Now, a team of researchers from Universidad de Magallanes in Punta Arenas, Chile, the Data Science Institute at Universidad del Desarrollo in Santiago, and PEACCEL in Paris has unveiled MAOMAO (Metadata-Aware Ontology for Multi-source Annotation Organization), an ontology-guided resource designed to bring coherence to the fragmented landscape of peptide toxicity data. Published in Frontiers in Pharmacology as a data report in the Predictive Toxicology section, the resource integrates toxicity information from 56 publicly available databases, literature-derived resources, supplementary datasets, and manually curated repositories into a unified, FAIR-oriented framework that preserves evidence, provenance, and metadata at every step.

The problem MAOMAO addresses is deceptively simple but technically profound. Toxicity annotations for peptides have accumulated across public databases, literature-derived resources, supplementary datasets, and individual experimental studies, but these sources employ heterogeneous terminology, annotation criteria, and metadata standards. A peptide described as haemolytic in one database might appear as hemolytic in another, while cytolytic activity in one source could correspond to cytolysis in a different resource. These inconsistencies complicate integration, reproducibility, and large-scale analysis, and they can introduce annotation ambiguity, conflicting evidence, incomplete metadata, and limited provenance tracking that reduce the reliability of downstream computational reuse.

MAOMAO standardizes toxicity terminology by mapping alternative spellings, synonyms, and source-dependent nomenclature to a unified vocabulary while preserving the original biological meaning. The resource then organizes harmonized annotations within a hierarchical toxicity ontology centered on the root category Toxic, which comprises four primary classes: Cytotoxic, Neurotoxic, Embryotoxic, and Ichthyotoxic. The Cytotoxic branch is further subdivided into Hemolytic, Cytolysis, and Anti-mammalian-cell activity, preserving more specific biological contexts. Hemolytic denotes toxicity involving erythrocyte lysis, whereas Cytolysis captures cell lysis or membrane-disruptive activity without a specified cell type, and anti-mammalian-cell activity denotes damage to mammalian cells or reduced cell viability.

The construction workflow began with structured exploration of the peptide toxicity landscape using scientific search engines, bibliographic repositories, specialized peptide databases, review articles, and citation tracking. Searches combined general terms such as peptide toxicity, toxic peptides, and toxic peptide databases with endpoint-specific queries for hemolytic, cytotoxic, cytolytic, neurotoxic, embryotoxic, ichthyotoxic, and anti-mammalian-cell peptides. Candidate resources were evaluated against predefined eligibility criteria, including relevance to at least one MAOMAO toxicity endpoint, availability of peptide sequences with corresponding categorical toxicity annotations or quantitative toxicity measurements, sufficient provenance information, and the possibility of programmatic or manual retrieval. Resources lacking retrievable sequences or containing only numerical representations, peptide names, or functional annotations without corresponding sequences were excluded.

Retrieved records underwent standardized preprocessing that included sequence validation, format standardization, annotation parsing, identifier normalization, and quality control. The final resource was restricted to peptides containing the 20 canonical amino acids and lengths between 5 and 70 residues. Exact duplicates within each source were removed, while identical sequences reported by different sources were retained to preserve provenance and support evidence integration. Each normalized sequence was assigned a deterministic identifier generated by applying SHA-256 to the UTF-8 encoding of the uppercase, whitespace-free sequence, providing a stable key for linking records across all resource layers. Quantitative HC50, LC50, LD50, and MHC observations were parsed while preserving their reported relation, value, unit, and provenance.

Because the same peptide sequence could be reported with different toxicity annotations across sources, MAOMAO applied an evidence-management framework to classify each peptide-endpoint pair as positive, negative, ambiguous, unlabeled, or no information. Positive states represented direct or hierarchy-supported toxicity evidence, whereas negative states corresponded to direct endpoint-specific non-toxic evidence. Ambiguous states captured conflicting or unresolved annotations, unlabeled states denoted sequences retained in endpoint-specific sources without an explicit class assignment, and no information indicated the absence of evidence for that endpoint. Negative evidence was further classified as strong or weak according to the supporting information available in the original source, with strong negatives explicitly supported by experimental evidence and weak negatives originating from randomly selected negative sets or lacking confirmation of the absence of toxicity annotations.

The final release comprises 71,857 unique peptide sequences across eight toxicity-related endpoints, organized into 574,856 sequence-endpoint combinations together with standardized metadata and provenance records. Of 1,372,190 non-deduplicated source-level records initially collected, 898,637 were retained after source-specific preprocessing, representing a reduction of approximately 34.5 percent. Across the resource, 31,153 states were positive, 91,506 negative, 15,717 ambiguous, and 5,044 unlabeled. The Toxic root category contained 57,080 available peptide-endpoint states, including 13,824 positives, while the Cytotoxic class held 25,454 available states with 7,122 positive. Hemolytic had the highest proportion of ambiguous states at 19.05 percent and was the only endpoint with unlabeled states, representing 14.03 percent of its available annotations. Notably, 7,331 of the 9,421 peptides with at least one positive annotation across the seven non-root endpoints were assigned to multiple toxicity categories, corresponding to 77.82 percent of positive peptides across these endpoints.

MAOMAO is organized into complementary functional layers that separate harmonized biological data, computational representations, benchmark resources, and supporting documentation while preserving interoperability. The Core Layer contains harmonized peptide sequences, toxicity annotations, ontology mappings, evidence classifications, provenance information, and metadata, along with 169 quantitative HC50, LC50, LD50, and MHC measurement records. The Descriptor Layer provides 41 features capturing sequence composition and physicochemical properties, generated using the Roxy package. The Embedding Layer includes 10 protein-language-model embedding matrices and a one-hot representation, all covering the complete set of 71,857 sequences, produced with the Sylphy package. The Benchmark Layer contains 300 endpoint-seed-strategy configurations generated through random and stratified five-fold partitioning of five eligible endpoints using 30 random seeds, corresponding to 1,500 data partitions and, across 11 numerical representations, 3,300 representation-specific configurations and 16,500 fold instances. The Documentation Layer includes ontology definitions, metadata schemas, resource descriptions, versioning records, and licensing information, with files distributed in standard formats including CSV, JSON, and YAML.

To facilitate direct reuse, MAOMAO provides five executable notebooks that operate on the released files rather than reconstructing the resource from original sources. These cover resource access, file validation, navigation of the Core Layer, and linkage of annotations, numerical representations, and benchmark partitions through stable sequence identifiers. Two additional notebooks illustrate downstream machine-learning reuse, including a neurotoxicity classification example that joins a released train-validation-test partition with an existing protein language model representation and evaluates logistic regression, a linear support vector machine, and random forest models using fixed settings, and a quantitative regression example showing how numerical toxicity measurements can be combined with a released sequence representation. The complete, versioned MAOMAO resource is archived on Zenodo, and code, workflows, notebooks, and environment specifications are available on GitHub under the MIT License. The researchers acknowledge that endpoint coverage depends on the availability of retrievable sequence-level evidence across curated sources, resulting in class imbalance and limited representation of some endpoints, particularly Cytolysis, Embryotoxic, and Ichthyotoxic, and that the predefined benchmark partitions rely on random and stratified splitting without homology- or similarity-based control. Future releases may address these limitations through the incorporation of newly available toxicity datasets, expanded endpoint and numerical-property coverage, metadata enrichment, updated numerical representations, and similarity-aware benchmark partitions.

Subject of Research: An ontology-guided FAIR resource for harmonized peptide toxicity data

Article Title: MAOMAO: an ontology-guided FAIR resource for harmonized peptide toxicity data

Article References: Soto-Garcia, N., García-Vinuesa, J., Uribe-Paredes, R., Murgas, L., Oróstica, K., González-Puelma, J., Navarrete, M. A., Cadet, F., & Medina-Ortiz, D. (2026). MAOMAO: an ontology-guided FAIR resource for harmonized peptide toxicity data. Frontiers in Pharmacology, 17, Article 1952924. https://doi.org/10.3389/fphar.2026.1952924

Image Credits: AI Generated

DOI: 10.3389/fphar.2026.1952924

Keywords: MAOMAO, peptide toxicity, FAIR principles, ontology, predictive toxicology, machine learning, therapeutic peptides, data harmonization, protein language models, bioinformatics, ontology-guided, FAIR

The challenge of experimentally characterizing transporters is formidable. Membrane proteins are notoriously unstable once removed from their native lipid environment, which complicates purification, crystallization, and nearly every classical biochemical assay used to establish what a protein binds. As a result, curated databases of experimentally verified transporter–substrate relationships cover only a small fraction of known transporters, and many solute carrier proteins—the largest transporter family in humans—remain so-called orphans with no assigned substrate. Computational tools could in principle fill this gap, but existing methods have tended to focus on narrow classes of substrates or have struggled to generalize to transporters that are evolutionarily distant from those in their training data.

