millions of protein families represented as concatenated, unaligned sets of sequences. ProFam-1 is competitive with state-of-the-art models on the ProteinGym zero-shot fitness prediction benchmark, achieving Spearman correlations of 0.47 for substitutions and 0.53 for indels. For homology-guided generation, ProFam-1 generates diverse sequences with predicted structural similarity, while preserving residue conservation and covariance patterns. All of ProFam’s training and inference pipelines, together with our curated, large-scale training dataset ProFam Atlas , are released fully open source, lowering the barrier to future method development. ### Competing Interest Statement The authors have declared no competing interest. Wellcome Trust, 218529/Z/19/Z Wellcome EPSRC Centre for Medical Engineering, https://ror.org/04s9kfd63, EP/S021566/

Understanding and comparing three-dimensional (3D) structures of proteins can advance bioinformatics, molecular biology, and drug discovery. While 3D models offer detailed insights, comparing multiple structures simultaneously remains challenging, especially on two-dimensional (2D) displays. Existing 2D visualization tools lack standardized approaches for pipelined inspection of large protein sets, limiting their utility in large-scale pre-filtering. We introduce FlatProt, a tool designed to complement 3D viewers by enabling standardized 2D visualization of individual protein structures or large sets thereof. By including Foldseek-based family rotation alignment or an inertia-based fallback, FlatProt creates consistent and scalable visual representations for user-defined protein structures. It supports domain-aware decomposition, family-level overlays, and lightweight visual abstraction of secondary structures. FlatProt processes proteins efficiently, as showcased on a subset of the human-proteome. FlatProt provides clear, consistent, user-friendly visualizations that support rapid, comparative inspection of protein structures at scale. By bridging the gap between interactive 3D tools and static visual summaries, it enables users to explore conserved features, detect outliers, and prioritize structures for further analysis. GitHub ( https://github.com/t03i/FlatProt ); Zenodo ( https://doi.org/10.5281/zenodo.15697296 ).

Protein language models (pLMs) capture some aspects of the grammar of the language of life as written in protein sequences. The so-called pLM embeddings implicitly contain this information. Therefore, embeddings can serve as the exclusive input into downstream supervised methods for protein prediction. Over the last 33 years, evolutionary information extracted through simple averaging for specific protein families from multiple sequence alignments (MSAs) has been the most successful universal key to the success of protein prediction. For many applications, MSA-free pLM-based predictions now have become significantly more accurate. The reason for this is often a combination of two aspects. Firstly, embeddings condense the grammar so efficiently that downstream prediction methods succeed with small models, i.e., they need few free parameters in particular in the era of exploding deep neural networks. Secondly, pLM-based methods provide protein-specific solutions. As additional benefit, once the pLM pre-training is complete, pLM-based solutions tend to consume much fewer resources than MSA-based solutions. In fact, we appeal to the community to rather optimize foundation models than to retrain new ones and to evolve incentives for solutions that require fewer resources even at some loss in accuracy. Although pLMs have not, yet, succeeded to entirely replace the body of solutions developed over three decades, they clearly are rapidly advancing as the universal key for protein prediction.

Large Language Models for proteins, namely protein Language Models (pLMs), have begun to provide an important alternative to capturing the information encoded in a protein sequence in computers. Arguably, pLMs have advanced importantly to understanding aspects of the language of life as written in proteins, and through this understanding, they are becoming an increasingly powerful means of advancing protein prediction, e.g., in the prediction of molecular function as expressed by identifying binding residues or variant effects. While benefitting from the same technology, protein structure prediction remains one of the few applications for which only using pLM embeddings from single sequences appears not to improve over or match the state-of-the-art. Fine-tuning foundation pLMs enhances efficiency and accuracy of solutions, in particular in cases with few experimental annotations. pLMs facilitate the integration of computational and experimental biology, of AI and wet-lab, in particular toward a new era of protein design.

Regular, systematic, and independent assessments of computational tools that are used to predict the pathogenicity of missense variants are necessary to evaluate their clinical and research utility and guide future improvements. The Critical Assessment of Genome Interpretation (CAGI) conducts the ongoing Annotate-All-Missense (Missense Marathon) challenge, in which missense variant effect predictors (also called variant impact predictors) are evaluated on missense variants added to disease-relevant databases following the prediction submission deadline. Here we assess predictors submitted to the CAGI 6 Annotate-All-Missense challenge, predictors commonly used in clinical genetics, and recently developed deep learning methods. We examine performance across a range of settings relevant for clinical and research applications, focusing on different subsets of the evaluation data as well as high-specificity and high-sensitivity regimes. Our evaluations reveal notable advances in current methods relative to older, well-cited tools in the field. While meta-predictors tend to outperform their constituent individual predictors, several newer individual predictors perform comparably to commonly used meta-predictors. Predictor performance varies between high-specificity and high-sensitivity regimes, highlighting that different methods may be optimal for different use cases. We also characterize two potential sources of bias. Predictors that incorporate allele frequency as a predictive feature tend to have reduced performance when distinguishing pathogenic variants from very rare benign variants, and predictors trained on pathogenicity labels from curated variant databases often inherit gene-level label imbalances. Our findings help illuminate the clinical and research utility of modern missense variant effect predictors and identify potential areas for future development.

Existing toxins datasets include a mixture of proteins and toxin peptides. In this study we present two curated datasets of toxic proteins free of associated proteins: bacterial exotoxins and animal toxins. Our stringent selection criteria resulted in two data sets with only toxins that directly target or disrupt vital molecular mechanisms of their target organism. To gain insight in their properties and differences, we compared both sets of toxins to controls, and used simple biophysical features such as protein length, and amino acid composition distinguishing between evolutionary kingdoms (phyla). This approach should reveal if there is a need to consider differences present in toxin sets depending on their origin. Our analysis revealed biophysical differences between animal and bacterial toxins that should not be ignored. Both toxin groups exhibited preferencial amino acid usage depending on their origin, together with higher cysteine content compared to their respective controls. The animal toxins set contains, on average, significantly shorter sequences than bacterial toxins, and had their isoelectric point shifted towards acidic pH values. We show that animal toxins and bacterial toxins possess intrinsic differences in general biophysical properties, which reinforce the necessity of segregating these datasets to ensure reliability of bioinformatics models aimed at understanding and predicting toxin characteristics.

Protein language models (pLMs) generate high-dimensional representations of proteins, so called embeddings, that capture complex information stored in the set of evolved sequences. Interpreting these embeddings remains an important challenge. ProtSpace provides one solution through an open-source Python package that visualizes protein embeddings interactively in 2D and 3D. The combination of embedding space with protein 3D structure view aids in discovering functional patterns readily missed by traditional sequence analysis.We present two examples to showcase ProtSpace. First, investigations of phage data sets showed distinct clusters of major functional groups and a mixed region, possibly suggesting bias in today’s protein sequences used to train pLMs. Second, the analysis of venom proteins revealed unexpected convergent evolution between scorpion and snake toxins; this challenges existing toxin family classifications and added evidence refuting the aculeatoxin family hypothesis.ProtSpace is freely available as a pip-installable Python package (source code & documentation) with examples on GitHub (https://github.com/tsenoner/protspace) and as a web interface (https://protspace.rostlab.org). The platform enables seamless collaboration through portable JSON session files.

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.