We present FeatureDCA, a statistical framework for protein sequence modeling and generation that extends Direct Coupling Analysis (DCA) with biologically meaningful conditioning. The method can leverage different kinds of information, such as phylogeny, optimal growth temperature, enzymatic activity or, as in the case presented here, principal components derived from multiple sequence alignments, and use it to improve the learning process and consequently efficiently condition the generative process. FeatureDCA allows sampling to be guided toward specific regions of sequence space while maintaining the efficiency and interpretability of Potts-based inference. Across multiple protein families, our autoregressive implementation of FeatureDCA matches or surpasses the generative accuracy of established models in reproducing higher-order sequence statistics while preserving substantial sequence diversity. Structural validation with AlphaFold and ESMFold confirms that generated sequences adopt folds consistent with their intended wild-type targets. In a detailed case study of the Response Regulator family (PF00072), which comprises distinct structural subclasses linked to different DNA-binding domains, FeatureDCA accurately reproduces class-specific architectures when conditioned on subtype-specific principal components, highlighting its potential for fine-grained structural control. Predictions of experimental deep mutational scanning data show accuracy comparable to that of unconditioned autoregressive Potts models, indicating that FeatureDCA also captures local functional constraints. These results position FeatureDCA as a flexible and transparent approach for targeted sequence generation, bridging statistical fidelity, structural realism, and interpretability in protein design.

Proteins are involved in nearly all cellular functions, encompassing roles in transport, signaling, enzymatic activity, and more. Their functionalities crucially depend on their complex three-dimensional arrangement. For this reason, being able to predict their structure from the amino acid sequence has been and still is a phenomenal computational challenge that the introduction of AlphaFold solved with unprecedented accuracy. However, the inherent complexity of AlphaFold's architectures makes it challenging to understand the rules that ultimately shape the protein's predicted structure. This study investigates a single-layer unsupervised model based on the attention mechanism. More precisely, we explore a Direct Coupling Analysis (DCA) method that mimics the attention mechanism of several popular Transformer architectures, such as AlphaFold itself. The model's parameters, notably fewer than those in standard DCA-based algorithms, can be directly used for extracting structural determinants such as the contact map of the protein family under study. Additionally, the functional form of the energy function of the model enables us to deploy a multi-family learning strategy, allowing us to effectively integrate information across multiple protein families, whereas standard DCA algorithms are typically limited to single protein families. Finally, we implemented a generative version of the model using an autoregressive architecture, capable of efficiently generating new proteins in silico.

The potential and promise of deep learning systems to provide an independent assessment and relieve radiologists' burden in screening mammography have been recognized in several studies. However, the low cancer prevalence, the need to process high-resolution images, and the need to combine information from multiple views and scales still pose technical challenges. Multi-view architectures that combine information from the four mammographic views to produce an exam-level classification score are a promising approach to the automated processing of screening mammography. However, training such architectures from exam-level labels, without relying on pixel-level supervision, requires very large datasets and may result in suboptimal accuracy. Emerging architectures such as Visual Transformers (ViT) and graph-based architectures can potentially integrate ipsi-lateral and contra-lateral breast views better than traditional convolutional neural networks, thanks to their stronger ability of modeling long-range dependencies. In this paper, we extensively evaluate novel transformer-based and graph-based architectures against state-of-the-art multi-view convolutional neural networks, trained in a weakly-supervised setting on a middle-scale dataset, both in terms of performance and interpretability. Extensive experiments on the CSAW dataset suggest that, while transformer-based architecture outperform other architectures, different inductive biases lead to complementary strengths and weaknesses, as each architecture is sensitive to different signs and mammographic features. Hence, an ensemble of different architectures should be preferred over a winner-takes-all approach to achieve more accurate and robust results. Overall, the findings highlight the potential of a wide range of multi-view architectures for breast cancer classification, even in datasets of relatively modest size, although the detection of small lesions remains challenging without pixel-wise supervision or ad-hoc networks.

