BACKGROUND: Protein large language models (LLM) have been used to extract representations of enzyme sequences to predict their function, which is encoded by enzyme commission (EC) numbers. However, a comprehensive comparison of different LLMs for this task is still lacking, leaving questions about their relative performance. Moreover, protein sequence alignments (e.g. BLASTp or DIAMOND) are often combined with machine learning models to assign EC numbers from homologous enzymes, thus compensating for the shortcomings of these models' predictions. In this context, LLMs and sequence alignment methods have not been extensively compared as individual predictors, raising unaddressed questions about LLMs' performance and limitations relative to the alignment methods. In this study, we set out to assess the performance of ESM2, ESM1b, and ProtBERT language models in their ability to predict EC numbers, comparing them with BLASTp, against each other and against models that rely on one-hot encodings of amino acid sequences. RESULTS: Our findings reveal that combining these LLMs with fully connected neural networks surpasses the performance of deep learning models that rely on one-hot encodings. Moreover, although BLASTp provided marginally better results overall, DL models provide results that complement BLASTp's, revealing that LLMs better predict certain EC numbers while BLASTp excels in predicting others. The ESM2 stood out as the best model among the LLMs tested, providing more accurate predictions on difficult annotation tasks and for enzymes without homologs. CONCLUSIONS: Crucially, this study demonstrates that LLMs still have to be improved to become the gold standard tool over BLASTp in mainstream enzyme annotation routines. On the other hand, LLMs can provide good predictions for more difficult-to-annotate enzymes, particularly when the identity between the query sequence and the reference database falls below 25%. Our results reinforce the claim that BLASTp and LLM models complement each other and can be more effective when used together.

Sesquiterpene synthases (STSs) catalyze the formation of a large class of plant volatiles called sesquiterpenes. While thousands of putative STS sequences from diverse plant species are available, only a small number of them have been functionally characterized. Sequence identity-based screening for desired enzymes, often used in biotechnological applications, is difficult to apply here as STS sequence similarity is strongly affected by species. This calls for more sophisticated computational methods for functionality prediction. We investigate the specificity of precursor cation formation in these elusive enzymes. By inspecting multi-product STSs, we demonstrate that STSs have a strong selectivity towards one precursor cation. We use a machine learning approach combining sequence and structure information to accurately predict precursor cation specificity for STSs across all plant species. We combine this with a co-evolutionary analysis on the wealth of uncharacterized putative STS sequences, to pinpoint residues and distant functional contacts influencing cation formation and reaction pathway selection. These structural factors can be used to predict and engineer enzymes with specific functions, as we demonstrate by predicting and characterizing two novel STSs from Citrus bergamia.

MOTIVATION: As the number of experimentally solved protein structures rises, it becomes increasingly appealing to use structural information for predictive tasks involving proteins. Due to the large variation in protein sizes, folds and topologies, an attractive approach is to embed protein structures into fixed-length vectors, which can be used in machine learning algorithms aimed at predicting and understanding functional and physical properties. Many existing embedding approaches are alignment based, which is both time-consuming and ineffective for distantly related proteins. On the other hand, library- or model-based approaches depend on a small library of fragments or require the use of a trained model, both of which may not generalize well. RESULTS: We present Geometricus, a novel and universally applicable approach to embedding proteins in a fixed-dimensional space. The approach is fast, accurate, and interpretable. Geometricus uses a set of 3D moment invariants to discretize fragments of protein structures into shape-mers, which are then counted to describe the full structure as a vector of counts. We demonstrate the applicability of this approach in various tasks, ranging from fast structure similarity search, unsupervised clustering and structure classification across proteins from different superfamilies as well as within the same family. AVAILABILITY AND IMPLEMENTATION: Python code available at https://git.wur.nl/durai001/geometricus.

