Evolutionary-scale prediction of atomic-level protein structure with a language model

Prediction of protein biophysical traits from limited data: a case study on nanobody thermostability through NanoMelt

In-silico prediction of protein biophysical traits is often hindered by the limited availability of experimental data and their heterogeneity. Training on limited data can lead to overfitting and poor generalizability to sequences distant from those in the training set. Additionally, inadequate use of scarce and disparate data can introduce biases during evaluation, leading to unreliable model performances being reported. Here, we present a comprehensive study exploring various approaches for protein fitness prediction from limited data, leveraging pre-trained embeddings, repeated stratified nested cross-validation, and ensemble learning to ensure an unbiased assessment of the performances. We applied our framework to introduce NanoMelt, a predictor of nanobody thermostability trained with a dataset of 640 measurements of apparent melting temperature, obtained by integrating data from the literature with 129 new measurements from this study. We find that an ensemble model stacking multiple regression using diverse sequence embeddings achieves state-of-the-art accuracy in predicting nanobody thermostability. We further demonstrate NanoMelt's potential to streamline nanobody development by guiding the selection of highly stable nanobodies. We make the curated dataset of nanobody thermostability freely available and NanoMelt accessible as a downloadable software and webserver.

Predicting purification process fit of monoclonal antibodies using machine learning

In early-stage development of therapeutic monoclonal antibodies, assessment of the viability and ease of their purification typically requires extensive experimentation. However, the work required for upstream protein expression and downstream purification development often conflicts with timeline pressures and material constraints, limiting the number of molecules and process conditions that can reasonably be assessed. Recently, high-throughput batch-binding screen data along with improved molecular descriptors have enabled development of robust quantitative structure-property relationship (QSPR) models that predict monoclonal antibody chromatographic binding behavior from the amino acid sequence. Here, we describe a QSPR strategy for in silico monoclonal antibody purification process fit assessment. Principal Component Analysis is applied to extract a one-dimensional basis for comparison of molecular chromatographic binding behavior from multi-dimensional high-throughput batch-binding screen data. Kernel Ridge Regression is used to predict the first principal component for new molecular sequences. This workflow is demonstrated with a set of 97 monoclonal antibodies for five chromatography resins in two salt types across a range of pH and salt concentrations. Model development benchmarks four descriptor sets from biophysical structural models and protein language models. The investigation illustrates the value QSPR models can provide to purification process fit assessment, and selection of resins and operating conditions from sequence alone.

How generative Artificial Intelligence can transform drug discovery?

Generative Artificial Intelligence (Generative AI) is transforming drug discovery by enabling advanced analysis of complex biological and chemical data. This review explores key Generative AI models, including Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), flow-based models and Transformer-based models, with Transformers gaining prominence due to the abundance of text-based biological data and the success of language models like ChatGPT. The paper discusses molecular representations, performance evaluation metrics, and current trends in Generative AI-driven drug discovery, such as protein-protein interactions (PPIs), drug-target interactions (DTIs) and de-novo drug design. However, these approaches face significant challenges, including applicability domain issues, lack of interpretability, data scarcity, novelty, scalability, computational resource limitations, and the absence of standardized evaluation metrics. These challenges hinder model performance, complicate decision-making, and limit the generation of novel and viable drug candidates. To address these issues, strategies such as hybrid models, integration of multiomics datasets, explainable AI (XAI) techniques, data augmentation, transfer learning, and cloud-based solutions are proposed. Additionally, a curated list of databases supporting drug discovery research is provided. The review concludes by emphasizing the need for optimized AI models, robust validation methods, interdisciplinary collaboration, and future academic efforts to fully realize the potential of Generative AI in advancing drug discovery.

PILOT: Deep Siamese network with hybrid attention improves prediction of mutation impact on protein stability

Evaluating the mutation impact on protein stability (ΔΔG) is essential in the study of protein engineering and understanding molecular mechanisms of disease-associated mutations. Here, we propose a novel deep learning framework, PILOT, for improved prediction of ΔΔG using a Siamese network with hybrid attention mechanism. The PILOT framework leverages multiple attention modules to effectively extract representations for amino acids, atoms, and protein sequences, respectively. This approach significantly ensures the deep fusion of structural information at both residue and atom levels, the seamless integration of structural and sequence representations, and the effective capture of both long-range and short-range dependencies among amino acids. Our extensive evaluations demonstrate that PILOT greatly outperforms other state-of-the-art methods. We also showcase that PILOT identifies exceptional patterns for different mutation types. Moreover, we illustrate the clinical applicability of PILOT in highlighting pathogenic variants from benign variants and VUS (variants of uncertain significance), and distinguishing de novo mutations in disease cases and controls. In summary, PILOT presents a robust deep learning tool that could offer significant insights into drug design, medical applications, and protein engineering studies.

A critical address to advancements and challenges in computational strategies for structural prediction of protein in recent past

Protein structure prediction has undergone significant advancements, driven by the limitations of experimental techniques like X-ray crystallography, NMR, and cryo-EM, which are costly and time-consuming. To bridge the gap between protein sequences and their structures, computational methods have emerged as essential tools. Traditional approaches such as homology modeling, threading, and ab initio folding made progress but often lacked atomic-level precision. The field has been revolutionized by deep learning-based models such as AlphaFold2, RoseTTAFold, and OpenFold, which have demonstrated unprecedented accuracy in predicting protein structures. These AI-driven models leverage vast datasets and neural networks to generate highly reliable structural predictions, sometimes rivaling experimental methods. This review explores the historical evolution of computational protein structure prediction, analyzing the strengths and weaknesses of state-of-the-art models. These models have broad applications in fields such as drug discovery, enzyme engineering, and disease-related protein modeling. However, challenges remain, including the need for extensive training data, computational resource requirements, and difficulties in modeling protein dynamics, intrinsically disordered regions, and protein-protein interactions. Future directions in the field include improving AI models to address current limitations, better integration with experimental techniques, and extending predictions to protein complexes and post-translational modifications. By continuing to refine these methods, computational protein structure prediction will further enhance biomedical research and therapeutic design, reshaping the landscape of structural biology and computational biophysics.

A comprehensive review of computational methods for Protein-DNA binding site prediction

Accurately identifying protein-DNA binding sites is essential for understanding the molecular mechanisms underlying biological processes, which in turn facilitates advancements in drug discovery and design. While biochemical experiments provide the most accurate way to locate DNA-binding sites, they are generally time-consuming, resource-intensive, and expensive. There is a pressing need to develop computational methods that are both efficient and accurate for DNA-binding site prediction. This study thoroughly reviews and categorizes major computational approaches for predicting DNA-binding sites, including template detection, statistical machine learning, and deep learning-based methods. The 14 state-of-the-art DNA-binding site prediction models have been benchmarked on 136 non-redundant proteins, where the deep learning-based, especially pre-trained large language model-based, methods achieve superior performance over the other two categories. Applications of these DNA-binding site prediction methods are also involved.

GPCRchimeraDB: A Database of Chimeric G Protein-Coupled Receptors (GPCRs) to Assist Their Design

G protein-coupled receptors (GPCRs) are membrane proteins crucial to numerous diseases, yet many remain poorly characterized and untargeted by drugs. Chimeric GPCRs have emerged as valuable tools for elucidating GPCR function by facilitating the identification of signaling pathways, resolving structures, and discovering novel ligands of poorly understood GPCRs. Such chimeric GPCRs are obtained by merging a well- and less-well-characterized GPCR at the intracellular limits of their transmembrane regions or intracellular loops, leveraging knowledge transfer from the well-characterized GPCR. However, despite the engineering of over 200 chimeric GPCRs to date, the design process remains largely trial-and-error and lacks a standardized approach. To address this gap, we introduce GPCRchimeraDB (https://www.bio2byte.be/gpcrchimeradb/), the first comprehensive database dedicated to chimeric GPCRs. It catalogs 212 chimeric receptors, identified through literature review, and includes 1,755 class A natural GPCRs, enabling connections between chimeras and their parent receptors while facilitating the exploration of novel parent combinations. Both chimeric and natural GPCR entries are extensively described at the sequence, structural, and biophysical level through a range of visualization tools, with annotations from resources like UniProt and GPCRdb and predictions from AlphaFold2 and b2btools. Additionally, GPCRchimeraDB offers a GPCR sequence aligner and a feature comparator to investigate differences between natural and chimeric receptors. It also provides design guidelines to support rational chimera engineering. GPCRchimeraDB is therefore a resource to facilitate and optimize the design of new chimeras, so helping to gain insights into poorly characterized receptors and contributing to advances in GPCR therapeutic development.

