Protein language models trained on multiple sequence alignments learn phylogenetic relationships

SSEmb: A joint embedding of protein sequence and structure enables robust variant effect predictions

The ability to predict how amino acid changes affect proteins has a wide range of applications including in disease variant classification and protein engineering. Many existing methods focus on learning from patterns found in either protein sequences or protein structures. Here, we present a method for integrating information from sequence and structure in a single model that we term SSEmb (Sequence Structure Embedding). SSEmb combines a graph representation for the protein structure with a transformer model for processing multiple sequence alignments. We show that by integrating both types of information we obtain a variant effect prediction model that is robust when sequence information is scarce. We also show that SSEmb learns embeddings of the sequence and structure that are useful for other downstream tasks such as to predict protein-protein binding sites. We envisage that SSEmb may be useful both for variant effect predictions and as a representation for learning to predict protein properties that depend on sequence and structure.

Computational tools to predict context-specific protein complexes

Interactions between thousands of proteins define cells' protein-protein interaction (PPI) network. Some of these interactions lead to the formation of protein complexes. It is challenging to identify a protein complex in a haystack of protein-protein interactions, and it is even more difficult to predict all protein complexes of the complexome. Simulations and machine learning approaches try to crack these problems by looking at the PPI network or predicted protein structures. Clustering of PPI networks led to the first protein complex predictions, while most recently, atomistic models of protein complexes and deep-learning-based structure prediction methods have also emerged. The simulation of PPI level interactions even enables the quantitative prediction of protein complexes. These methods, the required data sources, and their potential future developments are discussed in this review.

Impact of phylogeny on the inference of functional sectors from protein sequence data

Statistical analysis of multiple sequence alignments of homologous proteins has revealed groups of coevolving amino acids called sectors. These groups of amino-acid sites feature collective correlations in their amino-acid usage, and they are associated to functional properties. Modeling showed that nonlinear selection on an additive functional trait of a protein is generically expected to give rise to a functional sector. These modeling results motivated a principled method, called ICOD, which is designed to identify functional sectors, as well as mutational effects, from sequence data. However, a challenge for all methods aiming to identify sectors from multiple sequence alignments is that correlations in amino-acid usage can also arise from the mere fact that homologous sequences share common ancestry, i.e. from phylogeny. Here, we generate controlled synthetic data from a minimal model comprising both phylogeny and functional sectors. We use this data to dissect the impact of phylogeny on sector identification and on mutational effect inference by different methods. We find that ICOD is most robust to phylogeny, but that conservation is also quite robust. Next, we consider natural multiple sequence alignments of protein families for which deep mutational scan experimental data is available. We show that in this natural data, conservation and ICOD best identify sites with strong functional roles, in agreement with our results on synthetic data. Importantly, these two methods have different premises, since they respectively focus on conservation and on correlations. Thus, their joint use can reveal complementary information.

Understanding the Natural Language of DNA using Encoder-Decoder Foundation Models with Byte-level Precision

This paper presents the Ensemble Nucleotide Byte-level Encoder-Decoder (ENBED) foundation model, analyzing DNA sequences at byte-level precision with an encoder-decoder Transformer architecture. ENBED uses a sub-quadratic implementation of attention to develop an efficient model capable of sequence-to-sequence transformations, generalizing previous genomic models with encoder-only or decoder-only architectures. We use Masked Language Modeling to pre-train the foundation model using reference genome sequences and apply it in the following downstream tasks: (1) identification of enhancers, promotors and splice sites, (2) recognition of sequences containing base call mismatches and insertion/deletion errors, an advantage over tokenization schemes involving multiple base pairs, which lose the ability to analyze with byte-level precision, (3) identification of biological function annotations of genomic sequences, and (4) generating mutations of the Influenza virus using the encoder-decoder architecture and validating them against real-world observations. In each of these tasks, we demonstrate significant improvement as compared to the existing state-of-the-art results.