MOTIVATION: Polyketide synthases (PKSs) are enzymes that generate diverse molecules of great pharmaceutical importance, including a range of clinically used antimicrobials and antitumor agents. Many polyketides are synthesized by cis-AT modular PKSs, which are organized in assembly lines, in which multiple enzymes line up in a specific order. This order is defined by specific protein-protein interactions (PPIs). The unique modular structure and catalyzing mechanism of these assembly lines makes their products predictable and also spurred combinatorial biosynthesis studies to produce novel polyketides using synthetic biology. However, predicting the interactions of PKSs, and thereby inferring the order of their assembly line, is still challenging, especially for cases in which this order is not reflected by the ordering of the PKS-encoding genes in the genome. RESULTS: Here, we introduce PKSpop, which uses a coevolution-based PPI algorithm to infer protein order in PKS assembly lines. Our method accurately predicts protein orders (93% accuracy). Additionally, we identify new residue pairs that are key in determining interaction specificity, and show that coevolution of N- and C-terminal docking domains of PKSs is significantly more predictive for PPIs than coevolution between ketosynthase and acyl carrier protein domains. AVAILABILITY AND IMPLEMENTATION: The code is available on http://www.bif.wur.nl/ (under 'Software'). SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Predicting residue-residue contacts between interacting proteins is an important problem in bioinformatics. The growing wealth of sequence data can be used to infer these contacts through correlated mutation analysis on multiple sequence alignments of interacting homologs of the proteins of interest. This requires correct identification of pairs of interacting proteins for many species, in order to avoid introducing noise (i.e. non-interacting sequences) in the analysis that will decrease predictive performance. RESULTS: We have designed Ouroboros, a novel algorithm to reduce such noise in intermolecular contact prediction. Our method iterates between weighting proteins according to how likely they are to interact based on the correlated mutations signal, and predicting correlated mutations based on the weighted sequence alignment. We show that this approach accurately discriminates between protein interaction versus non-interaction and simultaneously improves the prediction of intermolecular contact residues compared to a naive application of correlated mutation analysis. This requires no training labels concerning interactions or contacts. Furthermore, the method relaxes the assumption of one-to-one interaction of previous approaches, allowing for the study of many-to-many interactions. AVAILABILITY AND IMPLEMENTATION: Source code and test data are available at www.bif.wur.nl/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Biology is rapidly turning into an information science, thanks to enormous advances in the ability to observe the molecular properties of cells, organs and individuals. This wealth of data allows us to model molecular systems at an unprecedented level of detail and to start to understand the underlying biological mechanisms. The burgeoning field of systems biology creates a huge need for methods from machine learning, which find statistical dependencies and patterns in these large-scale datasets and use these to establish models of complex molecular systems. MLSB is a successful series of workshops that aims to provide a scientific forum for the exchange between researchers from Systems Biology and Machine Learning, to promote the exchange of ideas, interactions and collaborations between these communities.

Correlated motif covering (CMC) is the problem of finding a set of motif pairs, i.e., pairs of patterns, in the sequences of proteins from a protein-protein interaction network (PPI-network) that describe the interactions in the network as concisely as possible. In other words, a perfect solution for CMC would be a minimal set of motif pairs that describes the interaction behavior perfectly in the sense that two proteins from the network interact if and only if their sequences match a motif pair in the minimal set. In this paper, we introduce and formally define CMC and show that it is closely related to the red-blue set cover (RBSC) problem and its weighted version (WRBSC)--both well-known NP-hard problems for that there exist several algorithms with known approximation factor guarantees. We prove the hardness of approximation of CMC by providing an approximation factor preserving reduction from RBSC to CMC. We show the existence of a theoretical approximation algorithm for CMC by providing an approximation factor preserving reduction from CMC to WRBSC. We adapt the latter algorithm into a functional heuristic for CMC, called CMC-approx, and experimentally assess its performance and biological relevance. The implementation in Java can be found at >http://bioinformatics.uhasselt.be.

BACKGROUND: In addition to sequence conservation, protein multiple sequence alignments contain evolutionary signal in the form of correlated variation among amino acid positions. This signal indicates positions in the sequence that influence each other, and can be applied for the prediction of intra- or intermolecular contacts. Although various approaches exist for the detection of such correlated mutations, in general these methods utilize only pairwise correlations. Hence, they tend to conflate direct and indirect dependencies. RESULTS: We propose RMRCM, a method for Regularized Multinomial Regression in order to obtain Correlated Mutations from protein multiple sequence alignments. Importantly, our method is not restricted to pairwise (column-column) comparisons only, but takes into account the network nature of relationships between protein residues in order to predict residue-residue contacts. The use of regularization ensures that the number of predicted links between columns in the multiple sequence alignment remains limited, preventing overprediction. Using simulated datasets we analyzed the performance of our approach in predicting residue-residue contacts, and studied how it is influenced by various types of noise. For various biological datasets, validation with protein structure data indicates a good performance of the proposed algorithm for the prediction of residue-residue contacts, in comparison to previous results. RMRCM can also be applied to predict interactions (in addition to only predicting interaction sites or contact sites), as demonstrated by predicting PDZ-peptide interactions. CONCLUSIONS: A novel method is presented, which uses regularized multinomial regression in order to obtain correlated mutations from protein multiple sequence alignments. AVAILABILITY: R-code of our implementation is available via http://www.ab.wur.nl/rmrcm.

Correlated motif mining (cmm) is the problem of finding overrepresented pairs of patterns, called motifs, in sequences of interacting proteins. Algorithmic solutions for cmm thereby provide a computational method for predicting binding sites for protein interaction. In this paper, we adopt a motif-driven approach where the support of candidate motif pairs is evaluated in the network. We experimentally establish the superiority of the Chi-square-based support measure over other support measures. Furthermore, we obtain that cmm is an np-hard problem for a large class of support measures (including Chi-square) and reformulate the search for correlated motifs as a combinatorial optimization problem. We then present the generic metaheuristic slider which uses steepest ascent with a neighborhood function based on sliding motifs and employs the Chi-square-based support measure. We show that slider outperforms existing motif-driven cmm methods and scales to large protein-protein interaction networks. The slider-implementation and the data used in the experiments are available on http://bioinformatics.uhasselt.be.