Systems biology of Haemonchus contortus - Advancing biotechnology for parasitic nematode control

Parasitic nematodes represent a substantial global burden, impacting animal health, agriculture and economies worldwide. Of these worms, Haemonchus contortus - a blood-feeding nematode of ruminants - is a major pathogen and a model for molecular and applied parasitology research. This review synthesises some key advances in understanding the molecular biology, genetic diversity and host-parasite interactions of H. contortus, highlighting its value for comparative studies with the free-living nematode Caenorhabditis elegans. Key themes include recent developments in genomic, transcriptomic and proteomic technologies and resources, which are illuminating critical molecular pathways, including the ubiquitination pathway, protease/protease inhibitor systems and the secretome of H. contortus. Some of these insights are providing a foundation for identifying essential genes and exploring their potential as targets for novel anthelmintics or vaccines, particularly in the face of widespread anthelmintic resistance. Advanced bioinformatic tools, such as machine learning (ML) algorithms and artificial intelligence (AI)-driven protein structure prediction, are enhancing annotation capabilities, facilitating and accelerating analyses of gene functions, and biological pathways and processes. This review also discusses the integration of these tools with cutting-edge single-cell sequencing and spatial transcriptomics to dissect host-parasite interactions at the cellular level. The discussion emphasises the importance of curated databases, improved culture systems and functional genomics platforms to translate molecular discoveries into practical outcomes, such as novel interventions. New research findings and resources not only advance research on H. contortus and related nematodes but may also pave the way for innovative solutions to the global challenges with anthelmintic resistance.

Annotating the microbial dark matter with HiFi-NN

The accurate computational annotation of protein sequences with enzymatic function remains a fundamental challenge in bioinformatics. Here, we present HiFi-NN (Hierarchically-Finetuned Nearest Neighbor search) which annotates protein sequences to the 4th level of Enzyme Commission (EC) number with greater precision and recall than state-of-the-art deep learning methods. Furthermore, we show that this method can correctly identify the EC number of a given sequence to lower identities than BLASTp. We show that performance can be improved by increasing the diversity of the lookup set in both sequence space and the environment the sequence has been sampled from. We proceed to show that we can correct specific mis-annotations in the BRENDA enzymes database reproducing results found by others. Finally, we use HiFi-NN to annotate functional dark-matter protein sequences from NMPFamDB. Our findings pave the way for more accurate functional annotation in silico, especially for proteins from distant sequence space.

Cross-species modeling of plant genomes at single-nucleotide resolution using a pretrained DNA language model

Interpreting function and fitness effects in diverse plant genomes requires transferable models. Language models (LMs) pretrained on large-scale biological sequences can capture evolutionary conservation and offer cross-species prediction better than supervised models through fine-tuning limited labeled data. We introduce PlantCaduceus, a plant DNA LM that learns evolutionary conservation patterns in 16 angiosperm genomes by modeling both DNA strands simultaneously. When fine-tuned on a small set of labeled Arabidopsis data for tasks such as predicting translation initiation/termination sites and splice donor/acceptor sites, PlantCaduceus demonstrated remarkable transferability to maize, which diverged 160 Mya. The model outperformed the best existing DNA language model by 1.45-fold in maize splice donor prediction and 7.23-fold in maize translation initiation site prediction. In variant effect prediction, PlantCaduceus showed performance comparative to state-of-the-art protein LMs. Mutations predicted to be deleterious by PlantCaduceus showed threefold lower average minor allele frequencies compared to those identified by multiple sequence alignment-based methods. Additionally, PlantCaduceus successfully identifies well-known causal variants in both Arabidopsis and maize. Overall, PlantCaduceus is a versatile DNA LM that can accelerate plant genomics and crop breeding applications.

Predicting high-fitness viral protein variants with Bayesian active learning and biophysics

The early detection of high-fitness viral variants is critical for pandemic response, yet limited experimental resources at the onset of variant emergence hinder effective identification. To address this, we introduce an active learning framework, VIRAL (Viral Identification via Rapid Active Learning), that integrates protein language model, Gaussian process with uncertainty estimation, and a biophysical model to predict the fitness of novel variants in a few-shot learning setting. By benchmarking on past SARS-CoV-2 data, we demonstrate that our method accelerates the identification of high-fitness variants by up to fivefold compared to random sampling while requiring experimental characterization of fewer than 1% of possible variants. We also demonstrate that our framework effectively identifies sites that are frequently mutated during natural viral evolution with a predictive advantage of up to two years compared to baseline strategies, particularly those enabling antibody escape while preserving ACE2 binding. Through systematic analysis of different acquisition strategies, we show that incorporating uncertainty in variant selection enables broader exploration of the sequence landscape, leading to the identification of evolutionarily distant but potentially dangerous variants. Our results suggest that VIRAL could serve as an effective early warning system for identifying concerning SARS-CoV-2 variants and potentially emerging viruses with pandemic potential before they achieve widespread circulation.

The risk of pathogenicity and antibiotic resistance in deep-sea cold seep microorganisms

Deep-sea cold seeps host high microbial biomass and biodiversity that thrive on hydrocarbon and inorganic compound seepage, exhibiting diverse ecological functions and unique genetic resources. However, potential health risks from pathogenic or antibiotic-resistant microorganisms in these environments remain largely overlooked, especially during resource exploitation and laboratory research. Here, we analyzed 165 metagenomes and 33 metatranscriptomes from 16 global cold seep sites to investigate the diversity and distribution of virulence factors (VFs), antibiotic resistance genes (ARGs), and mobile genetic elements (MGEs). A total of 2,353 VFs are retrieved in 689 metagenome-assembled genomes (MAGs), primarily associated with indirect pathogenesis like adherence. In addition, cold seeps harbor nearly 100,000 ARGs, as important reservoirs, with high-risk ARGs (11.22%) presenting at low abundance. Compared to other environments, microorganisms in cold seeps exhibit substantial differences in VF and ARG counts, with potential horizontal gene transfer facilitating their spread. These virulome and resistome profiles provide valuable insights into the evolutionary and ecological implications of pathogenicity and antibiotic resistance in extreme deep-sea ecosystems. Collectively, these results indicate that cold seep sediments pose minimal public health risks, shedding light on environmental safety in deep-sea resource exploitation and research.

IMPORTANCE

In the "One Health" era, understanding pathogenicity and antibiotic resistance in vast and largely unexplored regions like deep-sea cold seeps is critical for assessing public health risks. These environments serve as critical reservoirs where resistant and virulent bacteria can persist, adapt, and undergo genetic evolution. The increasing scope of human activities, such as deep-sea mining, is disrupting these previously isolated ecosystems, heightening the potential for microbial exchange between deep-sea communities and human or animal populations. This interaction poses a significant risk for the dissemination of resistance and virulence genes, with potential consequences for global public health and ecosystem stability. This study offers the first comprehensive analysis of virulome, resistome, and mobilome profiles in cold seep microbial communities. While cold seeps act as reservoirs for diverse ARGs, high-risk ARGs are rare, and most VFs were low risk that contribute to ecological functions. These results provide a reference for monitoring the spread of pathogenicity and resistance in extreme ecosystems, informing environmental safety assessments during deep-sea resource exploitation.

