Integrating physics in deep learning algorithms: a force field as a PyTorch module.

MOTIVATION: Deep learning algorithms applied to structural biology often struggle to converge to meaningful solutions when limited data is available, since they are required to learn complex physical rules from examples. State-of-the-art force-fields, however, cannot interface with deep learning algorithms due to their implementation. RESULTS: We present MadraX, a forcefield implemented as a differentiable PyTorch module, able to interact with deep learning algorithms in an end-to-end fashion. AVAILABILITY AND IMPLEMENTATION: MadraX documentation, together with tutorials and installation guide, is available at madrax.readthedocs.io.

HPMPdb: A machine learning-ready database of protein molecular phenotypes associated to human missense variants.

Current human Single Amino acid Variants (SAVs) databases provide a link between a SAVs and their effect on the carrier individual phenotype, often dividing them into Deleterious/Neutral variants. This is a very coarse-grained description of the genotype-to-phenotype relationship because it relies on un-realistic assumptions such as the perfect Mendelian behavior of each SAV and considers only dichotomic phenotypes. Moreover, the link between the effect of a SAV on a protein (its molecular phenotype) and the individual phenotype is often very complex, because multiple level of biological abstraction connect the protein and individual level phenotypes. Here we present HPMPdb, a manually curated database containing human SAVs associated with the detailed description of the molecular phenotype they cause on the affected proteins. With particular regards to machine learning (ML), this database can be used to let researchers go beyond the existing Deleterious/Neutral prediction paradigm, allowing them to build molecular phenotype predictors instead. Our class labels describe in a succinct way the effects that each SAV has on 15 protein molecular phenotypes, such as protein-protein interaction, small molecules binding, function, post-translational modifications (PTMs), sub-cellular localization, mimetic PTM, folding and protein expression. Moreover, we provide researchers with all necessary means to re-producibly train and test their models on our database. The webserver and the data described in this paper are available at hpmp.esat.kuleuven.be.

Prediction of Disordered Regions in Proteins with Recurrent Neural Networks and Protein Dynamics.

The role of intrinsically disordered protein regions (IDRs) in cellular processes has become increasingly evident over the last years. These IDRs continue to challenge structural biology experiments because they lack a well-defined conformation, and bioinformatics approaches that accurately delineate disordered protein regions remain essential for their identification and further investigation. Typically, these predictors use the protein amino acid sequence, without taking into account likely sequence-dependent emergent properties, such as protein backbone dynamics. Here we present DisoMine, a method that predicts protein'long disorder' with recurrent neural networks from simple predictions of protein dynamics, secondary structure and early folding. The tool is fast and requires only a single sequence, making it applicable for large-scale screening, including poorly studied and orphan proteins. DisoMine is a top performer in its category and compares well to disorder prediction approaches using evolutionary information. DisoMine is freely available through an interactive webserver at https://bio2byte.be/disomine/.

PyUUL provides an interface between biological structures and deep learning algorithms.

Structural bioinformatics suffers from the lack of interfaces connecting biological structures and machine learning methods, making the application of modern neural network architectures impractical. This negatively affects the development of structure-based bioinformatics methods, causing a bottleneck in biological research. Here we present PyUUL ( https://pyuul.readthedocs.io/ ), a library to translate biological structures into 3D tensors, allowing an out-of-the-box application of state-of-the-art deep learning algorithms. The library converts biological macromolecules to data structures typical of computer vision, such as voxels and point clouds, for which extensive machine learning research has been performed. Moreover, PyUUL allows an out-of-the box GPU and sparse calculation. Finally, we demonstrate how PyUUL can be used by researchers to address some typical bioinformatics problems, such as structure recognition and docking.

In silico prediction of in vitro protein liquid-liquid phase separation experiments outcomes with multi-head neural attention.

MOTIVATION: Proteins able to undergo liquid-liquid phase separation (LLPS) in vivo and in vitro are drawing a lot of interest, due to their functional relevance for cell life. Nevertheless, the proteome-scale experimental screening of these proteins seems unfeasible, because besides being expensive and time-consuming, LLPS is heavily influenced by multiple environmental conditions such as concentration, pH and temperature, thus requiring a combinatorial number of experiments for each protein. RESULTS: To overcome this problem, we propose a neural network model able to predict the LLPS behavior of proteins given specified experimental conditions, effectively predicting the outcome of in vitro experiments. Our model can be used to rapidly screen proteins and experimental conditions searching for LLPS, thus reducing the search space that needs to be covered experimentally. We experimentally validate Droppler's prediction on the TAR DNA-binding protein in different experimental conditions, showing the consistency of its predictions. AVAILABILITY AND IMPLEMENTATION: A python implementation of Droppler is available at https://bitbucket.org/grogdrinker/droppler. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

b2bTools: online predictions for protein biophysical features and their conservation.