Protein sequences encompass tertiary structures and contain information about specific molecular interactions, which in turn determine biological functions of proteins. Knowledge about how protein sequences define interaction specificity is largely missing, in particular for paralogous protein families with high sequence similarity, such as the plant MADS domain transcription factor family. In comparison to the situation in mammalian species, this important family of transcription regulators has expanded enormously in plant species and contains over 100 members in the model plant species Arabidopsis thaliana. Here, we provide insight into the mechanisms that determine protein-protein interaction specificity for the Arabidopsis MADS domain transcription factor family, using an integrated computational and experimental approach. Plant MADS proteins have highly similar amino acid sequences, but their dimerization patterns vary substantially. Our computational analysis uncovered small sequence regions that explain observed differences in dimerization patterns with reasonable accuracy. Furthermore, we show the usefulness of the method for prediction of MADS domain transcription factor interaction networks in other plant species. Introduction of mutations in the predicted interaction motifs demonstrated that single amino acid mutations can have a large effect and lead to loss or gain of specific interactions. In addition, various performed bioinformatics analyses shed light on the way evolution has shaped MADS domain transcription factor interaction specificity. Identified protein-protein interaction motifs appeared to be strongly conserved among orthologs, indicating their evolutionary importance. We also provide evidence that mutations in these motifs can be a source for sub- or neo-functionalization. The analyses presented here take us a step forward in understanding protein-protein interactions and the interplay between protein sequences and network evolution.

Correlated motif mining (CMM) is the problem to find overrepresented pairs of patterns, called motif pairs, in interacting protein sequences. Algorithmic solutions for CMM thereby provide a computational method for predicting binding sites for protein interaction. In this paper, we adopt a motif-driven approach where the support of candidate motif pairs is evaluated in the network. We experimentally establish the superiority of the Chi-square-based support measure over other support measures. Furthermore, we obtain that CMM is an NP-hard problem for a large class of support measures (including Chi-square) and reformulate the search for correlated motifs as a combinatorial optimization problem. We then present the method SLIDER which uses local search with a neighborhood function based on sliding motifs and employs the Chi-square-based support measure. We show that SLIDER outperforms existing motif-driven CMM methods and scales to large protein-protein interaction networks.

MOTIVATION: Recent research underlines the importance of finegrained knowledge on protein localization. In particular, subcompartmental localization in the Golgi apparatus is important, for example, for the order of reactions performed in glycosylation pathways or the sorting functions of SNAREs, but is currently poorly understood. RESULTS: We assemble a dataset of type II transmembrane proteins with experimentally determined sub-Golgi localizations and use this information to develop a predictor based on the transmembrane domain of these proteins, making use of a dedicated proteinstructure based kernel in an SVM. Various applications demonstrate the power of our approach. In particular, comparison with a large set of glycan structures illustrates the applicability of our predictions on a 'glycomic' scale and demonstrates a significant correlation between sub-Golgi localization and the ordering of different steps in glycan biosynthesis. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Transcription factor interactions are the cornerstone of combinatorial control, which is a crucial aspect of the gene regulatory system. Understanding and predicting transcription factor interactions based on their sequence alone is difficult since they are often part of families of factors sharing high sequence identity. Given the scarcity of experimental data on interactions compared to available sequence data, however, it would be most useful to have accurate methods for the prediction of such interactions. RESULTS: We present a method consisting of a Random Forest-based feature-selection procedure that selects relevant motifs out of a set found using a correlated motif search algorithm. Prediction accuracy for several transcription factor families (bZIP, MADS, homeobox and forkhead) reaches 60-90%. In addition, we identified those parts of the sequence that are important for the interaction specificity, and show that these are in agreement with available data. We also used the predictors to perform genome-wide scans for interaction partners and recovered both known and putative new interaction partners.

MOTIVATION: Interfacial water, which plays an important role in mediating biomolecular interactions, has been neglected in the modelling of biomolecular complexes. METHODS: We present a solvated docking approach that explicitly accounts for the presence of water in protein-protein complexes. Our solvated docking protocol is based on the concept of the first encounter complex in which a water layer is present in-between the molecules. It mimics the pathway from this initial complex towards the final assembly in which most waters have been expelled from the interface. Docking is performed from solvated biomolecules and waters are removed in a biased Monte Carlo procedure based on water-mediated contact propensities obtained from an analysis of high-resolution crystal structures. RESULTS: We demonstrate the feasibility of this approach for protein-protein complexes representing both 'wet' and 'dry' interfaces. Solvated docking leads to improvements both in quality and scoring. Water molecules are recovered that closely match the ones in the crystal structures. AVAILABILTY: Solvated docking will be made available in the future release of HADDOCK version 2.0 (http://www.nmr.chem.uu.nl/haddock).