Adeno-associated viruses 2 (AAV2) are minute viruses renowned for their capacity to infect human cells and akin organisms. They have recently emerged as prominent candidates in the field of gene therapy, primarily attributed to their inherent non-pathogenic nature in humans and the safety associated with their manipulation. The efficacy of AAV2 as gene therapy vectors hinges on their ability to infiltrate host cells, a phenomenon reliant on their competence to construct a capsid capable of breaching the nucleus of the target cell. To enhance their infection potential, researchers have extensively scrutinized various combinatorial libraries by introducing mutations into the capsid, aiming to boost their effectiveness. The emergence of high-throughput experimental techniques, like deep mutational scanning (DMS), has made it feasible to experimentally assess the fitness of these libraries for their intended purpose. Notably, machine learning is starting to demonstrate its potential in addressing predictions within the mutational landscape from sequence data. In this context, we introduce a biophysically-inspired model designed to predict the viability of genetic variants in DMS experiments. This model is tailored to a specific segment of the CAP region within AAV2's capsid protein. To evaluate its effectiveness, we conduct model training with diverse datasets, each tailored to explore different aspects of the mutational landscape influenced by the selection process. Our assessment of the biophysical model centers on two primary objectives: (i) providing quantitative forecasts for the log-selectivity of variants and (ii) deploying it as a binary classifier to categorize sequences into viable and non-viable classes.

Exquisite binding specificity is essential for many protein functions but is difficult to engineer. Many biotechnological or biomedical applications require the discrimination of very similar ligands, which poses the challenge of designing protein sequences with highly specific binding profiles. Experimental methods for generating specific binders rely on in vitro selection, which is limited in terms of library size and control over specificity profiles. Additional control was recently demonstrated through high-throughput sequencing and downstream computational analysis. Here we follow such an approach to demonstrate the design of specific antibodies beyond those probed experimentally. We do so in a context where very similar epitopes need to be discriminated, and where these epitopes cannot be experimentally dissociated from other epitopes present in the selection. Our approach involves the identification of different binding modes, each associated with a particular ligand against which the antibodies are either selected or not. Using data from phage display experiments, we show that the model successfully disentangles these modes, even when they are associated with chemically very similar ligands. Additionally, we demonstrate and validate experimentally the computational design of antibodies with customized specificity profiles, either with specific high affinity for a particular target ligand, or with cross-specificity for multiple target ligands. Overall, our results showcase the potential of leveraging a biophysical model learned from selections against multiple ligands to design proteins with tailored specificity, with applications to protein engineering extending beyond the design of antibodies.

The design of proteins with specific tasks is a major challenge in molecular biology with important diagnostic and therapeutic applications. High-throughput screening methods have been developed to systematically evaluate protein activity, but only a small fraction of possible protein variants can be tested using these techniques. Computational models that explore the sequence space in-silico to identify the fittest molecules for a given function are needed to overcome this limitation. In this article, we propose AnnealDCA, a machine-learning framework to learn the protein fitness landscape from sequencing data derived from a broad range of experiments that use selection and sequencing to quantify protein activity. We demonstrate the effectiveness of our method by applying it to antibody Rep-Seq data of immunized mice and screening experiments, assessing the quality of the fitness landscape reconstructions. Our method can be applied to several experimental cases where a population of protein variants undergoes various rounds of selection and sequencing, without relying on the computation of variants enrichment ratios, and thus can be used even in cases of disjoint sequence samples.

MOTIVATION: Being able to artificially design novel proteins of desired function is pivotal in many biological and biomedical applications. Generative statistical modeling has recently emerged as a new paradigm for designing amino acid sequences, including in particular models and embedding methods borrowed from natural language processing (NLP). However, most approaches target single proteins or protein domains, and do not take into account any functional specificity or interaction with the context. To extend beyond current computational strategies, we develop a method for generating protein domain sequences intended to interact with another protein domain. Using data from natural multidomain proteins, we cast the problem as a translation problem from a given interactor domain to the new domain to be generated, i.e. we generate artificial partner sequences conditional on an input sequence. We also show in an example that the same procedure can be applied to interactions between distinct proteins. RESULTS: Evaluating our model's quality using diverse metrics, in part related to distinct biological questions, we show that our method outperforms state-of-the-art shallow autoregressive strategies. We also explore the possibility of fine-tuning pretrained large language models for the same task and of using Alphafold 2 for assessing the quality of sampled sequences. AVAILABILITY AND IMPLEMENTATION: Data and code on https://github.com/barthelemymp/Domain2DomainProteinTranslation.