We provide integrated protein sequence-based predictions via https://bio2byte.be/b2btools/. The aim of our predictions is to identify the biophysical behaviour or features of proteins that are not readily captured by structural biology and/or molecular dynamics approaches. Upload of a FASTA file or text input of a sequence provides integrated predictions from DynaMine backbone and side-chain dynamics, conformational propensities, and derived EFoldMine early folding, DisoMine disorder, and Agmata ß-sheet aggregation. These predictions, several of which were previously not available online, capture 'emergent' properties of proteins, i.e. the inherent biophysical propensities encoded in their sequence, rather than context-dependent behaviour (e.g. final folded state). In addition, upload of a multiple sequence alignment (MSA) in a variety of formats enables exploration of the biophysical variation observed in homologous proteins. The associated plots indicate the biophysical limits of functionally relevant protein behaviour, with unusual residues flagged by a Gaussian mixture model analysis. The prediction results are available as JSON or CSV files and directly accessible via an API. Online visualisation is available as interactive plots, with brief explanations and tutorial pages included. The server and API employ an email-free token-based system that can be used to anonymously access previously generated results.

Role and therapeutic potential of liquid-liquid phase separation in amyotrophic lateral sclerosis.

Amyotrophic lateral sclerosis (ALS) is a late-onset neurodegenerative disease selectively affecting motor neurons, leading to progressive paralysis. Although most cases are sporadic, ∼10% are familial. Similar proteins are found in aggregates in sporadic and familial ALS, and over the last decade, research has been focused on the underlying nature of this common pathology. Notably, TDP-43 inclusions are found in almost all ALS patients, while FUS inclusions have been reported in some familial ALS patients. Both TDP-43 and FUS possess 'low-complexity domains' (LCDs) and are considered as 'intrinsically disordered proteins', which form liquid droplets in vitro due to the weak interactions caused by the LCDs. Dysfunctional 'liquid-liquid phase separation' (LLPS) emerged as a new mechanism linking ALS-related proteins to pathogenesis. Here, we review the current state of knowledge on ALS-related gene products associated with a proteinopathy and discuss their status as LLPS proteins. In addition, we highlight the therapeutic potential of targeting LLPS for treating ALS.

Structure-based machine-guided mapping of amyloid sequence space reveals uncharted sequence clusters with higher solubilities.

The amyloid conformation can be adopted by a variety of sequences, but the precise boundaries of amyloid sequence space are still unclear. The currently charted amyloid sequence space is strongly biased towards hydrophobic, beta-sheet prone sequences that form the core of globular proteins and by Q/N/Y rich yeast prions. Here, we took advantage of the increasing amount of high-resolution structural information on amyloid cores currently available in the protein databank to implement a machine learning approach, named Cordax (https://cordax.switchlab.org), that explores amyloid sequence beyond its current boundaries. Clustering by t-Distributed Stochastic Neighbour Embedding (t-SNE) shows how our approach resulted in an expansion away from hydrophobic amyloid sequences towards clusters of lower aliphatic content and higher charge, or regions of helical and disordered propensities. These clusters uncouple amyloid propensity from solubility representing sequence flavours compatible with surface-exposed patches in globular proteins, functional amyloids or sequences associated to liquid-liquid phase transitions.

ShiftCrypt: a web server to understand and biophysically align proteins through their NMR chemical shift values.

Nuclear magnetic resonance (NMR) spectroscopy data provides valuable information on the behaviour of proteins in solution. The primary data to determine when studying proteins are the per-atom NMR chemical shifts, which reflect the local environment of atoms and provide insights into amino acid residue dynamics and conformation. Within an amino acid residue, chemical shifts present multi-dimensional and complexly cross-correlated information, making them difficult to analyse. The ShiftCrypt method, based on neural network auto-encoder architecture, compresses the per-amino acid chemical shift information in a single, interpretable, amino acid-type independent value that reflects the biophysical state of a residue. We here present the ShiftCrypt web server, which makes the method readily available. The server accepts chemical shifts input files in the NMR Exchange Format (NEF) or NMR-STAR format, executes ShiftCrypt and visualises the results, which are also accessible via an API. It also enables the "biophysically-based" pairwise alignment of two proteins based on their ShiftCrypt values. This approach uses Dynamic Time Warping and can optionally include their amino acid code information, and has applications in, for example, the alignment of disordered regions. The server uses a token-based system to ensure the anonymity of the users and results. The web server is available at www.bio2byte.be/shiftcrypt.