Understanding the natural language of DNA using encoder-decoder foundation models with byte-level precision

Summary

This article presents the Ensemble Nucleotide Byte-level Encoder-Decoder (ENBED) foundation model, analyzing DNA sequences at byte-level precision with an encoder-decoder Transformer architecture. ENBED uses a subquadratic implementation of attention to develop an efficient model capable of sequence-to-sequence transformations, generalizing previous genomic models with encoder-only or decoder-only architectures. We use Masked Language Modeling to pretrain the foundation model using reference genome sequences and apply it in the following downstream tasks: (i) identification of enhancers, promotors, and splice sites, (ii) recognition of sequences containing base call mismatches and insertion/deletion errors, an advantage over tokenization schemes involving multiple base pairs, which lose the ability to analyze with byte-level precision, (iii) identification of biological function annotations of genomic sequences, and (iv) generating mutations of the Influenza virus using the encoder-decoder architecture and validating them against real-world observations. In each of these tasks, we demonstrate significant improvement as compared to the existing state-of-the-art results.

Availability and implementation

The source code used to develop and fine-tune the foundation model has been released on Github (https://github.itap.purdue.edu/Clan-labs/ENBED).

DR-BERT: A protein language model to annotate disordered regions

Despite their lack of a rigid structure, intrinsically disordered regions (IDRs) in proteins play important roles in cellular functions, including mediating protein-protein interactions. Therefore, it is important to computationally annotate IDRs with high accuracy. In this study, we present Disordered Region prediction using Bidirectional Encoder Representations from Transformers (DR-BERT), a compact protein language model. Unlike most popular tools, DR-BERT is pretrained on unannotated proteins and trained to predict IDRs without relying on explicit evolutionary or biophysical data. Despite this, DR-BERT demonstrates significant improvement over existing methods on the Critical Assessment of protein Intrinsic Disorder (CAID) evaluation dataset and outperforms competitors on two out of four test cases in the CAID 2 dataset, while maintaining competitiveness in the others. This performance is due to the information learned during pretraining and DR-BERT's ability to use contextual information.

Machine learning in biological physics: From biomolecular prediction to design

Machine learning has been proposed as an alternative to theoretical modeling when dealing with complex problems in biological physics. However, in this perspective, we argue that a more successful approach is a proper combination of these two methodologies. We discuss how ideas coming from physical modeling neuronal processing led to early formulations of computational neural networks, e.g., Hopfield networks. We then show how modern learning approaches like Potts models, Boltzmann machines, and the transformer architecture are related to each other, specifically, through a shared energy representation. We summarize recent efforts to establish these connections and provide examples on how each of these formulations integrating physical modeling and machine learning have been successful in tackling recent problems in biomolecular structure, dynamics, function, evolution, and design. Instances include protein structure prediction; improvement in computational complexity and accuracy of molecular dynamics simulations; better inference of the effects of mutations in proteins leading to improved evolutionary modeling and finally how machine learning is revolutionizing protein engineering and design. Going beyond naturally existing protein sequences, a connection to protein design is discussed where synthetic sequences are able to fold to naturally occurring motifs driven by a model rooted in physical principles. We show that this model is "learnable" and propose its future use in the generation of unique sequences that can fold into a target structure.

Pairing interacting protein sequences using masked language modeling

Predicting which proteins interact together from amino acid sequences is an important task. We develop a method to pair interacting protein sequences which leverages the power of protein language models trained on multiple sequence alignments (MSAs), such as MSA Transformer and the EvoFormer module of AlphaFold. We formulate the problem of pairing interacting partners among the paralogs of two protein families in a differentiable way. We introduce a method called Differentiable Pairing using Alignment-based Language Models (DiffPALM) that solves it by exploiting the ability of MSA Transformer to fill in masked amino acids in multiple sequence alignments using the surrounding context. MSA Transformer encodes coevolution between functionally or structurally coupled amino acids within protein chains. It also captures inter-chain coevolution, despite being trained on single-chain data. Relying on MSA Transformer without fine-tuning, DiffPALM outperforms existing coevolution-based pairing methods on difficult benchmarks of shallow multiple sequence alignments extracted from ubiquitous prokaryotic protein datasets. It also outperforms an alternative method based on a state-of-the-art protein language model trained on single sequences. Paired alignments of interacting protein sequences are a crucial ingredient of supervised deep learning methods to predict the three-dimensional structure of protein complexes. Starting from sequences paired by DiffPALM substantially improves the structure prediction of some eukaryotic protein complexes by AlphaFold-Multimer. It also achieves competitive performance with using orthology-based pairing.