SUMMARY: DCAlign is a new alignment method able to cope with the conservation and the co-evolution signals that characterize the columns of multiple sequence alignments of homologous sequences. However, the pre-processing steps required to align a candidate sequence are computationally demanding. We show in v1.0 how to dramatically reduce the overall computing time by including an empirical prior over an informative set of variables mirroring the presence of insertions and deletions. AVAILABILITY AND IMPLEMENTATION: DCAlign v1.0 is implemented in Julia and it is fully available at https://github.com/infernet-h2020/DCAlign.

SUMMARY: Topology determination is one of the most important intermediate steps toward building the atomic structure of proteins from their medium-resolution cryo-electron microscopy (cryo-EM) map. The main goal in the topology determination is to identify correct matches (i.e. assignment and direction) between secondary structure elements (SSEs) (α-helices and β-sheets) detected in a protein sequence and cryo-EM density map. Despite many recent advances in molecular biology technologies, the problem remains a challenging issue. To overcome the problem, this article proposes a linear programming-based topology determination (LPTD) method to solve the secondary structure topology problem in three-dimensional geometrical space. Through modeling of the protein's sequence with the aid of extracting highly reliable features and a distance-based scoring function, the secondary structure matching problem is transformed into a complete weighted bipartite graph matching problem. Subsequently, an algorithm based on linear programming is developed as a decision-making strategy to extract the true topology (native topology) between all possible topologies. The proposed automatic framework is verified using 12 experimental and 15 simulated α-β proteins. Results demonstrate that LPTD is highly efficient and extremely fast in such a way that for 77% of cases in the dataset, the native topology has been detected in the first rank topology in <2 s. Besides, this method is able to successfully handle large complex proteins with as many as 65 SSEs. Such a large number of SSEs have never been solved with current tools/methods. AVAILABILITY AND IMPLEMENTATION: The LPTD package (source code and data) is publicly available at https://github.com/B-Behkamal/LPTD. Moreover, two test samples as well as the instruction of utilizing the graphical user interface have been provided in the shared readme file. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: Boltzmann machines are energy-based models that have been shown to provide an accurate statistical description of domains of evolutionary-related protein and RNA families. They are parametrized in terms of local biases accounting for residue conservation, and pairwise terms to model epistatic coevolution between residues. From the model parameters, it is possible to extract an accurate prediction of the three-dimensional contact map of the target domain. More recently, the accuracy of these models has been also assessed in terms of their ability in predicting mutational effects and generating in silico functional sequences. RESULTS: Our adaptive implementation of Boltzmann machine learning, adabmDCA, can be generally applied to both protein and RNA families and accomplishes several learning set-ups, depending on the complexity of the input data and on the user requirements. The code is fully available at https://github.com/anna-pa-m/adabmDCA . As an example, we have performed the learning of three Boltzmann machines modeling the Kunitz and Beta-lactamase2 protein domains and TPP-riboswitch RNA domain. CONCLUSIONS: The models learned by adabmDCA are comparable to those obtained by state-of-the-art techniques for this task, in terms of the quality of the inferred contact map as well as of the synthetically generated sequences. In addition, the code implements both equilibrium and out-of-equilibrium learning, which allows for an accurate and lossless training when the equilibrium one is prohibitive in terms of computational time, and allows for pruning irrelevant parameters using an information-based criterion.