TraSPIN, developed by Tong Pan, Fuyi Li, Jiangning Song, and colleagues at institutions including the South Australian immunoGENomics Cancer Institute, Monash University, and the University of Melbourne, takes a fundamentally different approach. The framework is built on geometric deep learning, a branch of machine learning designed to operate on graph-structured data, and it jointly models both sides of the recognition problem: the transporter protein and the small-molecule substrate. By representing transporters and substrates in a shared learning framework, the model can learn the relational patterns that connect a protein’s sequence and structure to the chemical features of the molecules it accepts.

A central innovation is the use of protein language models as the foundation for transporter representations. These models, analogous to the large language models behind modern chatbots, are trained on enormous databases of protein sequences and learn to encode the statistical grammar of amino acid chains. In doing so, they capture subtle information about which positions in a sequence are functionally constrained, which residues are likely to form binding pockets, and how mutations are tolerated. TraSPIN leverages these learned embeddings to build rich, transferable representations of transporters, which are then combined with graph-based learning over protein–substrate interaction networks. The graph structure allows the model to propagate information across related transporters and related substrates, so that knowledge about one interaction informs predictions about others.

This architecture matters because specificity prediction is, at its core, a link-prediction problem on a sparse and biased network. The known interaction map is dense only for a handful of well-studied transporter families and substrate classes, such as sugar transporters or neurotransmitter carriers. A model that simply memorizes these dense regions performs poorly when asked about a transporter from an unexplored family. TraSPIN’s language-model-guided design is explicitly aimed at this generalization problem. In benchmark comparisons reported in the paper, the framework outperformed existing computational methods, and critically, it retained predictive accuracy on substrates it had never seen during training and on transporters whose sequences are dissimilar to anything in the training set—precisely the scenarios where earlier tools broke down.

Beyond raw prediction accuracy, the authors show that TraSPIN is interpretable in biologically meaningful ways. The model’s internal representations can be interrogated to reveal structural determinants of transporter–substrate recognition and sequence motifs associated with binding particular classes of molecules. In other words, the network does not behave as an inscrutable black box; it appears to have learned features that map onto real biochemical principles of molecular recognition. This kind of mechanistic insight is valuable in its own right, because it can guide experimentalists toward specific residues or structural regions to mutate and test, turning a purely computational prediction into a testable hypothesis.

The practical demonstrations in the study highlight two applications with immediate relevance to biology and medicine. The first is cross-species inference: using knowledge accumulated about transporters in well-characterized organisms to predict substrate specificities in species where little experimental data exists. This matters for microbiology, parasitology, and biotechnology, where transporter function in non-model organisms often underpins nutrient uptake, drug resistance, and metabolic engineering. The second application is de-orphanization—the assignment of substrates to uncharacterized solute carrier proteins. Human solute carriers regulate drug absorption, metabolism, and disposition, and orphan carriers are increasingly recognized as potential therapeutic targets and as mediators of unexplained drug–drug interactions. A reliable computational tool that can propose substrates for these orphans could dramatically accelerate functional annotation efforts that would otherwise require years of laboratory work.

The broader context is a rapidly growing intersection between artificial intelligence and structural biology. Protein language models have already transformed fields from structure prediction to protein design, and geometric deep learning has proven adept at tasks where the geometry of molecules—whether the three-dimensional fold of a protein or the two-dimensional scaffold of a small molecule—carries the essential signal. TraSPIN sits squarely at this confluence, treating transporter specificity not as an isolated classification task but as a relational inference problem across a heterogeneous graph of proteins and chemicals. The authors’ results suggest that this joint modeling strategy, rather than modeling proteins and substrates separately, is key to the framework’s generalization ability.

There are, of course, the usual caveats that accompany any machine learning advance in biology. Predictions of transporter–substrate interactions are hypotheses, not measurements, and the sparsity of ground-truth data means that performance estimates on rare substrate classes carry uncertainty. The authors acknowledge that experimental validation remains essential, and the model’s greatest value may lie in prioritizing which of thousands of untested transporter–substrate pairs should be examined first in the laboratory. Used this way—as an intelligent filter rather than an oracle—computational frameworks like TraSPIN can multiply the productivity of experimental programs focused on membrane proteins, where each assay is technically demanding and expensive.

Nevertheless, the arrival of TraSPIN marks a significant step toward solving one of the most persistent annotation problems in molecular biology. The genomes of humans and of every model organism encode hundreds of transporters whose substrates have never been determined, and each uncharacterized carrier is a potential link in a metabolic pathway, a drug target, or a determinant of drug response. By combining the sequence-reading power of protein language models with the relational reasoning of graph learning, TraSPIN offers a scalable route from sequence to function for this protein universe. If its predictions continue to hold up under experimental scrutiny, the era in which transporter specificity is inferred by computer before it is confirmed at the bench may arrive sooner than anyone expected.

Subject of Research: Computational prediction of membrane transporter substrate specificity using protein language models and geometric deep learning

Article Title: TraSPIN: inferring transporter specificity through language model-guided graph learning

Article References: Pan, T., Liu, J., Xu, J., Han, R., Ran, Z., Guo, X., Hao, Y., Bi, Y., Li, F., & Song, J. (2026). TraSPIN: inferring transporter specificity through language model-guided graph learning. Genome Biology. https://doi.org/10.1186/s13059-026-04309-2

Image Credits: AI Generated

DOI: 10.1186/s13059-026-04309-2

Keywords: membrane transporters, substrate specificity, TraSPIN, protein language models, geometric deep learning, graph learning, solute carriers, de-orphanization, protein-substrate interactions, functional annotation, Genome Biology, machine learning

The problem the authors set out to solve is one that anyone who has worked in peptide informatics will recognize immediately. Over the past two decades, the community has built curated databases for cell-penetrating peptides, anticancer peptides, antimicrobial peptides, neuropeptides, toxin peptides and many other functional classes. Each resource encodes genuine expert knowledge, but the knowledge is locked in silos. A researcher asking a simple question — does this sequence cross cell membranes, and is it toxic to cancer cells, and does it serve any other function? — would need to consult multiple databases, reconcile conflicting labels, and contend with inconsistent definitions of what counts as a positive example. The team behind Pep2Net, led by researchers at Hunan University with collaborators at Central South University, the University of Electronic Science and Technology of China and Shenzhen University of Advanced Technology, argues that this fragmentation has become one of the main bottlenecks for computational peptide design.

Their first answer is Pep2Net-DB, an ontology-inspired knowledge base that unifies the 52 peptide resources into a single, coherent structure. The word ontology-inspired is doing real work here: rather than simply dumping datasets into one table, the developers organized peptide types and their attributes according to structured relationships, so that a sequence annotated in one database as a cell-penetrating peptide and in another as an anticancer agent can be understood in a common conceptual framework. Label standardization, one of the keywords the authors attach to the paper, is central to this effort. Without shared definitions, a machine-learning model trained on merged data would learn noise rather than biology; with standardized labels, the merged corpus becomes a training resource far larger and more diverse than any single source database.

The second component, Pep2Net-XAtt, is the predictive engine. It is a multi-task model, meaning that instead of training a separate classifier for each peptide property, one network learns to predict several properties simultaneously. Multi-task learning has a well-established rationale in machine learning: tasks that share underlying structure can regularize one another, and a single model exposed to many labels can pick up generalizable features that a single-task model would miss. For peptides this is particularly apt, because the same physicochemical principles — charge, hydrophobicity, amphipathicity, conformational propensity — govern multiple functional behaviors at once.

What distinguishes Pep2Net-XAtt technically is how it fuses two very different kinds of information. On one side are embeddings from a protein language model, a neural network trained on massive corpora of natural protein sequences that learns to represent amino acid context in dense numerical vectors. These embeddings capture subtle evolutionary and structural patterns that no hand-crafted feature list can encode. On the other side are physicochemical descriptors, the classic quantitative measurements — molecular weight, isoelectric point, hydrophobic moment, charge distribution — that peptide chemists have relied on for decades. Previous models typically treated these two representations separately or simply concatenated them, which the authors argue fails to capture the interactions between semantic sequence context and chemical properties.