Applications of machine learning in phylogenetics

Machine learning has increasingly been applied to a wide range of questions in phylogenetic inference. Supervised machine learning approaches that rely on simulated training data have been used to infer tree topologies and branch lengths, to select substitution models, and to perform downstream inferences of introgression and diversification. Here, we review how researchers have used several promising machine learning approaches to make phylogenetic inferences. Despite the promise of these methods, several barriers prevent supervised machine learning from reaching its full potential in phylogenetics. We discuss these barriers and potential paths forward. In the future, we expect that the application of careful network designs and data encodings will allow supervised machine learning to accommodate the complex processes that continue to confound traditional phylogenetic methods.

Deep learning in modeling protein complex structures: From contact prediction to end-to-end approaches

Protein-protein interactions play crucial roles in many biological processes. Traditionally, protein complex structures are normally built by protein-protein docking. With the rapid development of artificial intelligence and its great success in monomer protein structure prediction, deep learning has widely been applied to modeling protein-protein complex structures through inter-protein contact prediction and end-to-end approaches in the past few years. This article reviews the recent advances of deep-learning-based approaches in modeling protein-protein complex structures as well as their advantages and limitations. Challenges and possible future directions are also briefly discussed in applying deep learning for the prediction of protein complex structures.

Computational drug development for membrane protein targets

The application of computational biology in drug development for membrane protein targets has experienced a boost from recent developments in deep learning-driven structure prediction, increased speed and resolution of structure elucidation, machine learning structure-based design and the evaluation of big data. Recent protein structure predictions based on machine learning tools have delivered surprisingly reliable results for water-soluble and membrane proteins but have limitations for development of drugs that target membrane proteins. Structural transitions of membrane proteins have a central role during transmembrane signaling and are often influenced by therapeutic compounds. Resolving the structural and functional basis of dynamic transmembrane signaling networks, especially within the native membrane or cellular environment, remains a central challenge for drug development. Tackling this challenge will require an interplay between experimental and computational tools, such as super-resolution optical microscopy for quantification of the molecular interactions of cellular signaling networks and their modulation by potential drugs, cryo-electron microscopy for determination of the structural transitions of proteins in native cell membranes and entire cells, and computational tools for data analysis and prediction of the structure and function of cellular signaling networks, as well as generation of promising drug candidates.

High-throughput prediction of enzyme promiscuity based on substrate-product pairs

The screening of enzymes for catalyzing specific substrate-product pairs is often constrained in the realms of metabolic engineering and synthetic biology. Existing tools based on substrate and reaction similarity predominantly rely on prior knowledge, demonstrating limited extrapolative capabilities and an inability to incorporate custom candidate-enzyme libraries. Addressing these limitations, we have developed the Substrate-product Pair-based Enzyme Promiscuity Prediction (SPEPP) model. This innovative approach utilizes transfer learning and transformer architecture to predict enzyme promiscuity, thereby elucidating the intricate interplay between enzymes and substrate-product pairs. SPEPP exhibited robust predictive ability, eliminating the need for prior knowledge of reactions and allowing users to define their own candidate-enzyme libraries. It can be seamlessly integrated into various applications, including metabolic engineering, de novo pathway design, and hazardous material degradation. To better assist metabolic engineers in designing and refining biochemical pathways, particularly those without programming skills, we also designed EnzyPick, an easy-to-use web server for enzyme screening based on SPEPP. EnzyPick is accessible at http://www.biosynther.com/enzypick/.

Protein Fitness Prediction Is Impacted by the Interplay of Language Models, Ensemble Learning, and Sampling Methods