Insight into the protein solubility driving forces with neural attention.

Protein solubility is a key aspect for many biotechnological, biomedical and industrial processes, such as the production of active proteins and antibodies. In addition, understanding the molecular determinants of the solubility of proteins may be crucial to shed light on the molecular mechanisms of diseases caused by aggregation processes such as amyloidosis. Here we present SKADE, a novel Neural Network protein solubility predictor and we show how it can provide novel insight into the protein solubility mechanisms, thanks to its neural attention architecture. First, we show that SKADE positively compares with state of the art tools while using just the protein sequence as input. Then, thanks to the neural attention mechanism, we use SKADE to investigate the patterns learned during training and we analyse its decision process. We use this peculiarity to show that, while the attention profiles do not correlate with obvious sequence aspects such as biophysical properties of the aminoacids, they suggest that N- and C-termini are the most relevant regions for solubility prediction and are predictive for complex emergent properties such as aggregation-prone regions involved in beta-amyloidosis and contact density. Moreover, SKADE is able to identify mutations that increase or decrease the overall solubility of the protein, allowing it to be used to perform large scale in-silico mutagenesis of proteins in order to maximize their solubility.

Accurate prediction of protein beta-aggregation with generalized statistical potentials.

MOTIVATION: Protein beta-aggregation is an important but poorly understood phenomena involved in diseases as well as in beneficial physiological processes. However, while this task has been investigated for over 50 years, very little is known about its mechanisms of action. Moreover, the identification of regions involved in aggregation is still an open problem and the state-of-the-art methods are often inadequate in real case applications. RESULTS: In this article we present AgMata, an unsupervised tool for the identification of such regions from amino acidic sequence based on a generalized definition of statistical potentials that includes biophysical information. The tool outperforms the state-of-the-art methods on two different benchmarks. As case-study, we applied our tool to human ataxin-3, a protein involved in Machado-Joseph disease. Interestingly, AgMata identifies aggregation-prone residues that share the very same structural environment. Additionally, it successfully predicts the outcome of in vitro mutagenesis experiments, identifying point mutations that lead to an alteration of the aggregation propensity of the wild-type ataxin-3. AVAILABILITY AND IMPLEMENTATION: A python implementation of the tool is available at https://bitbucket.org/bio2byte/agmata. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Exploring the limitations of biophysical propensity scales coupled with machine learning for protein sequence analysis.

Machine learning (ML) is ubiquitous in bioinformatics, due to its versatility. One of the most crucial aspects to consider while training a ML model is to carefully select the optimal feature encoding for the problem at hand. Biophysical propensity scales are widely adopted in structural bioinformatics because they describe amino acids properties that are intuitively relevant for many structural and functional aspects of proteins, and are thus commonly used as input features for ML methods. In this paper we reproduce three classical structural bioinformatics prediction tasks to investigate the main assumptions about the use of propensity scales as input features for ML methods. We investigate their usefulness with different randomization experiments and we show that their effectiveness varies among the ML methods used and the tasks. We show that while linear methods are more dependent on the feature encoding, the specific biophysical meaning of the features is less relevant for non-linear methods. Moreover, we show that even among linear ML methods, the simpler one-hot encoding can surprisingly outperform the "biologically meaningful" scales. We also show that feature selection performed with non-linear ML methods may not be able to distinguish between randomized and "real" propensity scales by properly prioritizing to the latter. Finally, we show that learning problem-specific embeddings could be a simple, assumptions-free and optimal way to perform feature learning/engineering for structural bioinformatics tasks.

Author Correction: Exploring the Sequence-based Prediction of Folding Initiation Sites in Proteins.

An amendment to this paper has been published and can be accessed via a link at the top of the paper.

Auto-encoding NMR chemical shifts from their native vector space to a residue-level biophysical index.

Chemical shifts (CS) are determined from NMR experiments and represent the resonance frequency of the spin of atoms in a magnetic field. They contain a mixture of information, encompassing the in-solution conformations a protein adopts, as well as the movements it performs. Due to their intrinsically multi-faceted nature, CS are difficult to interpret and visualize. Classical approaches for the analysis of CS aim to extract specific protein-related properties, thus discarding a large amount of information that cannot be directly linked to structural features of the protein. Here we propose an autoencoder-based method, called ShiftCrypt, that provides a way to analyze, compare and interpret CS in their native, multidimensional space. We show that ShiftCrypt conserves information about the most common structural features. In addition, it can be used to identify hidden similarities between diverse proteins and peptides, and differences between the same protein in two different binding states.

Computational identification of prion-like RNA-binding proteins that form liquid phase-separated condensates.