Pep2Net-XAtt solves the fusion problem with cross-attention and low-rank bilinear pooling. Cross-attention, the mechanism made famous by transformer architectures, allows one representation to selectively focus on relevant parts of another: the physicochemical descriptor vector can attend to specific positions in the language-model embedding, and vice versa, so that the model learns which chemical features matter in which sequence contexts. Low-rank bilinear pooling then compresses the pairwise interactions between the two representations into a tractable form, multiplying the expressiveness of simple concatenation without the combinatorial explosion that full bilinear interaction would require. The combination lets the network model multiplicative, context-dependent relationships — for instance, how a positive charge at one end of a peptide changes its meaning depending on the hydrophobic pattern along the rest of the chain.

According to the study, this architecture achieves state-of-the-art performance on three prediction tasks: identifying cell-penetrating peptides, predicting anticancer toxicity, and recognizing multi-functional peptides that carry several biological activities simultaneously. The multi-functional task is arguably the most important, because the most therapeutically interesting peptides are rarely one-trick molecules. A candidate drug that crosses cell membranes, kills tumor cells and modulates an immune pathway is exactly the kind of compound that fragmented single-task databases and single-task models systematically overlook. Supplementary material accompanying the paper reports extensive benchmarking, including comparisons with baseline methods, class-wise performance analyses, sequence-length and ensemble robustness checks, and interpretability analyses probing what the model actually learned.

The practical delivery vehicle is Pep2Net Web, a browser-based platform that lets researchers submit sequences and receive multi-task predictions without installing anything. For experimental groups without computational infrastructure, this lowers the barrier substantially: a medicinal chemist screening a panel of candidate peptides can upload them and get functional annotations back in a single pass. The authors position the platform as a tool for large-scale peptide characterization as well as targeted design, and the open-access publication under a Creative Commons license means the underlying data and methods are freely inspectable.

Looking at the bigger picture, Pep2Net arrives at a moment when peptide therapeutics are enjoying a genuine renaissance. GLP-1 receptor agonists have reshaped the treatment of diabetes and obesity, antimicrobial peptides are being pursued as answers to antibiotic resistance, and cell-penetrating peptides are explored as delivery vehicles for cargo that cannot cross membranes on its own. The bottleneck in all these areas is the same: synthesizing and testing peptides is expensive, and the search space of possible sequences is astronomically large. Reliable computational filters that rank candidates before they ever reach the bench can save enormous time and cost, but only if those filters are trained on comprehensive, well-labeled data and can evaluate the full functional profile of a molecule rather than one property at a time.

There are, of course, the usual caveats that accompany any machine-learning advance in biology. Benchmark performance does not guarantee performance on genuinely novel sequence families, and experimental validation remains the final arbiter of any predicted function. The exploratory graph-based structural modeling noted in the supplementary information suggests the authors themselves see structural representation as a frontier for future work. Still, the conceptual contribution is clear and considerable. By unifying 52 scattered resources into one ontology-grounded knowledge base and pairing it with a fusion architecture that finally lets semantic sequence understanding and physicochemical chemistry talk to each other inside a single multi-task model, Pep2Net offers the peptide community something it has lacked: a shared foundation. If the field adopts it, the era in which every peptide property required its own database, its own model and its own paper may finally be drawing to a close.

Subject of Research: An integrated peptide knowledge base and cross-attention deep-learning model for multi-task prediction of peptide functions

Article Title: Pep2Net: an integrated peptide knowledge base and cross-attention model for multi-task prediction of peptides

Article References: Zheng, B., Yuan, J., Wang, Y., Cao, D., Zou, Q., Chen, H., & Fu, X. (2026). Pep2Net: an integrated peptide knowledge base and cross-attention model for multi-task prediction of peptides. Genome Biology. https://doi.org/10.1186/s13059-026-04270-0

Image Credits: AI Generated

DOI: 10.1186/s13059-026-04270-0

Keywords: peptides, Pep2Net, knowledge base, cross-attention, multi-task learning, protein language models, low-rank bilinear pooling, cell-penetrating peptides, anticancer peptides, multi-functional peptides, database integration, peptide design

The review was conducted using a PRISMA-guided systematic literature review methodology, following the Kitchenham guidelines for reproducible evidence synthesis. The authors searched three academic databases, Web of Science, PubMed, and IEEE Xplore, on 2 March 2026, using structured Boolean queries built around large language model terms, protein language model terms such as transformer, BERT, and ESM, and protein structure prediction concepts. The initial search identified 284 records; after deduplication, title and abstract screening, and full-text evaluation against predefined inclusion criteria, the final corpus settled at 38 studies spanning 2020 to 2025. Two reviewers conducted screening with adjudication by a third author, and the complete data extraction sheet has been released as supplementary material to allow independent verification of the entire process.

At the heart of the review lies a technical distinction that is often blurred in popular coverage. Protein language models, or PLMs, are transformer-based models pretrained on protein sequence or structure-augmented sequence data, exemplified by systems such as ESM, ProtBERT, and ProtTrans. General-purpose large language models, by contrast, are trained primarily on natural-language corpora and appear in this literature only when adapted to biological tasks such as literature mining or annotation. The review’s corpus is dominated by dedicated PLMs, which account for 19 of the 38 studies, or 50 percent, followed by general-purpose transformer adaptations at 18.4 percent, with the remainder comprising review and perspective articles. This concentration signals that domain-specific sequence representation learning has become the field’s dominant paradigm.

The technical machinery behind these models is elegant. Transformers employ self-attention mechanisms that capture long-range dependencies within sequential data, and when amino acid sequences are treated as tokens, the models learn contextual relationships between residues directly from massive protein sequence databases. The resulting embeddings encode latent biochemical, evolutionary, and structural information, capturing evolutionary conservation patterns and functional motifs without explicit supervision. Yet the review is careful to note a crucial caveat: sequence-only transformers do not explicitly encode the physicochemical and geometric constraints that govern protein folding, and high-dimensional embedding similarity should not be overinterpreted as evidence of biological realism, particularly for dynamic or poorly characterised protein systems.

To evaluate the field systematically, the authors propose a Biological-Readiness Taxonomy organised around six complementary dimensions: the source of biological knowledge, the representation mechanism, structural reasoning capacity, the biological prediction tasks addressed, validation maturity, and interpretability. Rather than classifying studies by model family, the taxonomy asks whether a model can support biologically meaningful structural inference. The framework is operationalised as an explicit, reproducible scoring rubric, with each dimension scored on an ordinal scale and a composite biological readiness score reported for every study in the corpus. The authors draw an analogy to Technology Readiness Level frameworks used in applied science, and to arguments that AlphaFold predictions constitute testable hypotheses rather than validated structures, but their instrument is domain-specific and multi-dimensional rather than generic.

Applying this framework reveals a clear evolutionary trajectory. The literature has shifted decisively from sequence-only representation learning toward hybrid and multimodal frameworks that integrate evolutionary information from multiple sequence alignments, structural priors, graph-based reasoning, and other biological signals. The review’s quantitative comparison of nine representative foundation-scale models, including AlphaFold2, RoseTTAFold, ESMFold, OmegaFold, AlphaFold-Multimer, AlphaFold3, RoseTTAFold All-Atom, Chai-1, and Boltz-1, traces a progression from single-chain tertiary structure prediction toward comprehensive biomolecular interaction modelling encompassing protein complexes, nucleic acids, and ligands. Earlier methods predicted the shape of individual proteins; the newest systems attempt to model the entire molecular choreography of the cell.

But the validation maturity analysis delivers the review’s most striking numbers. Among the nine representative models, benchmark evaluation remains the predominant validation strategy, accounting for 88.9 percent, or eight of nine models. Only one, AlphaFold3, extended evaluation beyond conventional benchmarking by explicitly comparing predictions against experimentally determined structures. None of the nine reported independent external dataset validation or prospective wet-laboratory validation. The authors are careful to frame this as an observation within an illustrative subset rather than a statistically representative estimate for the entire field, but the pattern is consistent with the corpus-wide taxonomy scores: research priorities remain heavily weighted toward benchmark accuracy, embedding quality, and computational performance, while interpretability, mechanistic biological reasoning, and experimental verification receive comparatively little attention.