It is well known that, in order to preserve its structure and function, a protein cannot change its sequence at random, but only by mutations occurring preferentially at specific locations. We here investigate quantitatively the amount of variability that is allowed in protein sequence evolution, by computing the intrinsic dimension (ID) of the sequences belonging to a selection of protein families. The ID is a measure of the number of independent directions that evolution can take starting from a given sequence. We find that the ID is practically constant for sequences belonging to the same family, and moreover it is very similar in different families, with values ranging between 6 and 12. These values are significantly smaller than the raw number of amino acids, confirming the importance of correlations between mutations in different sites. However, we demonstrate that correlations are not sufficient to explain the small value of the ID we observe in protein families. Indeed, we show that the ID of a set of protein sequences generated by maximum entropy models, an approach in which correlations are accounted for, is typically significantly larger than the value observed in natural protein families. We further prove that a critical factor to reproduce the natural ID is to take into consideration the phylogeny of sequences.

The immune system has developed a number of distinct complex mechanisms to shape and control the antibody repertoire. One of these mechanisms, the affinity maturation process, works in an evolutionary-like fashion: after binding to a foreign molecule, the antibody-producing B-cells exhibit a high-frequency mutation rate in the genome region that codes for the antibody active site. Eventually, cells that produce antibodies with higher affinity for their cognate antigen are selected and clonally expanded. Here, we propose a new statistical approach based on maximum entropy modeling in which a scoring function related to the binding affinity of antibodies against a specific antigen is inferred from a sample of sequences of the immune repertoire of an individual. We use our inference strategy to infer a statistical model on a data set obtained by sequencing a fairly large portion of the immune repertoire of an HIV-1 infected patient. The Pearson correlation coefficient between our scoring function and the IC50 neutralization titer measured on 30 different antibodies of known sequence is as high as 0.77 (p-value 10-6), outperforming other sequence- and structure-based models.

Correlation patterns in multiple sequence alignments of homologous proteins can be exploited to infer information on the three-dimensional structure of their members. The typical pipeline to address this task, which we in this paper refer to as the three dimensions of contact prediction, is to (i) filter and align the raw sequence data representing the evolutionarily related proteins; (ii) choose a predictive model to describe a sequence alignment; (iii) infer the model parameters and interpret them in terms of structural properties, such as an accurate contact map. We show here that all three dimensions are important for overall prediction success. In particular, we show that it is possible to improve significantly along the second dimension by going beyond the pair-wise Potts models from statistical physics, which have hitherto been the focus of the field. These (simple) extensions are motivated by multiple sequence alignments often containing long stretches of gaps which, as a data feature, would be rather untypical for independent samples drawn from a Potts model. Using a large test set of proteins we show that the combined improvements along the three dimensions are as large as any reported to date.

We present a powerful experimental-computational technology for inferring network models that predict the response of cells to perturbations, and that may be useful in the design of combinatorial therapy against cancer. The experiments are systematic series of perturbations of cancer cell lines by targeted drugs, singly or in combination. The response to perturbation is quantified in terms of relative changes in the measured levels of proteins, phospho-proteins and cellular phenotypes such as viability. Computational network models are derived de novo, i.e., without prior knowledge of signaling pathways, and are based on simple non-linear differential equations. The prohibitively large solution space of all possible network models is explored efficiently using a probabilistic algorithm, Belief Propagation (BP), which is three orders of magnitude faster than standard Monte Carlo methods. Explicit executable models are derived for a set of perturbation experiments in SKMEL-133 melanoma cell lines, which are resistant to the therapeutically important inhibitor of RAF kinase. The resulting network models reproduce and extend known pathway biology. They empower potential discoveries of new molecular interactions and predict efficacious novel drug perturbations, such as the inhibition of PLK1, which is verified experimentally. This technology is suitable for application to larger systems in diverse areas of molecular biology.