PLPTP: A Motif-based Interpretable Deep Learning Framework Based on Protein Language Models for Peptide Toxicity Prediction

Peptide toxicity prediction holds significant importance in drug development and biotechnology, as accurately identifying toxic peptide sequences is crucial for designing safer peptide-based drugs. This study proposes a deep learning-based model for peptide toxicity prediction, integrating Evolutionary Scale Modeling (ESM2), Bidirectional Long Short-Term Memory (BiLSTM), and Deep Neural Network (DNN). The ESM2 model captures evolutionary information from peptide sequences, providing a rich context for the sequences; the BiLSTM network focuses on extracting contextual dependencies, thereby capturing long-range dependencies within the sequence; and the DNN further classifies the extracted features to achieve the final toxicity prediction. To enhance the reliability and transparency of the model, we also conducted motif analysis to identify key patterns in the data, which helps to explain the model's attention mechanism and its classification performance. To address the class imbalance in the dataset, we employed Focal Loss as the loss function, which enhances the model's ability to identify minority class samples by reducing the contribution of easily classified samples. Experimental results demonstrate that the proposed model performs exceptionally well across multiple evaluation metrics, particularly in handling imbalanced data, achieving significant improvements over traditional methods. This result highlights the model's potential to improve the accuracy of peptide toxicity prediction and its valuable role in drug development and biotechnology research. The PLPTP web server is available at https://www.bioai-lab.com/PLPTP.

Exploring 4th generation EGFR inhibitors: A review of clinical outcomes and structural binding insights

Epidermal growth factor receptor (EGFR) is a potential target for anticancer therapies and plays a crucial role in cell growth, survival, and metastasis. EGFR gene mutations trigger aberrant signaling, leading to non-small cell lung cancer (NSCLC). Tyrosine kinase inhibitors (TKIs) effectively target these mutations to treat NSCLC. While the first three generations of EGFR TKIs have been proven effective, the emergence of the EGFR-C797S resistance mutation poses a new challenge. To address this, various synthetic EGFR TKIs have been developed. In this review, we have summarized the EGFR TKIs reported in the past five years, focusing on their clinical outcomes and structure-activity relationship analysis. We have also explored binding modes and interactions between the binding pocket and ligands to provide insights into the mechanisms of these inhibitors, which contribute to advancements in targeted cancer therapy. Additionally, artificial Intelligence-driven methods, including recursive neural networks and reinforcement learning, have revolutionized EGFR inhibitor design by facilitating rapid screening, predicting EGFR mutations, and novel compound generation.

AbEpiTope-1.0: Improved antibody target prediction by use of AlphaFold and inverse folding

B cell epitope prediction tools are crucial for designing vaccines and disease diagnostics. However, predicting which antigens a specific antibody binds to and their exact binding sites (epitopes) remains challenging. Here, we present AbEpiTope-1.0, a tool for antibody-specific B cell epitope prediction, using AlphaFold for structural modeling and inverse folding for machine learning models. On a dataset of 1730 antibody-antigen complexes, AbEpiTope-1.0 outperforms AlphaFold in predicting modeled antibody-antigen interface accuracy. By creating swapped antibody-antigen complex structures for each antibody-antigen complex using incorrect antibodies, we show that predicted accuracies are sensitive to antibody input. Furthermore, a model variant optimized for antibody target prediction-differentiating true from swapped complexes-achieved an accuracy of 61.21% in correctly identifying antibody-antigen pairs. The tool evaluates hundreds of structures in minutes, providing researchers with a resource for screening antibodies targeting specific antigens. AbEpiTope-1.0 is freely available as a web server and software.

Structural Comparisons of Bifunctional Fatty Acid Dioxygenases with Allene Oxide, Epoxy Alcohol, or Diol Synthase Activities

Cyclooxygenase, lipoxygenases, and other dioxygenases (DOXs) transform fatty acids to endoperoxides or hydroperoxides, which can be isomerized by cytochromes P450 (CYPs) class III. These enzymes produce biological mediators in health and disease. Prostaglandin H2 is isomerized by thromboxane (CYP5A1) and prostacyclin synthases (CYP8A1), and hydroperoxides produced by plants by allene oxide synthases (AOS) (Cyp74A). Bifunctional DOXs belong to the peroxidase-COX superfamily and contain CYP fusion partners. They sequentially oxidize unsaturated C18 fatty acids to diols, epoxy alcohols, and allene oxides via hetero- or homolytic scissions of hydroperoxides. AlphaFold2 (AF2) models of bifunctional 8R/8S-, 9R/9S-, and 10R-DOXs illustrate the substrate recognition sites (SRSs) of the fused AOS, linoleate diol (LDS), and 10R-epoxy alcohol synthases (EAS). An Asn residue (SRS4) was positioned opposite to the heme iron of LDS, 10R-EAS, CYP8A1, and Cyp74A, but replaced at this position by Thr or nonpolar residues (Ile/Val/Ala) of 8R/9R- and 8S/9S-AOS, respectively, and by Ile of CYP5A1. Replacements of Asn and Gln residues of the SRS4 of LDSs altered the position of oxygenation and the homolytic scission of the hydroperoxide, but the amide residues were not required. The AF2 models of the CYP fusion partners illustrated the active sites in unprecedented details.

A generative language model decodes contextual constraints on codon choice for mRNA design

The genetic code allows multiple synonymous codons to encode the same amino acid, creating a vast sequence space for protein-coding regions. Codon choice can impact mRNA function and protein output, a consideration newly relevant with advances in mRNA technology. Genomes preferentially use some codons, but simple optimization methods that select preferred codons miss complex contextual patterns. We present Trias, an encoder-decoder language model trained on millions of eukaryotic coding sequences. Trias learns codon usage rules directly from sequence data, integrating local and global dependencies to generate species-specific codon sequences that align with biological constraints. Without explicit training on protein expression, Trias generates sequences and scores that correlate strongly with experimental measurements of mRNA stability, ribosome load, and protein output. The model outperforms commercial codon optimization tools in generating sequences resembling high-expression codon sequence variants. By modeling codon usage in context, Trias offers a data-driven framework for synthetic mRNA design and for understanding the molecular and evolutionary principles behind codon choice.

Structural diversity and oligomerization of bacterial ubiquitin-like proteins

Bacteria possess a variety of operons with homology to eukaryotic ubiquitination pathways that encode predicted E1, E2, E3, deubiquitinase, and ubiquitin-like proteins. Some of these pathways have recently been shown to function in anti-bacteriophage immunity, but the biological functions of others remain unknown. Here, we show that ubiquitin-like proteins in two bacterial operon families show surprising architectural diversity, possessing one to three β-grasp domains preceded by diverse N-terminal domains. We find that a large group of bacterial ubiquitin-like proteins possess three β-grasp domains and form homodimers and helical filaments mediated by conserved Ca2+ ion binding sites. Our findings highlight a distinctive mode of self-assembly for ubiquitin-like proteins and suggest that Ca2+-mediated ubiquitin-like protein filament assembly and/or disassembly enables cells to sense and respond to stress conditions that alter intracellular metal ion concentration.

Improving AlphaFold2- and AlphaFold3-Based Protein Complex Structure Prediction With MULTICOM4 in CASP16

With AlphaFold achieving high-accuracy tertiary structure prediction for most single-chain proteins (monomers), the next major challenge in protein structure prediction is to accurately model multichain protein complexes (multimers). We developed MULTICOM4, the latest version of the MULTICOM system, to improve protein complex structure prediction by integrating transformer-based AlphaFold2, diffusion model-based AlphaFold3, and our in-house techniques. These include protein complex stoichiometry prediction, diverse multiple sequence alignment (MSA) generation leveraging both sequence and structure comparison, modeling exception handling, and deep learning-based protein model quality assessment. MULTICOM4 was blindly evaluated in the 16th Critical Assessment of Techniques for Protein Structure Prediction (CASP16) in 2024. In Phase 0 of CASP16, where stoichiometry information was unavailable, MULTICOM predictors performed best, with MULTICOM_human achieving a TM-score of 0.752 and a DockQ score of 0.584 for top-ranked predictions on average. In Phase 1 of CASP16, with stoichiometry information provided, MULTICOM_human remained among the top predictors, attaining a TM-score of 0.797 and a DockQ score of 0.558 on average. The CASP16 results demonstrate that integrating complementary AlphaFold2 and AlphaFold3 with enhanced MSA inputs, comprehensive model ranking, exception handling, and accurate stoichiometry prediction can effectively improve protein complex structure prediction.