The review identifies several systemic challenges that recur across many independent studies. Benchmark dependence is prominent: standardised datasets such as CASP and the Protein Data Bank enable reproducible comparison but draw on well-characterised protein families with extensive homologous sequence information, potentially yielding overly optimistic assessments of robustness on structurally heterogeneous proteins, intrinsically disordered regions, and dynamic conformational states. Transformer architectures also carry quadratic computational complexity with respect to sequence length, limiting scalability to large proteins and proteome-wide applications. Interpretability remains perhaps the most persistent weakness: most models achieve high predictive performance while the reasoning behind their predictions stays opaque, and the few explainability efforts that exist typically offer model-level information such as attention weights rather than biologically valid, experimentally testable explanations.

These limitations matter far beyond academic benchmarking. Accurate protein structure prediction underpins drug discovery, therapeutic protein engineering, biomarker identification, disease diagnosis, and personalised medicine, and the review’s clinical discussion makes clear that the documented gaps in biological interpretability, external validation, and experimental verification would need to close before language-model-based structural predictions could be routinely trusted in therapeutic contexts. The authors also acknowledge the limitations of their own work with unusual candour: the search was restricted to English-language studies in three databases, taxonomy scoring was performed by a single author without inter-rater reliability statistics, and the taxonomy itself remains an unvalidated instrument intended as a starting point for discussion rather than a settled standard.

The review’s ultimate message is a call to rebalance the field’s priorities. Future progress, the authors argue, should be evaluated not only by benchmark accuracy but also by biological validation, structural reasoning capability, interpretability, and translational applicability. They point toward biologically grounded modelling frameworks that integrate structural constraints, evolutionary information, physicochemical knowledge, and multimodal evidence, alongside continued advances in computational efficiency, lightweight foundation models, and explainable AI. The Biological-Readiness Taxonomy offers a practical instrument for tracking that progress, giving researchers and funders a structured way to distinguish models that merely top the leaderboards from those genuinely ready to accelerate discovery in the laboratory. For a field that has reshaped structural biology in half a decade, the next frontier may not be predicting structures better, but proving that the predictions can be believed.

Subject of Research: Systematic review of protein language models for protein structure prediction and a proposed biological readiness taxonomy

Article Title: A systematic review of protein language models for protein structure prediction and a proposed biological readiness taxonomy

Article References: Al-Quraishi, T., Zaboon, W. A., Abawajy, J., & Nowrozy, R. (2026). A systematic review of protein language models for protein structure prediction and a proposed biological readiness taxonomy. Discover Artificial Intelligence, 6(1), Article 1338. https://doi.org/10.1007/s44163-026-02333-6

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02333-6

Keywords: protein language models, protein structure prediction, AlphaFold, transformers, biological validation, interpretability, systematic review, benchmarking, multimodal learning, drug discovery, computational biology, taxonomy

The challenge that motivated the work is deceptively simple to state. Effector proteins are extraordinarily diverse. They often share little recognizable sequence similarity with one another, even when they perform related functions, because bacteria have evolved them repeatedly from different starting points. Traditional bioinformatic tools that scan for conserved sequence motifs therefore miss many genuine effectors while occasionally flagging innocent proteins. In recent years, pretrained protein language models, which learn statistical representations of amino acid sequences from vast databases, have substantially improved effector prediction by capturing contextual patterns that simpler methods overlook. Yet even these powerful models share a fundamental blind spot: they see proteins as one-dimensional strings of letters, while the biological activity of an effector depends on how its chain folds in three-dimensional space.

Structure, in principle, holds the key. The virulence functions of effectors, such as binding host proteins, mimicking host enzymes, or disrupting cellular compartments, are encoded in conformational motifs: recurring arrangements of residues in space that may appear at widely separated positions along the linear sequence. The problem has always been data. Experimentally resolved structures, obtained through crystallography or cryo-electron microscopy, exist for only a small fraction of bacterial proteins, and effectors are particularly underrepresented because they are difficult to purify and crystallize. Conventional structure-based prediction methods, however elegant, have been starved of the structural information they need to work at scale.

GeoEPred, developed by Shouzhen Song, Hua Shi, Hongfeng Wu, Dachen Liu, Yihang Lin, Nor Ashidi Mat Isa, Quan Zou, Leyi Wei and colleagues, threads this needle with a clever compromise. Instead of waiting for experimental structures, the framework uses ESMFold, a deep learning system that predicts three-dimensional protein structures directly from sequence, to generate structural models for every protein it examines. These predicted structures are then processed by geometric vector perceptrons, neural network components that operate on the local geometry of the protein: the orientations between neighboring residues, the distances separating them, and the topology of the spatial neighborhood each residue occupies. In effect, the model learns what local conformational neighborhoods look like in proteins that function as effectors, capturing spatial signatures that no sequence-only model can perceive.

But structure alone is not enough either, and this is where the multimodal design of GeoEPred becomes distinctive. The framework runs two parallel streams of representation. The first draws on a pretrained protein language model to produce sequence-contextual embeddings, rich numerical descriptions of each amino acid in its linear context. The second builds the geometric representations from the predicted structure. A dedicated feature projection network then refines the fine-grained sequence signals that are most relevant to effector function, filtering the language model’s general-purpose knowledge toward the specific patterns that matter for secretion and virulence.

The hardest part of any multimodal system is making the two modalities talk to each other. A sequence embedding and a structural embedding describe the same protein, but they live in different mathematical spaces with different statistics, and naive concatenation often fails to exploit their complementary strengths. The GeoEPred team addressed this with a cross-modal alignment and feature-tokenized self-attention module. The alignment component uses contrastive learning, a technique that pulls matching representations from the two modalities closer together in the learned space while pushing mismatched ones apart, thereby enforcing consistency between sequence semantics and structural geometry. The tokenized self-attention component then models associations at the level of individual residues, allowing the network to link linear functional motifs, short stretches of sequence with known or learned significance, to the spatial conformational patterns those stretches adopt in the folded protein.

The result is a model that can recognize an effector not merely because its sequence resembles known effectors, but because its residues arrange themselves in space the way effector residues tend to. That distinction matters most at the frontier of discovery, where new effectors share no detectable ancestry with characterized ones. The authors evaluated GeoEPred on multiple benchmark datasets covering the three major secretion systems of Gram-negative bacteria, known as type III, type IV, and type VI secretion systems, each of which delivers effectors in its own characteristic way. Across all three prediction tasks, abbreviated T3SE, T4SE, and T6SE in the field’s shorthand, GeoEPred outperformed existing leading models, demonstrating that the structural signal adds genuine predictive power rather than redundant information.

Perhaps more important than the headline benchmarks is the model’s behavior in remote homolog recognition scenarios, where the test proteins are evolutionarily distant from anything seen during training. This is the regime that determines whether a predictor is useful for real discovery or merely good at memorizing known families. GeoEPred maintained stable performance under these demanding conditions, suggesting that the geometric representations it learns capture functional constraints that persist across deep evolutionary time even as sequences drift beyond recognition. The modular architecture also proved extensible, and the authors report strong generalization ability with substantial application potential for genome-scale effector discovery, the scenario in which a newly sequenced bacterial genome can be scanned end to end for candidate virulence factors.

The practical implications reach well beyond the benchmark leaderboards. Rapid identification of effectors accelerates the elucidation of bacterial pathogenic mechanisms, giving microbiologists a prioritized shortlist of proteins to characterize experimentally instead of a genome full of untested hypotheses. It also informs the development of precise anti-infective strategies: effectors and the machinery that delivers them are attractive drug targets, and knowing the full effector repertoire of a pathogen reveals which host pathways it depends on. In agriculture, where Gram-negative pathogens devastate crops, the same logic applies to engineering resistance. As predicted structures from tools like ESMFold become routine inputs to biological machine learning, frameworks like GeoEPred illustrate a broader shift in computational biology, from reading genomes as text to understanding proteins as physical objects whose shapes, not just their sequences, carry the instructions for what they do.

None of this replaces experimental validation, and the authors are careful to frame GeoEPred as a discovery engine whose predictions guide, rather than conclude, the scientific process. Predicted structures, however accurate on average, carry uncertainties that propagate into any downstream model, and effector biology remains full of surprises that no current dataset fully captures. But the trajectory is clear. By fusing the semantic richness of protein language models with the physical realism of geometric deep learning, and by solving the fusion problem with principled contrastive alignment rather than brute-force concatenation, GeoEPred offers a template for how the next generation of protein function predictors will be built. For the hidden weapons of Gram-negative bacteria, many of which have evaded detection precisely because they look like nothing we have sequenced before, the era of shape-aware prediction may finally be opening the door.