BACKGROUND: Transcriptional gene regulation is one of the most important mechanisms in controlling many essential cellular processes, including cell development, cell-cycle control, and the cellular response to variations in environmental conditions. Genes are regulated by transcription factors and other genes/proteins via a complex interconnection network. Such regulatory links may be predicted using microarray expression data, but most regulation models suppose transcription factor independence, which leads to spurious links when many genes have highly correlated expression levels. RESULTS: We propose a new algorithm to infer combinatorial control networks from gene-expression data. Based on a simple model of combinatorial gene regulation, it includes a message-passing approach which avoids explicit sampling over putative gene-regulatory networks. This algorithm is shown to recover the structure of a simple artificial cell-cycle network model for baker's yeast. It is then applied to a large-scale yeast gene expression dataset in order to identify combinatorial regulations, and to a data set of direct medical interest, namely the Pleiotropic Drug Resistance (PDR) network. CONCLUSIONS: The algorithm we designed is able to recover biologically meaningful interactions, as shown by recent experimental results 1. Moreover, new cases of combinatorial control are predicted, showing how simple models taking this phenomenon into account can lead to informative predictions and allow to extract more putative regulatory interactions from microarray databases.

BACKGROUND: Cellular metabolism is one of the most investigated system of biological interactions. While the topological nature of individual reactions and pathways in the network is quite well understood there is still a lack of comprehension regarding the global functional behavior of the system. In the last few years flux-balance analysis (FBA) has been the most successful and widely used technique for studying metabolism at system level. This method strongly relies on the hypothesis that the organism maximizes an objective function. However only under very specific biological conditions (e.g. maximization of biomass for E. coli in reach nutrient medium) the cell seems to obey such optimization law. A more refined analysis not assuming extremization remains an elusive task for large metabolic systems due to algorithmic limitations. RESULTS: In this work we propose a novel algorithmic strategy that provides an efficient characterization of the whole set of stable fluxes compatible with the metabolic constraints. Using a technique derived from the fields of statistical physics and information theory we designed a message-passing algorithm to estimate the size of the affine space containing all possible steady-state flux distributions of metabolic networks. The algorithm, based on the well known Bethe approximation, can be used to approximately compute the volume of a non full-dimensional convex polytope in high dimensions. We first compare the accuracy of the predictions with an exact algorithm on small random metabolic networks. We also verify that the predictions of the algorithm match closely those of Monte Carlo based methods in the case of the Red Blood Cell metabolic network. Then we test the effect of gene knock-outs on the size of the solution space in the case of E. coli central metabolism. Finally we analyze the statistical properties of the average fluxes of the reactions in the E. coli metabolic network. CONCLUSION: We propose a novel efficient distributed algorithmic strategy to estimate the size and shape of the affine space of a non full-dimensional convex polytope in high dimensions. The method is shown to obtain, quantitatively and qualitatively compatible results with the ones of standard algorithms (where this comparison is possible) being still efficient on the analysis of large biological systems, where exact deterministic methods experience an explosion in algorithmic time. The algorithm we propose can be considered as an alternative to Monte Carlo sampling methods.

MOTIVATION: In the last few years, a growing interest in biology has been shifting toward the problem of optimal information extraction from the huge amount of data generated via large-scale and high-throughput techniques. One of the most relevant issues has recently emerged that of correctly and reliably predicting the functions of a given protein with that of functions exploiting information coming from the whole network of proteins physically interacting with the functionally undetermined one. In the present work, we will refer to an 'observed' protein as the one present in the protein-protein interaction networks published in the literature. METHODS: The method proposed in this paper is based on a message passing algorithm known as Belief Propagation, which accepts the network of protein's physical interactions and a catalog of known protein's functions as input, and returns the probabilities for each unclassified protein of having one chosen function. The implementation of the algorithm allows for fast online analysis, and can easily be generalized into more complex graph topologies taking into account hypergraphs, i.e. complexes of more than two interacting proteins. RESULTS: Benchmarks of our method are the two Saccharomyces cerevisiae protein-protein interaction networks and the Database of Interacting Proteins. The validity of our approach is successfully tested against other available techniques. CONTACT: [email protected] SUPPLEMENTARY INFORMATION: http://isiosf.isi.it/~pagnani