Leveraging protein language models for robust antimicrobial peptide detection

Antimicrobial peptides (AMPs) are promising candidates for addressing the global challenge of antibiotic resistance due to their broad-spectrum antimicrobial properties. Traditional AMP identification methods, while effective, are labor-intensive and time-consuming. Recent advancements in deep learning and large language models (LLMs), especially protein language models (PLMs) present a transformative approach for AMP prediction. In this study, we propose PLAPD, a novel framework leveraging a pre-trained ESM2 protein language model for AMP classification. Besides, PLAPD combines local feature extraction via convolutional layers and global feature extraction with a residual Transformer module. We benchmarked PLAPD against state-of-the-art AMP prediction models using a dataset comprising 8,268 peptide sequences, achieving superior performance in Accuracy (0.87), Precision (0.9359), Specificity (0.9456), MCC (0.7486), and AUC (0.9225). The results highlight the potential of PLAPD as a high-throughput and accurate tool for AMP discovery.

Predicting protein-protein interaction with interpretable bilinear attention network

BACKGROUND AND OBJECTIVE

Protein-protein interactions (PPIs) play the key roles in myriad biological processes, helping to understand the protein function and disease pathology. Identification of PPIs and their interaction types through wet experimental methods are costly and time-consuming. Therefore, some computational methods (e.g., sequence-based deep learning method) have been proposed to predict PPIs. However, these methods predominantly focus on protein sequence information, neglecting the protein structure information, while the protein structure is closely related to its function. In addition, current PPI prediction methods that introduce the protein structure information use independent encoders to learn the sequence and structure representations from protein sequences and structures, respectively, without explicitly learn the important local interaction representation of two proteins, making the prediction results hard to interpret.

METHODS

Considering that current protein structure prediction methods (e.g., AlphaFold2) can accurately predict protein 3D structures and also provide a large number of protein 3D structures, here we present a novel end-to-end framework (called PPI-BAN) to predict PPIs and their interaction types by integrating protein sequence information and 3D structure information. PPI-BAN uses one-dimensional convolution operation (Conv1D) to extract the protein sequence features, employes GeomEtry-Aware Relational Graph Neural Network (GearNet) to learn protein 3D structure features, and adopts a deep bilinear attention network (BAN) to learn the joint features between one protein sequence and its 3D structure. The sequence features, structure features and joint features are concatenated to fed into a fully connected network for predicting PPIs and their interaction types.

RESULTS

Experimental results show that PPI-BAN achieves the best overall performance against other state-of-the-art methods.

CONCLUSIONS

PPI-BAN can effectively predict PPIs and their interaction types, and identify the significant interaction sites by computing attention weight maps and mapping them to specific amino acid residues.

Prediction of protein subcellular localization in single cells

The subcellular localization of a protein is important for its function, and its mislocalization is linked to numerous diseases. Existing datasets capture limited pairs of proteins and cell lines, and existing protein localization prediction models either miss cell-type specificity or cannot generalize to unseen proteins. Here we present a method for Prediction of Unseen Proteins' Subcellular localization (PUPS). PUPS combines a protein language model and an image inpainting model to utilize both protein sequence and cellular images. We demonstrate that the protein sequence input enables generalization to unseen proteins, and the cellular image input captures single-cell variability, enabling cell-type-specific predictions. Experimental validation shows that PUPS can predict protein localization in newly performed experiments outside the Human Protein Atlas used for training. Collectively, PUPS provides a framework for predicting differential protein localization across cell lines and single cells within a cell line, including changes in protein localization driven by mutations.

The Pseudoenzyme β-Amylase9 From Arabidopsis Activates α-Amylase3: A Possible Mechanism to Promote Stress-Induced Starch Degradation

Starch accumulation in plants provides carbon for nighttime use, for regrowth after periods of dormancy, and for times of stress. Both ɑ- and β-amylases (AMYs and BAMs, respectively) catalyze starch hydrolysis, but their functional roles are unclear. Moreover, the presence of catalytically inactive amylases that show starch excess phenotypes when deleted presents questions on how starch degradation is regulated. Plants lacking one of these catalytically inactive β-amylases, BAM9, have enhanced starch accumulation when combined with mutations in BAM1 and BAM3, the primary starch degrading BAMs in response to stress and at night, respectively. BAM9 has been reported to be transcriptionally induced by stress although the mechanism for BAM9 function is unclear. From yeast two-hybrid experiments, we identified the plastid-localized AMY3 as a potential interaction partner for BAM9. We found that BAM9 interacted with AMY3 in vitro and that BAM9 enhances AMY3 activity about three-fold. Modeling of the AMY3-BAM9 complex predicted a previously undescribed alpha-alpha hairpin in AMY3 that could serve as a potential interaction site. Additionally, AMY3 lacking the alpha-alpha hairpin is unaffected by BAM9. Structural analysis of AMY3 showed that it can form a homodimer in solution and that BAM9 appears to replace one of the AMY3 monomers to form a heterodimer. The presence of both BAM9 and AMY3 in many vascular plant lineages, along with model-based evidence that they heterodimerize, suggests that the interaction is conserved. Collectively these data suggest that BAM9 is a pseudoamylase that activates AMY3 in response to cellular stress, possibly facilitating stress recovery.

Unveiling novel antioxidant peptides from silk fibroin proteins: An integrated in silico and in vitro study

Antioxidant peptides exhibit significant potential in combating degenerative diseases by effectively mitigating oxidative stress. In this study, we developed a machine-learning model for screening antioxidant peptides, achieving a Matthews correlation coefficient of 0.892 ± 0.033 and surpassing the state-of-the-art (SOTA) models. Through in silico screening, seven novel antioxidant peptides derived from silk fibroin proteins (SFP) were identified (i.e., DEDY, NEEY, GAGRGY, ITRNHDQCR, VDHNL, QGDY, and DDY) and subsequently synthesized. Among them, all except for GAGRGY and QGDY demonstrated notable antioxidant activity in ABTS free radical assays, which were 1.26-3.25 times higher than that of glutathione. All seven antioxidant peptides effectively protected erythrocytes from oxidative damage. This protective capacity is likely attributed to their ability to bind free radicals and regulate the Keap1-Nrf2 pathway. Overall, this study presents an effective strategy for discovering antioxidant peptides from SFP and provides strong experimental validation for testing the effectiveness of the machine learning model.

A Computational Approach to Characterize the Protein S-Mer Tyrosine Kinase (PROS1-MERTK) Protein-Protein Interaction Dynamics

Protein S (PROS1) has recently been identified as a ligand for the TAM receptor MERTK, influencing immune response and cell survival. The PROS1-MERTK interaction plays a role in cancer progression, promoting immune evasion and metastasis in multiple cancers by fostering a tumor-supportive microenvironment. Despite its importance, limited structural insights into this interaction underscore the need for computational studies to explore their binding dynamics, potentially guiding targeted therapies. In this study, we investigated the PROS1-MERTK interaction using advanced computational analyses to support immunotherapy research. High-resolution structural models from ColabFold, an AlphaFold2 adaptation, provided a baseline structure, allowing us to examine the PROS1-MERTK interface with ChimeraX and map residue interactions through Van der Waals criteria. Molecular dynamics (MD) simulations were conducted in GROMACS over 100 ns to assess stability and conformational changes using RMSD, RMSF, and radius of gyration (Rg). The PROS1-MERTK interface was predicted to contain a heterogeneous mix of amino acid contacts, with lysine and leucine as frequent participants. MD simulations demonstrated prominent early structural shifts, stabilizing after approximately 50 ns with small conformational shifts occurring as the simulation completed. In addition, there are various regions in each protein that are predicted to have greater conformational fluctuations as compared to others, which may represent attractive areas to target to halt the progression of the interaction. These insights deepen our understanding of the PROS1-MERTK interaction role in immune modulation and tumor progression, unveiling potential targets for cancer immunotherapy.