Subject of Research: Multimodal geometric deep learning for predicting Gram-negative bacterial secreted effector proteins

Article Title: GeoEPred: A multimodal structure-aware geometric deep learning framework for Gram-negative bacterial secreted effector prediction with sequence semantics

Article References: Song, S., Shi, H., Wu, H., Liu, D., Lin, Y., Mat Isa, N. A., Zou, Q., & Wei, L. (2026). GeoEPred: A multimodal structure-aware geometric deep learning framework for Gram-negative bacterial secreted effector prediction with sequence semantics. PLOS Computational Biology, 22(9), e1014344. https://doi.org/10.1371/journal.pcbi.1014344

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014344

Keywords: GeoEPred, Gram-negative bacteria, effector proteins, deep learning, protein language models, ESMFold, protein structure prediction, type III secretion system, type IV secretion system, type VI secretion system, virulence, computational biology

Now, a research team led by Chunyan Ao, Shihu Jiao, Xi Su, and Huan Yang has unveiled a deep learning framework called pLM-HP that promises to change that. Writing in the Journal of Advanced Research, the authors describe a system that combines a pre-trained protein language model, ESM2, with a bidirectional long short-term memory network, or BiLSTM, to predict whether a given peptide sequence is a hormone. On an independent test set, the model achieved a balanced accuracy of 95.64 percent, with sensitivity of 94.48 percent and specificity of 96.79 percent, a Matthews correlation coefficient of 0.824, and an area under the ROC curve of 0.991. Those numbers place it well ahead of existing tools, and they hint at a broader shift in how computational biology identifies functional molecules.

The challenge the researchers set out to solve is not new. Traditional identification of peptide hormones relies on experimental techniques such as mass spectrometry, affinity purification, and activity screening, all of which are complex and time-consuming. Computational methods emerged as a complementary approach, but early efforts leaned on sequence similarity searches or motif recognition algorithms, which falter when homologous sequences are scarce. Machine learning brought improvements: the HOPPred model, for example, integrates BLAST similarity, MERCI motifs, and logistic regression, achieving an AUROC of 0.96 and an MCC of 0.80 on an independent test set, while mHPpred adopted a multi-view feature representation and a meta-modeling strategy to build a more robust ensemble. Still, many methods remained tethered to hand-crafted features or predefined motifs, limiting their ability to represent sequences fully, and class imbalance in real data often biased models toward the majority class, dulling their sensitivity to genuine hormone peptides.

Protein language models have upended that paradigm. Models such as ProtBERT, ProtT5, and ESM2 are pre-trained on enormous collections of unlabeled protein sequences using masked language modeling, in which the model learns to predict hidden amino acids from their surrounding context. In doing so, they acquire high-dimensional embeddings that encode evolutionary, structural, and physicochemical properties of proteins. These embeddings have already powered advances in post-translational modification prediction, peptide structure modeling, and functional peptide classification. But until now, their use in peptide hormone prediction had been limited, and no systematic study had evaluated how well they perform on this specific task.

The pLM-HP framework treats a peptide sequence like a sentence and each amino acid like a word. ESM2, built on the Transformer encoder architecture and pre-trained on UniRef50 sequences, processes the sequence with a multi-head self-attention mechanism that captures long-range dependencies between residues, combined with feed-forward networks for nonlinear feature transformation. Each input is augmented with special beginning-of-sequence and end-of-sequence tokens, and token and positional embeddings are combined to capture residue order and context. Crucially, the researchers used ESM2 solely as a frozen feature extractor, without fine-tuning, and passed the residue-level embeddings directly to the downstream classifier.

One of the study’s most striking findings concerns model size. The team benchmarked four ESM2 variants, ranging from the 8-million-parameter esm2_t6_8M to the 650-million-parameter esm2_t33_650M, with output dimensions of 320, 480, 640, and 1280 respectively. All achieved balanced accuracy of roughly 95 percent in both five-fold cross-validation and independent testing, but the mid-sized esm2_t12_35M model, with 480-dimensional outputs, performed best, reaching a balanced accuracy of 95.61 percent and an MCC of 0.802 in cross-validation and 95.64 percent and 0.824 on the independent test set. The largest model, despite its far greater capacity, actually performed slightly worse, with a balanced accuracy of 95.15 percent and an MCC of 0.804, suggesting that for short peptide sequences of 11 to 41 residues, bigger is not necessarily better and may even introduce redundant features or mild overfitting.

The choice of classifier mattered just as much. When the researchers compared deep learning architectures and traditional machine learning methods on top of the ESM2 features, the BiLSTM emerged as the clear winner. By running two recurrent networks in forward and backward directions and concatenating their hidden states, the BiLSTM exploits both past and future context around each residue, capturing contextual dependencies that other architectures miss. It outperformed convolutional neural networks, multilayer perceptrons, support vector machines, logistic regression, random forests, XGBoost, and LightGBM across the board. Tree-based models fared worst: random forest achieved only about 73 percent balanced accuracy on the independent test set, with a strong bias toward predicting negative samples, while XGBoost and LightGBM reached high specificity near 98 percent but sensitivity below 81 percent, limiting their ability to spot true hormone peptides.

Class imbalance posed another formidable obstacle, since hormone peptides are vastly outnumbered by non-hormone sequences in real datasets. The team built their training data from 5,729 plant- and animal-derived peptide hormone sequences curated in the Hmrbase2 database, reduced to 1,174 non-redundant positives using CD-HIT at a 60 percent similarity threshold, paired with 11,740 non-hormone peptides drawn from PeptideAtlas. Rather than relying on resampling tricks, they incorporated class weights into a weighted binary cross-entropy loss, boosting the contribution of the rare hormone class. When they stress-tested the model on training sets with positive-to-negative ratios ranging from 1:1 to 1:10, performance remained remarkably stable, with balanced accuracy consistently above 94 percent, MCC values between 0.741 and 0.824, AUC values from 0.983 to 0.992, and AUPRC values from 0.841 to 0.915. Notably, the gap between sensitivity and specificity stayed small, at just 2.31 percentage points on the final test set, indicating a balanced classifier rather than one that simply favors the majority class.

The advantages over traditional approaches were dramatic. Handcrafted sequence descriptors such as pseudo amino acid composition, composition of k-spaced amino acid pairs, and quasi-sequence-order descriptors, when paired with classifiers like XGBoost, showed high specificity but weak sensitivity, and classical imbalance-handling strategies such as SMOTE and ADASYN often improved sensitivity at the cost of specificity, producing unstable results. In head-to-head comparisons on the same independent test set, pLM-HP outperformed HOPPred on every metric, improving balanced accuracy by 24.33 percentage points, sensitivity by 35.00 points, and specificity by 13.64 points, while raising the MCC by 0.522 and the AUC by 0.264. Against mHPpred, it improved balanced accuracy by 9.95 points, specificity by 18.51 points, MCC by 0.368, and AUC by 0.055. Visualization analyses using t-SNE and UMAP reinforced the story: while handcrafted features left positive and negative samples substantially overlapping, and raw ESM2 embeddings only partly separated them, the representations refined by the BiLSTM formed compact, well-separated clusters.

The implications reach well beyond one prediction task. As high-throughput omics and peptide-based drug development accelerate, efficiently and reliably identifying hormone-functional peptides from vast sequence spaces has become a key bridge between basic research and drug discovery. Peptide hormones, capable of probing and modulating protein-protein interactions, are considered ideal scaffolds for drug design, and a tool that can screen candidate sequences with near-99 percent AUC could dramatically narrow the experimental search space. The authors are candid about limitations: the current framework relies on linear sequence information and does not explicitly integrate structural context that shapes peptide conformation and signaling. Future work, they suggest, could incorporate structure predictions from AlphaFold2, apply geometric deep learning with graph neural networks, and explore focal loss, contrastive learning, and fine-tuning strategies to improve robustness under extreme imbalance and cross-dataset transfer. Even so, pLM-HP offers a compelling demonstration that when protein language models meet the right sequence-aware classifier, the hidden language of hormonal peptides becomes far easier to read.