Advances in machine learning (ML) and the availability of protein sequences via high-throughput sequencing techniques have transformed the ability to design novel diagnostic and therapeutic proteins. ML allows protein engineers to capture complex trends hidden within protein sequences that would otherwise be difficult to identify in the context of the immense and rugged protein fitness landscape. Despite this potential, there persists a need for guidance during the training and evaluation of ML methods over sequencing data. Two key challenges for training discriminative models and evaluating their performance include handling severely imbalanced datasets (e.g., few high-fitness proteins among an abundance of non-functional proteins) and selecting appropriate protein sequence representations (numerical encodings). Here, we present a framework for applying ML over assay-labeled datasets to elucidate the capacity of sampling techniques and protein encoding methods to improve binding affinity and thermal stability prediction tasks. For protein sequence representations, we incorporate two widely used methods (One-Hot encoding and physiochemical encoding) and two language-based methods (next-token prediction, UniRep; masked-token prediction, ESM). Elaboration on performance is provided over protein fitness, protein size, and sampling techniques. In addition, an ensemble of protein representation methods is generated to discover the contribution of distinct representations and improve the final prediction score. We then implement multiple criteria decision analysis (MCDA; TOPSIS with entropy weighting), using multiple metrics well-suited for imbalanced data, to ensure statistical rigor in ranking our methods. Within the context of these datasets, the synthetic minority oversampling technique (SMOTE) outperformed undersampling while encoding sequences with One-Hot, UniRep, and ESM representations. Moreover, ensemble learning increased the predictive performance of the affinity-based dataset by 4% compared to the best single-encoding candidate (F1-score = 97%), while ESM alone was rigorous enough in stability prediction (F1-score = 92%).

Generative power of a protein language model trained on multiple sequence alignments

Computational models starting from large ensembles of evolutionarily related protein sequences capture a representation of protein families and learn constraints associated to protein structure and function. They thus open the possibility for generating novel sequences belonging to protein families. Protein language models trained on multiple sequence alignments, such as MSA Transformer, are highly attractive candidates to this end. We propose and test an iterative method that directly employs the masked language modeling objective to generate sequences using MSA Transformer. We demonstrate that the resulting sequences score as well as natural sequences, for homology, coevolution, and structure-based measures. For large protein families, our synthetic sequences have similar or better properties compared to sequences generated by Potts models, including experimentally validated ones. Moreover, for small protein families, our generation method based on MSA Transformer outperforms Potts models. Our method also more accurately reproduces the higher-order statistics and the distribution of sequences in sequence space of natural data than Potts models. MSA Transformer is thus a strong candidate for protein sequence generation and protein design.

Impact of phylogeny on structural contact inference from protein sequence data

Local and global inference methods have been developed to infer structural contacts from multiple sequence alignments of homologous proteins. They rely on correlations in amino acid usage at contacting sites. Because homologous proteins share a common ancestry, their sequences also feature phylogenetic correlations, which can impair contact inference. We investigate this effect by generating controlled synthetic data from a minimal model where the importance of contacts and of phylogeny can be tuned. We demonstrate that global inference methods, specifically Potts models, are more resilient to phylogenetic correlations than local methods, based on covariance or mutual information. This holds whether or not phylogenetic corrections are used, and may explain the success of global methods. We analyse the roles of selection strength and of phylogenetic relatedness. We show that sites that mutate early in the phylogeny yield false positive contacts. We consider natural data and realistic synthetic data, and our findings generalize to these cases. Our results highlight the impact of phylogeny on contact prediction from protein sequences and illustrate the interplay between the rich structure of biological data and inference.

MuLan-Methyl-multiple transformer-based language models for accurate DNA methylation prediction

Transformer-based language models are successfully used to address massive text-related tasks. DNA methylation is an important epigenetic mechanism, and its analysis provides valuable insights into gene regulation and biomarker identification. Several deep learning-based methods have been proposed to identify DNA methylation, and each seeks to strike a balance between computational effort and accuracy. Here, we introduce MuLan-Methyl, a deep learning framework for predicting DNA methylation sites, which is based on 5 popular transformer-based language models. The framework identifies methylation sites for 3 different types of DNA methylation: N6-adenine, N4-cytosine, and 5-hydroxymethylcytosine. Each of the employed language models is adapted to the task using the "pretrain and fine-tune" paradigm. Pretraining is performed on a custom corpus of DNA fragments and taxonomy lineages using self-supervised learning. Fine-tuning aims at predicting the DNA methylation status of each type. The 5 models are used to collectively predict the DNA methylation status. We report excellent performance of MuLan-Methyl on a benchmark dataset. Moreover, we argue that the model captures characteristic differences between different species that are relevant for methylation. This work demonstrates that language models can be successfully adapted to applications in biological sequence analysis and that joint utilization of different language models improves model performance. Mulan-Methyl is open source, and we provide a web server that implements the approach.