Evaluating sequence and structural similarity metrics for predicting shared paralog functions

Gene duplication is the primary source of new genes, resulting in most genes having identifiable paralogs. Over time, paralog pairs may diverge in some respects but many retain the ability to perform the same functional role. Protein sequence identity is often used as a proxy for functional similarity and can predict shared functions between paralogs as revealed by synthetic lethal experiments. However, the advent of alternative protein representations, including embeddings from protein language models (PLMs) and predicted structures from AlphaFold, raises the possibility that alternative similarity metrics could better capture functional similarity between paralogs. Here, using two species (budding yeast and human) and two different definitions of shared functionality (shared protein-protein interactions and synthetic lethality), we evaluated a variety of alternative similarity metrics. For some tasks, predicted structural similarity or PLM similarity outperform sequence identity, but more importantly these similarity metrics are not redundant with sequence identity, i.e. combining them with sequence identity leads to improved predictions of shared functionality. By adding contextual features, representing similarity to homologous proteins within and across species, we can significantly enhance our predictions of shared paralog functionality. Overall, our results suggest that alternative similarity metrics capture complementary aspects of functional similarity beyond sequence identity alone.

Fast, sensitive detection of protein homologs using deep dense retrieval

The identification of protein homologs in large databases using conventional methods, such as protein sequence comparison, often misses remote homologs. Here, we offer an ultrafast, highly sensitive method, dense homolog retriever (DHR), for detecting homologs on the basis of a protein language model and dense retrieval techniques. Its dual-encoder architecture generates different embeddings for the same protein sequence and easily locates homologs by comparing these representations. Its alignment-free nature improves speed and the protein language model incorporates rich evolutionary and structural information within DHR embeddings. DHR achieves a >10% increase in sensitivity compared to previous methods and a >56% increase in sensitivity at the superfamily level for samples that are challenging to identify using alignment-based approaches. It is up to 22 times faster than traditional methods such as PSI-BLAST and DIAMOND and up to 28,700 times faster than HMMER. The new remote homologs exclusively found by DHR are useful for revealing connections between well-characterized proteins and improving our knowledge of protein evolution, structure and function.

Designing diverse and high-performance proteins with a large language model in the loop

We present a protein engineering approach to directed evolution with machine learning that integrates a new semi-supervised neural network fitness prediction model, Seq2Fitness, and an innovative optimization algorithm, biphasic annealing for diverse and adaptive sequence sampling (BADASS) to design sequences. Seq2Fitness leverages protein language models to predict fitness landscapes, combining evolutionary data with experimental labels, while BADASS efficiently explores these landscapes by dynamically adjusting temperature and mutation energies to prevent premature convergence and to generate diverse high-fitness sequences. Compared to alternative models, Seq2Fitness improves Spearman correlation with experimental fitness measurements, increasing from 0.34 to 0.55 on sequences containing mutations at positions entirely not seen during training. BADASS requires less memory and computation compared to gradient-based Markov Chain Monte Carlo methods, while generating more high-fitness and diverse sequences across two protein families. For both families, 100% of the top 10,000 sequences identified by BADASS exceed the wildtype in predicted fitness, whereas competing methods range from 3% to 99%, often producing far fewer than 10,000 sequences. BADASS also finds higher-fitness sequences at every cutoff (top 1, 100, and 10,000). Additionally, we provide a theoretical framework explaining BADASS's underlying mechanism and behavior. While we focus on amino acid sequences, BADASS may generalize to other sequence spaces, such as DNA and RNA.

High-Throughput Evaluation of Natural Diversity of F-Type ATP Synthase Rotor Ring Stoichiometries

Adenosine triphosphate (ATP) synthases are large enzymes present in every living cell. They consist of a transmembrane and a soluble domain, each comprising multiple subunits. The transmembrane part contains an oligomeric rotor ring (c-ring), whose stoichiometry defines the ratio between the number of synthesized ATP molecules and the number of ions transported through the membrane. Currently, c-rings of F-Type ATP synthases consisting of 8-17 (except 16) subunits have been experimentally demonstrated, but it is not known whether other stoichiometries are present in natural organisms. Here, we present an easy-to-use high-throughput computational approach based on AlphaFold that allows us to estimate the stoichiometry of all homo-oligomeric c-rings, whose sequences are present in genomic databases. We validate the approach on the available experimental data, obtaining the correlation as high as 0.94 for the reference dataset and use it to predict the existence of c-rings with stoichiometry varying at least from 8 to 27. We then conduct molecular dynamics simulations of two c-rings with stoichiometry above 17 to corroborate the machine learning-based predictions. Our work strongly suggests existence of rotor rings with previously undescribed high stoichiometry in natural organisms and highlights the utility of AlphaFold-based approaches for studying homo-oligomeric proteins.

AcrDB update: Predicted 3D structures of anti-CRISPRs in human gut viromes

Anti-CRISPR (Acr) proteins play a key role in phage-host interactions and hold great promise for advancing genome-editing technologies. However, finding new Acrs has been challenging due to their low sequence similarity. Recent advances in protein structure prediction have opened new pathways for Acr discovery by using 3D structure similarity. This study presents an updated AcrDB, with the following new features not available in other databases: (1) predicted Acrs from human gut virome databases, (2) Acr structures predicted by AlphaFold2, (3) a structural similarity search function to allow users to submit new sequences and structures to search against 3D structures of experimentally known Acrs. The updated AcrDB contains predicted 3D structures of 795 candidate Acrs with structural similarity (TM-score ≥0.7) to known Acrs supported by at least two of the three non-sequence similarity-based tools (TM-Vec, Foldseek, AcrPred). Among these candidate Acrs, 121 are supported by all three tools. AcrDB also includes 3D structures of 122 experimentally characterized Acr proteins. The 121 most confident candidate Acrs were combined with the 122 known Acrs and clustered into 163 sequence similarity-based Acr families. The 163 families were further subject to a structure similarity-based hierarchical clustering, revealing structural similarity between 44 candidate Acr (cAcr) families and 119 known Acr families. The bacterial hosts of these 163 Acr families are mainly from Bacillota, Pseudomonadota, and Bacteroidota, which are all dominant gut bacterial phyla. Many of these 163 Acr families are also co-localized in Acr operons. All the data and visualization are provided on our website: https://pro.unl.edu/AcrDB.

DeepCR: predicting cytokine receptor proteins through pretrained language models and deep learning networks

Cytokine receptors play a pivotal role in mediating the immune response and are critical in cytokine storms, which underlie the pathogenesis of conditions such as acute respiratory distress syndrome (ARDS) and autoimmune disorders. Identifying cytokine receptors is essential for understanding their biological functions, exploring therapeutic targets, and guiding clinical interventions. Traditional biochemical methods to identify cytokine receptors are labor-intensive, costly, and time-consuming, prompting the need for more efficient alternatives. Recent advances in computational biology have enabled the use of machine learning to classify cytokine receptor proteins. Most existing approaches focused on homologous features and protein composition to classify cytokine families, but no dedicated studies have been conducted on cytokine receptor proteins. This gap presents an opportunity to develop a method specifically for classifying cytokine receptors among other membrane proteins. In this study, we present a novel classification framework combining pre-trained language models (PLMs) with a multi-window convolutional neural network (mCNN) architecture for the fast and accurate identification of cytokine receptor proteins. PLMs, such as ProtTrans and ESM variants, capture biochemical context directly from raw protein sequences, while mCNN efficiently extracts local and global sequence patterns using convolutional layers with varying window sizes. Our model achieved an AUC of 0.96 in the training as well as 0.97 and 0.93 in two independent tests, demonstrating its effectiveness in distinguishing cytokine receptors from non-cytokine receptor proteins. By eliminating the need for manual feature extraction, this approach offers a robust and scalable solution for protein classification, paving the way for its application in drug discovery and understanding cytokine-mediated diseases.