Subject of Research: Computational prediction of peptide hormones using pre-trained protein language model embeddings and deep learning

Article Title: pLM-HP: Peptide hormone prediction using pre-trained protein language model representations

Article References: Ao, C., Jiao, S., Su, X., & Yang, H. (2026). pLM-HP: Peptide hormone prediction using pre-trained protein language model representations. Journal of Advanced Research. https://doi.org/10.1016/j.jare.2026.10.002

Image Credits: AI Generated

DOI: 10.1016/j.jare.2026.10.002

Keywords: peptide hormones, protein language models, ESM2, BiLSTM, deep learning, drug discovery, machine learning, bioinformatics, class imbalance, sequence analysis, peptide drugs, computational biology

Developed by Polina Polunina and Wolfgang Maier of the University of Freiburg together with Alan Rubin of the Walter and Eliza Hall Institute of Medical Research, VEFill is built on LightGBM, a gradient boosting framework well suited to structured, high-dimensional tabular data. Rather than relying on a single source of information, the model integrates a rich set of biologically informed features: evolutionary conservation scores from the EVE generative model, amino acid substitution matrices such as BLOSUM62 and PAM250, physicochemical descriptors of the wild-type and variant residues, and contextual sequence embeddings from ESM-1v, a transformer-based protein language model trained on millions of natural sequences. The authors trained and optimized the model using Bayesian hyperparameter search and stored all underlying data in a structured PostgreSQL database to ensure reproducibility.

The training ground for VEFill was the Human Domainome 1 dataset, a standardized collection of stability measurements generated with an abundance-based protein fragment complementation assay across 522 human protein domains, using site-saturation mutagenesis to introduce every possible amino acid substitution. From this resource, the team assembled a feature-complete subset of 140 domains comprising 136,854 mutations for the full model, and used the broader set of 521 domains, totaling 562,208 mutations, to train reduced-feature versions that depend only on ESM-1v embeddings and positional mean scores.

The performance figures are striking. In a leave-protein-out evaluation, where the model must predict variant effects for protein domains entirely absent from training, the best configuration achieved a coefficient of determination of 0.64 and a Pearson correlation of 0.80. Feature ablation experiments revealed a clear hierarchy of information: models relying solely on evolutionary scores, substitution matrices, or biochemical descriptors performed poorly, with R-squared values below 0.15. Adding ESM-1v embeddings lifted performance substantially, and combining the embeddings with the mean experimental score per position proved even more powerful, indicating that learned sequence representations and data-driven positional priors capture complementary aspects of mutational tolerance.

Perhaps the most practically important finding concerns data scarcity. In controlled subsampling experiments on 28 high-quality datasets, VEFill consistently outperformed a battery of existing imputation methods, including Envision, FactorizeDMS, AALasso, and nearest-neighbor approaches, once at least 20 percent of variants had been experimentally measured. Below that threshold, predictions became markedly harder, and the authors are candid that true zero-shot prediction without any positional context remains challenging, particularly for functionally complex proteins. This matters because a survey of 841 public score sets from the MaveDB repository showed that fully complete datasets are vanishingly rare, with only 44 out of 841 achieving complete coverage of all possible substitutions. The vast majority of real-world experiments therefore fall squarely within the regime where VEFill delivers its greatest benefit.

The team also probed how few measurements are actually needed per position. When per-protein models were trained with a restricted number of substitutions per site, accuracy rose rapidly and began to plateau at roughly four mutations per position. Intriguingly, models trained exclusively on five carefully chosen amino acids, histidine, glutamic acid, asparagine, isoleucine, and glycine, each representing a distinct physicochemical class, matched the performance of models trained on randomly selected substitutions at equivalent coverage. This suggests that sparse, information-efficient mutational libraries could be designed deliberately, prioritizing a small representative set of substitutions per site without sacrificing predictive power.

Noise ceiling analyses added an important dose of realism. Because experimental variability imposes a fundamental upper bound on achievable accuracy, the researchers simulated replicate experiments using reported measurement uncertainties and found that estimated ceilings ranged from 0.68 to 0.99 across domains. VEFill’s per-protein correlations, between 0.52 and 0.91 under an 80/20 split, sat consistently below but generally close to these limits, indicating that much of the remaining discrepancy between predicted and observed scores reflects measurement noise rather than model failure. Error profiles were lowest for near-neutral variants, where data are densest, and predictions at the extreme deleterious and high-activity tails should be read as conservative approximations rather than precise estimates.

Generalization beyond the training data showed both promise and limits. When the cross-protein model was applied to eight full-length proteins from independent MaveDB datasets, including Parkin, PTEN, aspartoacylase, calmodulin, TPK1, and TP53, it performed substantially better on stability-based assays than on activity-based ones, a mismatch the authors attribute to differences between training and test assay modalities and the complexity of cellular phenotypes. A lightweight two-feature version using only ESM-1v embeddings and positional mean scores performed comparably to the full model in most settings, offering a practical alternative when evolutionary scores or other annotations are unavailable. Analysis of protein family composition revealed that a few Pfam families dominated the training data, but retraining on a Pfam-unique subset increased error only modestly, suggesting the findings are robust rather than inflated by family-level redundancy.

Fine-grained evaluations exposed the model’s blind spots in biologically meaningful ways. Proline substitutions consistently produced elevated errors, likely because their unusual conformational constraints disrupt secondary structure in ways that sequence-based features do not fully capture. In the TRIM44 zinc finger domain, histidine-to-cysteine mutations were poorly predicted when entire positions were held out but accurately recovered when single variants were withheld, hinting at context-dependent chemistry, such as zinc coordination, that only fine-grained positional information can resolve. The authors suggest that future versions could incorporate explicit structural features, cross-species transfer learning, or systematic assay harmonization to close these gaps.

Overall, VEFill arrives at a moment when the variant effect field is scaling rapidly, with MaveDB now listing more than 2,600 public datasets and over 1,100 for human proteins. By providing an interpretable, scalable tool that turns partially complete mutational maps into denser, more usable resources, the study lowers the experimental barrier for variant prioritization, protein engineering, and the clinical interpretation of missense mutations. The code and trained models have been released openly, inviting the community to apply sparse-library design strategies and, ultimately, to build the comprehensive Atlas of Variant Effects that the field has long envisioned.

Subject of Research: Machine learning-based imputation of missing scores in deep mutational scanning datasets across human protein domains

Article Title: VEFill: accurate and generalizable deep mutational scanning score imputation across protein domains

Article References: Polunina, P. V., Maier, W., & Rubin, A. F. (2026). VEFill: accurate and generalizable deep mutational scanning score imputation across protein domains. Molecular Systems Biology, 22(6), 979-1002. https://doi.org/10.1038/s44320-026-00203-y

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00203-y

Keywords: deep mutational scanning, variant effect prediction, machine learning, gradient boosting, protein stability, ESM-1v, protein language models, MaveDB, Human Domainome, imputation, missense variants, bioinformatics

The stakes are higher than they might first appear. CRISPR and the Cas proteins that accompany it form the adaptive immune systems of prokaryotes, the vast group of single-celled organisms that dominate life on Earth. These systems store fragments of invading viruses and plasmids in the famous clustered regularly interspaced short palindromic repeats, then use them as molecular wanted posters to recognize and destroy foreign genetic material. Because different classes of Cas proteins perform different jobs within these systems, identifying which Cas proteins are present in a genome offers a route to classifying the whole system. The trouble is that direct experimental identification of CRISPR-Cas systems remains difficult, so computational classification carries much of the load, and its accuracy matters.

The core insight behind PrePssmCas is that no single numerical representation of a protein captures everything a classifier needs to know. The team therefore extracted two complementary families of features from Cas protein sequences. The first came from pre-trained protein language models, the deep neural networks that have transformed computational biology in recent years by learning statistical patterns from millions of sequences. The second came from position-specific scoring matrices, or PSSMs, a much older technique that encodes the evolutionary conservation of each position in a sequence by comparing it against related proteins in a database.

The language model side of the comparison was ambitious. The researchers systematically evaluated five pre-trained models: Prot_BERT, ALBERT, ProtXLNet, ProtT5 and ESM1b. Each of these models reads a protein sequence like a sentence and produces a high-dimensional vector, known as an embedding, for every amino acid in it. These embeddings encode information about structure, function and evolutionary context that the models absorbed during training on vast protein databases. On the PSSM side, the team compared six different variants, methods that compress the raw scoring matrix into fixed-dimensional descriptors that a classifier can consume. The raw PSSM itself has dimensions that scale with protein length, which makes it unwieldy for machine learning, so the compression step is where much of the engineering happens.