Evolutionary insights into Interleukin-12 family targets across 405 species

The Interleukin-12 (IL-12) family ligand subunits (IL-12s) and receptor subunits (IL-12Rs) constitute pivotal regulators of immune homeostasis, with direct implications in autoimmune pathologies and oncogenesis. Through phylogenetic reconstruction, synteny analysis, and sequence alignment across 400+ animal species, we delineated the evolutionary trajectories and functional diversification of these immune mediators. Phylogenetic analysis revealed IL-12Rs originated prior to the mollusk era (514-686.2 million years ago, Mya), while ligand subunits p19/p28 emerged during the mammalian and avian epoch (180-225 Mya). Structural characterization identified three evolutionarily invariant signature motifs within the fibronectin type III (fn3) domain essential for receptor-ligand interface stability. Furthermore, phylogenetically ultra-conserved residue and motif configurations were mapped, revealing candidate therapeutic epitopes. These findings establish an evolutionary framework explaining functional conservation/divergence in IL-12 signaling components. The identified ancient receptor architectures coupled with derived ligand innovations provide a blueprint for cross-species immunotherapy design targeting conserved interaction interfaces. The conserved molecular signatures offer dual utility in developing precision therapies and interspecies disease management strategies, particularly for translational applications across human medicine, agriculture, and aquaculture.

Protein Sequence Analysis landscape: A Systematic Review of Task Types, Databases, Datasets, Word Embeddings Methods, and Language Models

Protein sequence analysis examines the order of amino acids within protein sequences to unlock diverse types of a wealth of knowledge about biological processes and genetic disorders. It helps in forecasting disease susceptibility by finding unique protein signatures, or biomarkers that are linked to particular disease states. Protein Sequence analysis through wet-lab experiments is expensive, time-consuming and error prone. To facilitate large-scale proteomics sequence analysis, the biological community is striving for utilizing AI competence for transitioning from wet-lab to computer aided applications. However, Proteomics and AI are two distinct fields and development of AI-driven protein sequence analysis applications requires knowledge of both domains. To bridge the gap between both fields, various review articles have been written. However, these articles focus revolves around few individual tasks or specific applications rather than providing a comprehensive overview about wide tasks and applications. Following the need of a comprehensive literature that presents a holistic view of wide array of tasks and applications, contributions of this manuscript are manifold: It bridges the gap between Proteomics and AI fields by presenting a comprehensive array of AI-driven applications for 63 distinct protein sequence analysis tasks. It equips AI researchers by facilitating biological foundations of 63 protein sequence analysis tasks. It enhances development of AI-driven protein sequence analysis applications by providing comprehensive details of 68 protein databases. It presents a rich data landscape, encompassing 627 benchmark datasets of 63 diverse protein sequence analysis tasks. It highlights the utilization of 25 unique word embedding methods and 13 language models in AI-driven protein sequence analysis applications. It accelerates the development of AI-driven applications by facilitating current state-of-the-art performances across 63 protein sequence analysis tasks.

BindUP-Alpha: A Webserver for Predicting DNA- and RNA-Binding Proteins based on Experimental and Computational Structural Models

Structural data provides important information on the proteins' function. Recent development of advanced machine learning and artificial intelligence tools, such as AlphaFold, have led to an explosion of predicted protein structures. However, many of the computed protein models contain unstructured and disordered regions, posing challenges in protein function characterization. Here we present BindUP-Alpha, an upgraded webserver for predicting nucleic acid binding proteins. Our structure-based algorithm utilizes the electrostatic features of the protein surface and other physiochemical and structural properties extracted from the protein sequence. Using a Support Vector Machine (SVM) learning approach, BindUP-Alpha successfully predicts DNA- and RNA-binding proteins from both experimentally solved structures and predicted models. In addition, BindUP-Alpha identifies electrostatic patches on the protein's surface that represent potential nucleic-acid binding interfaces. BindUP-Alpha is freely accessible at https://bindup.technion.ac.il, providing interactive three-dimensional visualizations and downloadable text-based results.

NMR Structural Characterization of SARS-CoV-2 ORF6 Reveals an N-Terminal Membrane Anchor

SARS-CoV-2, the virus responsible for the COVID-19 pandemic, encodes several accessory proteins, among which ORF6, a potent interferon inhibitor, is recognized as one of the most cytotoxic. Here, we investigated the structure, oligomeric state, and membrane interactions of ORF6 using NMR spectroscopy and molecular dynamics simulations. Using chemical-shift-ROSETTA, we show that ORF6 in proteoliposomes adopts a straight α-helical structure with an extended, rigid N-terminal part and flexible C-terminal residues. Cross-linking experiments indicate that ORF6 forms oligomers within lipid bilayers, and paramagnetic spin labeling suggests an antiparallel arrangement in its multimers. The amphipathic ORF6 helix establishes multiple contacts with the membrane surface with its N-terminal residues acting as membrane anchors. Our work demonstrates that ORF6 is an integral monotopic membrane protein and provides key insights into its conformation and the importance of the N-terminal region for the interaction with the membrane.

Computer-Aided Drug Discovery for Undruggable Targets

Undruggable targets are those of therapeutical significance but challenging for conventional drug design approaches. Such targets often exhibit unique features, including highly dynamic structures, a lack of well-defined ligand-binding pockets, the presence of highly conserved active sites, and functional modulation by protein-protein interactions. Recent advances in computational simulations and artificial intelligence have revolutionized the drug design landscape, giving rise to innovative strategies for overcoming these obstacles. In this review, we highlight the latest progress in computational approaches for drug design against undruggable targets, present several successful case studies, and discuss remaining challenges and future directions. Special emphasis is placed on four primary target categories: intrinsically disordered proteins, protein allosteric regulation, protein-protein interactions, and protein degradation, along with discussion of emerging target types. We also examine how AI-driven methodologies have transformed the field, from applications in protein-ligand complex structure prediction and virtual screening to de novo ligand generation for undruggable targets. Integration of computational methods with experimental techniques is expected to bring further breakthroughs to overcome the hurdles of undruggable targets. As the field continues to evolve, these advancements hold great promise to expand the druggable space, offering new therapeutic opportunities for previously untreatable diseases.

LucaPCycle: Illuminating microbial phosphorus cycling in deep-sea cold seep sediments using protein language models

Phosphorus is essential for life and critically influences marine productivity. Despite geochemical evidence of active phosphorus cycling in deep-sea cold seeps, the microbial processes involved remain poorly understood. Traditional sequence-based searches often fail to detect proteins with remote homology. To address this, we developed a deep learning model, LucaPCycle, integrating raw sequences and contextual embeddings based on the protein language model ESM2-3B. LucaPCycle identified 5241 phosphorus-cycling protein families from global cold seep gene and genome catalogs, substantially enhancing our understanding of their diversity, ecology, and function. Among previously unannotated sequences, we discovered three alkaline phosphatase families that feature unique domain organizations and preserved enzymatic capabilities. These results highlight previously overlooked ecological importance of phosphorus cycling within cold seeps, corroborated by data from porewater geochemistry, metatranscriptomics, and metabolomics. We revealed a previously unrecognized diversity of archaea, including Asgardarchaeota, anaerobic methanotrophic archaea and Thermoproteota, which contribute to organic phosphorus mineralization and inorganic phosphorus solubilization through various mechanisms. Additionally, auxiliary metabolic genes of cold seep viruses primarily encode the PhoR-PhoB regulatory system and PhnCDE transporter, potentially enhancing their hosts' phosphorus utilization. Overall, LucaPCycle are capable of accessing previously 'hidden' sequence spaces for microbial phosphorus cycling and can be applied to various ecosystems.