After systematic comparison, a clear winner emerged: embeddings from ESM1b, the protein language model developed by Meta’s fundamental AI research group, combined with a PSSM variant called RPM-PSSM gave the most comprehensive representation of Cas proteins. The researchers also incorporated an attention-based aggregation strategy, a mechanism that lets the model learn which parts of a sequence deserve the most weight when building its final representation. Attention has become a cornerstone of modern deep learning, and its use here reflects a broader trend of importing techniques from natural language processing into protein analysis.

Fusing the two feature families produced a large pool of candidate descriptors, and not all of them carried useful signal. To distill the mixture, the team applied random forest-based feature selection, a technique that uses an ensemble of decision trees to score each feature by how much it contributes to accurate classification. The result was a compact 143-dimensional feature vector, comprising 87 features drawn from the pre-trained language model embeddings and 56 features from the RPM-PSSM representation. That reduction, from potentially thousands of dimensions down to 143, is what allows the final classifier to generalize rather than memorize, focusing on the descriptors that genuinely distinguish one type of Cas protein from another.

The performance gains were substantial. On the independent validation set, the selected features achieved 97.98 percent accuracy and an MCC of 0.962. The Matthews correlation coefficient is widely regarded as a more honest metric than raw accuracy because it accounts for all four cells of the confusion matrix and remains informative even when classes are imbalanced, as they typically are in biological datasets. An MCC above 0.95 indicates near-perfect agreement between predictions and ground truth. Crucially, the improvements were measured against CRISPRCasStack, itself a recent and strong baseline, meaning the gains are not merely the result of comparing against weak predecessors.

The benchmarking went further. PrePssmCas was evaluated against seven existing methods: HMMCAS, CASPredict, CRISPRone, CRISPRCasFinder, CRISPRCasTyper, CRISPRloci and CRISPRCasStack. These tools represent a range of strategies, from hidden Markov models that capture profile signatures of protein families to more recent machine learning pipelines. On the independent validation set, PrePssmCas outperformed all of them, a result the authors attribute to the complementary nature of the fused features. Language model embeddings excel at capturing deep semantic patterns learned across the protein universe, while PSSM-derived features anchor the predictions in the specific evolutionary history of each sequence, and the combination appears to cover blind spots that neither approach addresses alone.

The work also contributes to a lively debate in computational biology about whether large pre-trained language models have made classical sequence analysis techniques obsolete. The answer from this study is a qualified no. Although the ESM1b embeddings contributed the larger share of the final feature vector, more than half of the selected dimensions came from the language model, the 56 RPM-PSSM features that survived selection clearly added discriminative power that the embeddings alone did not provide. Evolutionary profiles, computed by aligning a query sequence against a database of homologs with tools like BLAST, encode information about which positions tolerate mutation, and that information is not always fully captured by a language model trained on sequences without explicit alignment data. The fusion framework suggests that the most powerful classifiers of the near future will be hybrids rather than replacements.

The authors are candid about the limits of their achievement. They note that the generalization of PrePssmCas to large-scale datasets and to remotely homologous sequences, proteins whose similarity to known Cas proteins is so distant that standard detection methods struggle, remains to be established. This is a familiar caveat in protein classification, where models trained on curated datasets can falter when confronted with the messy diversity of real genomic data, particularly the rapidly evolving catalog of CRISPR-Cas variants being discovered through metagenomics. Still, the researchers frame their contribution as both a practical tool for classifying CRISPR-Cas systems and a promising feature-fusion framework that could support future Cas protein discovery, potentially helping biologists spot novel defense systems hiding in the growing torrent of microbial genome sequences. The study, led by Ningyi Zhang with corresponding author Qianqian Shi, was conducted without dedicated funding and relied on the High-Performance Computing Center of Northwest A&F University, the publicly available ESM1b model and BLAST software, and the Gold Standard dataset for Cas protein analysis maintained by the research community.

Subject of Research: Machine learning classification of CRISPR-Cas proteins using fused pre-trained language model and PSSM features

Article Title: PrePssmCas: fusing pre-trained features and PSSM features for Cas protein classification

Article References: Zhang, N., Zhao, Y., Luo, C., Peng, Z., & Shi, Q. (2026). PrePssmCas: fusing pre-trained features and PSSM features for Cas protein classification. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06569-y

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06569-y

Keywords: CRISPR-Cas, Cas proteins, protein classification, machine learning, ESM1b, protein language models, PSSM, feature fusion, bioinformatics, BMC Bioinformatics, random forest, deep learning

The problem that Alper Karagöl and Taner Karagöl of Istanbul University set out to solve is familiar to anyone who works on membrane proteins. Human transporters are fragile, unstable and expensive to produce in quantities suitable for structural studies. Bacterial cousins, by contrast, are often far more tractable. The Escherichia coli transporter DtpA, for example, is a stable and well-behaved structural surrogate for PepT1 — but it is still a bacterial protein. The binding pocket inside DtpA differs in crucial ways from the human version, so drug-screening results obtained with the bacterial surrogate do not always translate faithfully to human pharmacology. The obvious fix is to graft the human binding-site residues onto the bacterial scaffold, a strategy sometimes described as humanizing the transporter. In practice, that fix has repeatedly failed, because swapping in non-native side chains disrupts the finely co-evolved packing of the surrounding protein, producing what the authors describe as severe thermodynamic frustration.

The scale of the difficulty is staggering. A transporter’s binding microenvironment involves dozens of residues whose interactions are interdependent: change one, and the energetic consequences ripple through its neighbors. Traditional structure-based engineering confronts a combinatorial explosion — the number of possible compensatory mutations grows exponentially with each humanizing substitution — and laboratory directed evolution, which mutates proteins randomly and screens for improved variants, is slow, expensive and poorly suited to membrane proteins that are hard to express and assay. The Karagöls’ answer was to move the entire evolutionary search into silicon, building what they call a fully in silico co-evolutionary pipeline powered by deep contextual protein language models.

Protein language models, or PLMs, are neural networks trained on vast databases of natural protein sequences. Like large language models that learn the statistical grammar of human text, PLMs learn the grammar of amino acid sequences — which residues tend to appear where, and which combinations are evolutionarily tolerated. The researchers used ESM-2, a 650-million-parameter model developed by Meta researchers and described in Science in 2023, to perform zero-shot mutational profiling. Zero-shot means the model predicts the effect of a mutation it has never explicitly been trained on, simply by asking how likely the mutated sequence is under the statistical patterns it has internalized. The team mapped 19 pharmacologically critical residues from human PepT1 onto the corresponding positions in the DtpA sequence and asked ESM-2 to score each humanizing substitution.

The results were sobering but informative. Fourteen of the primary humanizing mutations received very low likelihood scores — below 10 to the power of minus 5 — indicating that the model considered them essentially incompatible with the rest of the bacterial protein. In evolutionary terms, these substitutions would be deleterious on their own. This is precisely the thermodynamic frustration that has doomed previous humanization attempts: the human residues carry the pharmacological information researchers want, but the bacterial scaffold cannot accommodate them without help. The question became whether that help could be found computationally, without running a single experiment.

To find it, the authors deployed an automated epistatic rescue algorithm. Epistasis refers to the phenomenon in which the effect of one mutation depends on the presence of another — a deleterious change can be buffered, or rescued, by a compensatory mutation elsewhere in the protein. The algorithm swept through the flanking microenvironment of each high-risk substitution, systematically testing nearby positions for secondary mutations that would restore the model’s confidence in the humanized sequence. The heuristic proved computationally efficient and, remarkably, succeeded for all 14 high-risk targets, identifying localized non-native compensatory mutations for each one — all without recourse to traditional molecular dynamics simulations, which are far more expensive.

Two examples illustrate the power of the approach. The substitution Q41R — replacing glutamine with arginine at position 41 — was flagged as highly deleterious, but pairing it with a secondary mutation, Y38D, increased the sequence likelihood assigned to the primary substitution by a factor of 5,215. Even more dramatic was R305F, an arginine-to-phenylalanine change, which was buffered by a companion I304K mutation yielding an 11,614-fold increase in likelihood. The authors are careful to note an important technical caveat: these figures describe changes in sequence likelihood as assigned by the language model, not calculated free energies. The likelihood gains indicate that the model considers the rescued sequences evolutionarily plausible, but they are not direct thermodynamic measurements.