TransABseq: A Two-Stage Approach for Predicting Antigen-Antibody Binding Affinity Changes upon Mutation Based on Protein Sequences

The antigen-antibody interaction represents a critical mechanism in host defense, contributing to pathogen neutralization, tumor surveillance, immunotherapy, and in vitro disease detection. Owing to their exceptional specificity, affinity, and selectivity, antibodies have been extensively utilized in the development of clinical diagnostic, therapeutic, and prophylactic strategies. In this study, we propose TransABseq, a novel computational framework specifically designed to predict the effects of missense mutations on antigen-antibody interactions. The model's innovative two-stage architecture enables comprehensive feature analysis: in the first stage, multiple embeddings of protein language models are processed through a Transformer encoder module and a multiscale convolutional module; in the second stage, the XGBOOST model is used to perform quantitative output based on the deeply fused features. A critical advancement contributing to the effectiveness of TransABseq is the deep feature fusion strategy, which reveals the biochemical properties of proteins. By leveraging the multilayer self-attention mechanism of the Transformer to capture complex global dependencies within sequences and mining features at different hierarchical levels through multiscale convolution, the feature abstraction capability of TransABseq is significantly enhanced. We evaluated TransABseq through three distinct cross-validation strategies on two established benchmarks and a newly reconstructed data set. As a result, TransABseq achieved average PCC values of 0.607, 0.843, and 0.794 and average RMSE values of 1.166, 1.314, and 1.337 kcal/mol in 10-fold cross-validation. Furthermore, its robustness and predictive accuracy were validated on blind test data sets, where TransABseq outperformed existing methods, enabling it to attain a PCC of 0.721 and an RMSE of 0.925 kcal/mol. The relevant data and code have been made publicly available for academic research at: https://github.com/cuifengLI/TransABseq.

Advancements in Ligand-Based Virtual Screening through the Synergistic Integration of Graph Neural Networks and Expert-Crafted Descriptors

The fusion of traditional chemical descriptors with graph neural networks (GNNs) offers a compelling strategy for enhancing ligand-based virtual screening methodologies. A comprehensive evaluation revealed that the benefits derived from this integrative strategy vary significantly among different GNNs. Specifically, while GCN and SchNet demonstrate pronounced improvements by incorporating descriptors, SphereNet exhibits only marginal enhancement. Intriguingly, despite SphereNet's modest gain, all three models-GCN, SchNet, and SphereNet-achieve comparable performance levels when leveraging this combination strategy. This observation underscores a pivotal insight: sophisticated GNN architectures may be substituted with simpler counterparts without sacrificing efficacy, provided that they are augmented with descriptors. Furthermore, our analysis reveals a set of expert-crafted descriptors' robustness in scaffold-split scenarios, frequently outperforming the combined GNN-descriptor models. Given the critical importance of scaffold splitting in accurately mimicking real-world drug discovery contexts, this finding accentuates an imperative for GNN researchers to innovate models that can adeptly navigate and predict within such frameworks. Our work not only validates the potential of integrating descriptors with GNNs in advancing ligand-based virtual screening but also illuminates pathways for future enhancements in model development and application. Our implementation can be found at https://github.com/meilerlab/gnn-descriptor.

Advancing Enzyme Optimal pH Prediction via Retrieved Embedding Data Augmentation

The optimal enzyme pH is a critical factor that directly influences the catalytic efficiency of the enzymes. Accurate computational prediction of the optimal pH can greatly advance our understanding and design of enzymes for diverse scientific and industrial applications. However, current prediction tools often fall short in terms of accuracy and robustness. In this study, we propose OpHReda, a novel method that significantly improves enzyme optimal pH prediction by leveraging a retrieved embedding data augmentation mechanism. Given an enzyme sequence, OpHReda first retrieves similar sequence embeddings from a preconstructed augmentation database. It then jointly analyzes the original and retrieved embeddings through the Multiple Embedding Alignment transformer to narrow the prediction range. Finally, the calibrator integrates residue-level information with the refined prediction range to make the final prediction. By moving beyond the limitations of single-sequence-based models, OpHReda achieves a 55% improvement in F1-score compared to that of state-of-the-art methods. Extensive ablation studies demonstrate that this enhancement arises from the synergy between our tailored architecture and the augmentation mechanism. Overall, OpHReda offers a promising advancement in enzyme optimal pH prediction and holds potential for downstream applications such as enzyme engineering and rational design.

Recent advances and future trends for protein-small molecule interaction predictions with protein language models

In recent years, the application of natural language models to protein amino acid sequences, referred to as protein language models (PLMs), has demonstrated a significant potential for uncovering hidden patterns related to protein structure, function, and stability. The critical functions of proteins in biological processes often arise through interactions with small molecules; central examples are enzymes, receptors, and transporters. Understanding these interactions is particularly important for drug design, for bioengineering, and for understanding cellular metabolism. In this review, we present state-of-the-art PLMs and explore how they can be integrated with small molecule information to predict protein-small molecule interactions. We present several such prediction tasks and discuss current limitations and potential areas for improvement.

ActiMut-XGB: Predicting thermodynamic stability of point mutations for CALB with protein language model

Predicting the functional impact of single-point mutations on protein residual activity, especially after high-temperature incubation, is critical in protein engineering. We present an innovative machine learning model based on eXtreme Gradient Boosting that leverages protein sequence data to predict thermostability, circumventing the need for three-dimensional structural information. Our model integrates features from the ESM2 language model, physicochemical properties, evolutionary features, and positional features. A key advancement is the use of transfer learning with thermal stability data from various proteins, which enhances prediction accuracy and generalizability. To fine-tune and validate the model, we used experimental data from Candida antarctica lipase B single-point mutants, a widely studied enzyme in biocatalysis and industrial applications. Despite potential limitations of Gibbs free energy values in capturing all factors influencing thermostability, our model represents a significant improvement over traditional approaches, providing valuable insights for protein engineering, enzyme optimization, and therapeutic protein development.

Self-supervised learning of molecular representations from millions of tandem mass spectra using DreaMS

Characterizing biological and environmental samples at a molecular level primarily uses tandem mass spectroscopy (MS/MS), yet the interpretation of tandem mass spectra from untargeted metabolomics experiments remains a challenge. Existing computational methods for predictions from mass spectra rely on limited spectral libraries and on hard-coded human expertise. Here we introduce a transformer-based neural network pre-trained in a self-supervised way on millions of unannotated tandem mass spectra from our GNPS Experimental Mass Spectra (GeMS) dataset mined from the MassIVE GNPS repository. We show that pre-training our model to predict masked spectral peaks and chromatographic retention orders leads to the emergence of rich representations of molecular structures, which we named Deep Representations Empowering the Annotation of Mass Spectra (DreaMS). Further fine-tuning the neural network yields state-of-the-art performance across a variety of tasks. We make our new dataset and model available to the community and release the DreaMS Atlas-a molecular network of 201 million MS/MS spectra constructed using DreaMS annotations.

A multimodal deep learning framework for enzyme turnover prediction with missing modality

Accurate prediction of the turnover number (kcat), which quantifies the maximum rate of substrate conversion at an enzyme's active site, is essential for assessing catalytic efficiency and understanding biochemical reaction mechanisms. Traditional wet-lab measurements of kcat are time-consuming and resource-intensive, making deep learning (DL) methods an appealing alternative. However, existing DL models often overlook the impact of reaction products on kcat due to feedback inhibition, resulting in suboptimal performance. The multimodal nature of this kcat prediction task, involving enzymes, substrates, and products as inputs, presents additional challenges when certain modalities are unavailable during inference due to incomplete data or experimental constraints, leading to the inapplicability of existing DL models. To address these limitations, we introduce MMKcat, a novel framework employing a prior-knowledge-guided missing modality training mechanism, which treats substrates and enzyme sequences as essential inputs while considering other modalities as maskable terms. Moreover, an innovative auxiliary regularizer is incorporated to encourage the learning of informative features from various modal combinations, enabling robust predictions even with incomplete multimodal inputs. We demonstrate the superior performance of MMKcat compared to state-of-the-art methods, including DLKcat, TurNup, UniKP, EITLEM-Kinetic, DLTKcat and GELKcat, using BRENDA and SABIO-RK. Our results show significant improvements under both complete and missing modality scenarios in RMSE, R2, and SRCC metrics, with average improvements of 6.41%, 22.18%, and 8.15%, respectively. Codes are available at https://github.com/ProEcho1/MMKcat.

Computational discovery and systematic analysis of protein entangling motifs in nature: from algorithm to database

Nontrivial protein topology has the potential to revolutionize protein engineering by enabling the manipulation of proteins' stability and dynamics. However, the rarity of topological proteins in nature poses a challenge for their design, synthesis and application, primarily due to the limited number of available entangling motifs as synthetic templates. Discovering these motifs is particularly difficult, as entanglement is a subtle structural feature that is not readily discernible from protein sequences. In this study, we developed a streamlined workflow enabling efficient and accurate identification of structurally reliable and applicable entangling motifs from protein sequences. Through this workflow, we automatically curated a database of 1115 entangling protein motifs from over 100 thousand sequences in the UniProt Knowledgebase. In our database, 73.3% of C2 entangling motifs and 80.1% of C3 entangling motifs exhibited low structural similarity to known protein structures. The entangled structures in the database were categorized into different groups and their functional and biological significance were analyzed. The results were summarized in an online database accessible through a user-friendly web platform, providing researchers with an expanded toolbox of entangling motifs. This resource is poised to significantly advance the field of protein topology engineering and inspire new research directions in protein design and application.

Accurate prediction of virulence factors using pre-train protein language model and ensemble learning

BACKGROUND

As bacterial pathogens develop increasing resistance to antibiotics, strategies targeting virulence factors (VFs) have emerged as a promising and effective approach for treating bacterial infections. Existing methods mainly relied on sequence similarity, and remote homology relationships cannot be discovered by sequence analysis alone.

RESULTS

To address this limitation, we developed a protein language model and ensemble learning approach for VF identification (PLMVF). Specifically, we extracted features from protein sequences using ESM-2 and their three-dimensional (3D) structures using ESMFold. We calculated the true TM-score of the proteins based on their 3D structures and trained a TM-predictor model to predict structural similarity, thereby capturing hidden remote homology information within the sequences. Subsequently, we concatenated the sequence-level features extracted by ESM-2 with the predicted TM-score features to form a comprehensive feature set for prediction. Extensive experimental validation demonstrated that PLMVF achieved an accuracy (ACC) of 86.1%, significantly outperforming existing models across multiple evaluation metrics. This study provided an ideal tool for identifying novel targets in the development of anti-virulence therapies, offering promise for the effective prevention and control of pathogenic bacterial infections.

CONCLUSIONS

The proposed PLMVF model offers an efficient computational approach for VF identification.

A probabilistic graphical model for estimating selection coefficients of nonsynonymous variants from human population sequence data

Accurately predicting the effect of missense variants is important in discovering disease risk genes and clinical genetic diagnostics. Commonly used computational methods predict pathogenicity, which does not capture the quantitative impact on fitness in humans. We develop a method, MisFit, to estimate missense fitness effect using a graphical model. MisFit jointly models the effect at a molecular level ( d ) and a population level (selection coefficient, s ), assuming that in the same gene, missense variants with similar d have similar s . We train it by maximizing probability of observed allele counts in 236,017 individuals of European ancestry. We show that s is informative in predicting allele frequency across ancestries and consistent with the fraction of de novo mutations in sites under strong selection. Further, s outperforms previous methods in prioritizing de novo missense variants in individuals with neurodevelopmental disorders. In conclusion, MisFit accurately predicts s and yields new insights from genomic data.

Comprehensive genomic and evolutionary analysis of biofilm matrix clusters and proteins in the Vibrio genus

Vibrio cholerae pathogens cause cholera, an acute diarrheal disease resulting in significant morbidity and mortality worldwide. Biofilms in vibrios enhance their survival in natural ecosystems and facilitate transmission during cholera outbreaks. Critical components of the biofilm matrix include the Vibrio polysaccharides produced by the vps-1 and vps-2 gene clusters and the biofilm matrix proteins encoded in the rbm gene cluster, together comprising the biofilm matrix cluster. However, the biofilm matrix clusters and their evolutionary patterns in other Vibrio species remain underexplored. In this study, we systematically investigated the distribution, diversity, and evolution of biofilm matrix clusters and proteins across the Vibrio genus. Our findings reveal that these gene clusters are sporadically distributed throughout the genus, even appearing in species phylogenetically distant from Vibrio cholerae. Evolutionary analysis of the major biofilm matrix proteins RbmC and Bap1 shows that they are structurally and sequentially related, having undergone structural domain and modular alterations. Additionally, a novel loop-less Bap1 variant was identified, predominantly represented in two phylogenetically distant V. cholerae subspecies clades that share specific gene groups associated with the presence or absence of the protein. Furthermore, our analysis revealed that rbmB, a gene involved in biofilm dispersal, shares a recent common ancestor with Vibriophage tail proteins, suggesting that phages may mimic host functions to evade biofilm-associated defenses. Our study offers a foundational understanding of the diversity and evolution of biofilm matrix clusters in vibrios, laying the groundwork for future biofilm engineering through genetic modification.

IMPORTANCE

Biofilms help vibrios survive in nature and spread cholera. However, the genes that control biofilm formation in vibrios other than Vibrio cholerae are not well understood. In this study, we analyzed the biofilm matrix gene clusters and proteins across diverse Vibrio species to explore their patterns and evolution. We discovered that these genes are spread across different Vibrio species, including those not closely related to V. cholerae. We also found various forms of key biofilm proteins with different structures. Additionally, we identified genes involved in biofilm dispersal that are related to vibriophage genes, highlighting the role of phages in biofilm development. This study not only provides a foundational understanding of biofilm diversity and evolution in vibrios but also leads to new strategies for engineering biofilms through genetic modification, which is crucial for managing cholera outbreaks and improving the environmental resilience of these bacteria.

Enhancing Transthyretin Binding Affinity Prediction with a Consensus Model: Insights from the Tox24 Challenge

Transthyretin (TTR) plays a vital role in thyroid hormone transport and homeostasis in both the blood and target tissues. Interactions between exogenous compounds and TTR can disrupt the function of the endocrine system, potentially causing toxicity. In the Tox24 challenge, we leveraged the data set provided by the organizers to develop a deep learning-based consensus model, integrating sPhysNet, KANO, and GGAP-CPI for predicting TTR binding affinity. Each model utilized distinct levels of molecular information, including 2D topology, 3D geometry, and protein-ligand interactions. Our consensus model achieved favorable performance on the blind test set, yielding an RMSE of 20.8 and ranking fifth among all submissions. Following the release of the blind test set, we incorporated the leaderboard test set into our training data, further reducing the RMSE to 20.6 in an offlineretrospective study. These results demonstrate that combining three regression models across different modalities significantly enhances the predictive accuracy. Furthermore, we employ the standard deviation of the consensus model's ensemble outputs as an uncertainty estimate. Our analysis reveals that both the RMSE and interval error of predictions increase with rising uncertainty, indicating that the uncertainty can serve as a useful measure of prediction confidence. We believe that this consensus model can be a valuable resource for identifying potential TTR binders and predicting their binding affinity in silico. The source code for data preparation, model training, and prediction can be accessed at https://github.com/xiaolinpan/tox24_challenge_submission_yingkai_lab.

IL-1 Superfamily Across 400+ Species: Therapeutic Targets and Disease Implications

An important area of interest for therapeutic development is the IL-1 superfamily, a critical group of immune regulators with profound implications in a variety of disorders. This study clarifies the evolutionary patterns of IL-1 family members by thoroughly analyzing more than 400 animal species, demonstrating their ancient roots that extend back to the earliest vertebrates. Important results show that, although IL-1 ligands expanded significantly over the evolution of mammals, their corresponding receptors remained remarkably structurally conserved. Identifying both lineage-specific adaptations and evolutionarily conserved residues provides vital information for treatment design. These findings point to the possibility of two different therapeutic strategies: addressing species-specific variants may allow for more targeted interventions, whereas focusing on conserved motifs may result in broad-acting treatments. The study also identified less well-known species as useful models for comprehending early immune systems. In addition to advancing our knowledge of the function of the IL-1 family in autoimmune, inflammatory, and carcinogenic illnesses, this research lays the groundwork for the development of more potent targeted therapeutics by creating an evolutionary framework for the IL-1 family.