With the rescued variants in hand, the team turned to structural validation. Binding dynamics were assessed through docking calculations, using tools including SwissDock, and through all-atom molecular dynamics simulations of the transporter embedded in a realistic membrane environment, built with the CHARMM-GUI Membrane Builder and run with the GROMACS simulation package using the CHARMM36m force field. These simulations allowed the researchers to examine whether the humanized binding pocket maintained a stable, transport-competent architecture in a lipid bilayer — a critical check, since membrane proteins can misbehave badly when their environment is not modeled faithfully.

The practical payoff could be substantial for drug development. A humanized DtpA surrogate, if it works, would give pharmacologists a stable bacterial scaffold carrying the human PepT1 binding pocket, enabling more faithful screening of peptide-mimetic drugs and inhibitors of the SLC15 transporter family. By computationally rationalizing which humanizing mutations are viable and which compensatory changes make them viable, the framework dramatically reduces the experimental screening space that laboratory researchers would need to explore. The authors also point toward applications beyond drug screening: engineered peptide transporters are relevant to live biotherapeutic products, including probiotic strains of E. coli Nissle 1917 that are being developed as delivery vehicles for therapeutic peptides.

The team is careful about what has and has not been achieved. The humanized surrogate exists only as a computational construct so far, and the authors state explicitly that experimental validation of folding, binding and transport is required before the design can be used in pharmacological screening. All the statistical and computational analyses, including AlphaFold calculations and the code to regenerate them, have been made publicly available on GitHub to ensure reproducibility. Still, the study offers a striking preview of how protein engineering may be done in the years ahead: rather than mutating and screening in the lab, researchers may first ask a language model which mutations evolution would reject, then let an epistatic rescue algorithm find the partners that make the impossible tolerable — and only then head to the bench, with a shortlist of designs that artificial intelligence has already vetted against billions of years of evolutionary constraint.

Subject of Research: Computational humanization of the bacterial peptide transporter DtpA using protein language models and epistatic rescue to create a surrogate of human PepT1 for drug screening

Article Title: In silico directed evolution of humanized peptide transporters via computational epistatic rescue

Article References: Karagöl, A., & Karagöl, T. (2026). In silico directed evolution of humanized peptide transporters via computational epistatic rescue. Molecular Diversity. https://doi.org/10.1007/s11030-026-11710-3

Image Credits: AI Generated

DOI: 10.1007/s11030-026-11710-3

Keywords: PepT1, SLC15A1, DtpA, protein language models, ESM-2, epistasis, directed evolution, membrane transporters, drug bioavailability, molecular dynamics, protein engineering, in silico mutagenesis

The study, led by Jorge F. Beltrán and colleagues, presents a multi-scale benchmark of ESM2 models for binary classification of viral entry proteins. The researchers compared three versions of the model containing approximately 650 million, 3 billion, and 15 billion parameters. These systems were not retrained on the viral-protein dataset. Instead, the investigators used “frozen embeddings”—fixed numerical descriptions generated by ESM2 for each protein sequence—and supplied those representations to conventional machine-learning classifiers. This approach separates the task of learning general protein patterns from the task of distinguishing entry proteins from other viral proteins.

Protein language models are inspired by the way large language models process written text. In a protein sequence, amino acids function like an alphabet, while recurring sequence patterns reflect structural elements, interaction sites, cellular targeting signals, and evolutionary constraints. During training, models such as ESM2 learn to predict missing or masked amino acids from their surrounding context. Although the model is not explicitly told that a sequence belongs to a fusion protein or receptor-binding protein, its internal representation can encode relationships associated with folding, topology, and molecular function. The resulting embedding is a high-dimensional numerical profile that can be analyzed by a separate classifier.

To create the benchmark, the researchers assembled 1,092 reviewed viral proteins from UniProt/Swiss-Prot. The dataset was balanced between viral entry proteins and non-entry proteins, reducing the risk that a classifier could appear accurate simply by favoring the more common class. To limit the influence of closely related sequences, the proteins were clustered with CD-HIT at 40 percent sequence identity. The resulting data were divided into a training set of 873 proteins and an independent test set of 219 proteins. This sequence-redundancy control is important because random splits can produce deceptively high scores when nearly identical proteins occur on both sides of the evaluation.

Five supervised algorithms were tested on the ESM2 embeddings: a support vector machine with a radial basis function kernel, random forest, XGBoost, LightGBM, and a multilayer perceptron. These classifiers differ in how they model decision boundaries. The radial-basis-function support vector machine can separate classes in a transformed feature space, while tree-based methods identify combinations of informative variables through ensembles of decision trees. The multilayer perceptron, by contrast, learns nonlinear transformations through artificial neural-network layers. Comparing several algorithms allowed the researchers to assess whether performance depended primarily on the protein representation or on the choice of downstream classifier.

Performance increased as the ESM2 model grew larger, although the improvement became modest beyond the intermediate scale. The strongest result came from the 15-billion-parameter ESM2 model combined with the radial-basis-function support vector machine. On the independent test set, this configuration achieved 88.6 percent accuracy, a Matthews correlation coefficient of 0.772, and a receiver operating characteristic area under the curve of 0.956. The Matthews correlation coefficient is particularly useful for binary classification because it incorporates true and false positives and negatives into a single measure and remains informative when class distributions are uneven. The ROC-AUC value indicates strong discrimination across a range of classification thresholds.

The 3-billion-parameter model produced nearly the same outcome when paired with the same classifier, reaching 88.1 percent accuracy, an MCC of 0.765, and a ROC-AUC of 0.953. This small difference suggests that most of the useful information for this task was captured before the largest model scale. In practical terms, the 3B model may offer a more attractive balance between accuracy, memory requirements, and computational cost. The findings also illustrate an important feature of protein language models: larger networks can improve biological representation, but their gains may diminish when the dataset is relatively small or when the classification problem is already well represented by an intermediate model.

The ESM2 systems also outperformed traditional sequence-based descriptors. The study evaluated amino-acid composition, dipeptide composition, and pseudo-amino-acid composition, which summarize proteins using frequencies of individual residues, adjacent residue pairs, or combinations of composition and sequence-order information. Such descriptors are computationally efficient and remain useful in many bioinformatics applications, but they discard much of the contextual information present in a full sequence. The best descriptor-based baseline reached an MCC of 0.537 and a ROC-AUC of 0.846 on the independent test set, considerably below the strongest ESM2 configurations. The comparison supports the idea that contextual embeddings can capture signals that simple frequency-based features overlook.

The authors further explored how the method behaved when evaluated across taxonomic boundaries. In a taxid-disjoint analysis using the 650-million-parameter ESM2 model and the support vector machine, the mean MCC was 0.734 and the mean ROC-AUC was 0.929. This experiment was designed to test whether the classifier could retain predictive ability when proteins associated with particular taxonomic identifiers were separated between training and evaluation. The result is encouraging for surveillance applications, where newly encountered viruses may not be represented in the training data. However, the researchers emphasize that the analysis does not establish performance on completely unseen viral families. Taxonomic separation at the identifier level can still leave evolutionary relationships between training and test proteins.

The study’s practical outcome is ViralEntryPred, a web application intended for rapid screening of protein sequences for possible viral entry function. Such a tool could help prioritize candidates for laboratory testing, structural modeling, receptor-binding studies, or vaccine research. It is not a replacement for experimental validation: a computational prediction cannot by itself demonstrate receptor usage, membrane-fusion activity, cellular tropism, or pathogenicity. The results are also bounded by the quality and scope of the curated dataset, the definition of an entry protein, and the possibility that sequence patterns associated with known viral groups may not generalize to deeply divergent organisms. Even with these limitations, the benchmark demonstrates how frozen protein-language-model embeddings can extend functional annotation into regions of viral sequence space where conventional similarity searches provide little guidance. As genomic surveillance expands, intermediate-scale models such as ESM2 3B could become useful components of a broader pipeline for identifying molecular features that shape the earliest and most decisive step of infection: viral entry.

Subject of Research: Viral entry protein prediction using protein language model embeddings and machine learning

Article Title: Protein language models for viral entry protein prediction: a multi-scale ESM2 benchmark

Article References: Beltrán, J. F., Belén, L. H., Lugo, A., et al. “Protein language models for viral entry protein prediction: a multi-scale ESM2 benchmark.” BMC Bioinformatics (2026).

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06622-w

Keywords: Protein language models, ESM2, viral entry proteins, membrane fusion, machine learning, protein function prediction, embeddings