MOTIVATION: Eukaryotic cells contain different membrane-delimited compartments, which are crucial for the biochemical reactions necessary to sustain cell life. Recent studies showed that cells can also trigger the formation of membraneless organelles composed by phase-separated proteins to respond to various stimuli. These condensates provide new ways to control the reactions and phase-separation proteins (PSPs) are thus revolutionizing how cellular organization is conceived. The small number of experimentally validated proteins, and the difficulty in discovering them, remain bottlenecks in PSPs research. RESULTS: Here we present PSPer, the first in-silico screening tool for prion-like RNA-binding PSPs. We show that it can prioritize PSPs among proteins containing similar RNA-binding domains, intrinsically disordered regions and prions. PSPer is thus suitable to screen proteomes, identifying the most likely PSPs for further experimental investigation. Moreover, its predictions are fully interpretable in the sense that it assigns specific functional regions to the predicted proteins, providing valuable information for experimental investigation of targeted mutations on these regions. Finally, we show that it can estimate the ability of artificially designed proteins to form condensates (r=-0.87), thus providing an in-silico screening tool for protein design experiments. AVAILABILITY AND IMPLEMENTATION: PSPer is available at bio2byte.com/psp. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Large-scale in-silico statistical mutagenesis analysis sheds light on the deleteriousness landscape of the human proteome.

Next generation sequencing technologies are providing increasing amounts of sequencing data, paving the way for improvements in clinical genetics and precision medicine. The interpretation of the observed genomic variants in the light of their phenotypic effects is thus emerging as a crucial task to solve in order to advance our understanding of how exomic variants affect proteins and how the proteins' functional changes affect human health. Since the experimental evaluation of the effects of every observed variant is unfeasible, Bioinformatics methods are being developed to address this challenge in-silico, by predicting the impact of millions of variants, thus providing insight into the deleteriousness landscape of entire proteomes. Here we show the feasibility of this approach by using the recently developed DEOGEN2 variant-effect predictor to perform the largest in-silico mutagenesis scan to date. We computed the deleteriousness score of 170 million variants over 15000 human proteins and we analysed the results, investigating how the predicted deleteriousness landscape of the proteins relates to known functionally and structurally relevant protein regions and biophysical properties. Moreover, we qualitatively validated our results by comparing them with two mutagenesis studies targeting two specific proteins, showing the consistency of DEOGEN2 predictions with respect to experimental data.

Ultra-fast global homology detection with Discrete Cosine Transform and Dynamic Time Warping.

Motivation: Evolutionary information is crucial for the annotation of proteins in bioinformatics. The amount of retrieved homologs often correlates with the quality of predicted protein annotations related to structure or function. With a growing amount of sequences available, fast and reliable methods for homology detection are essential, as they have a direct impact on predicted protein annotations. Results: We developed a discriminative, alignment-free algorithm for homology detection with quasi-linear complexity, enabling theoretically much faster homology searches. To reach this goal, we convert the protein sequence into numeric biophysical representations. These are shrunk to a fixed length using a novel vector quantization method which uses a Discrete Cosine Transform compression. We then compute, for each compressed representation, similarity scores between proteins with the Dynamic Time Warping algorithm and we feed them into a Random Forest. The WARP performances are comparable with state of the art methods. Availability and implementation: The method is available at http://ibsquare.be/warp. Supplementary information: Supplementary data are available at Bioinformatics online.

Exploring the Sequence-based Prediction of Folding Initiation Sites in Proteins.

Protein folding is a complex process that can lead to disease when it fails. Especially poorly understood are the very early stages of protein folding, which are likely defined by intrinsic local interactions between amino acids close to each other in the protein sequence. We here present EFoldMine, a method that predicts, from the primary amino acid sequence of a protein, which amino acids are likely involved in early folding events. The method is based on early folding data from hydrogen deuterium exchange (HDX) data from NMR pulsed labelling experiments, and uses backbone and sidechain dynamics as well as secondary structure propensities as features. The EFoldMine predictions give insights into the folding process, as illustrated by a qualitative comparison with independent experimental observations. Furthermore, on a quantitative proteome scale, the predicted early folding residues tend to become the residues that interact the most in the folded structure, and they are often residues that display evolutionary covariation. The connection of the EFoldMine predictions with both folding pathway data and the folded protein structure suggests that the initial statistical behavior of the protein chain with respect to local structure formation has a lasting effect on its subsequent states.

SVM-dependent pairwise HMM: an application to protein pairwise alignments.

MOTIVATION: Methods able to provide reliable protein alignments are crucial for many bioinformatics applications. In the last years many different algorithms have been developed and various kinds of information, from sequence conservation to secondary structure, have been used to improve the alignment performances. This is especially relevant for proteins with highly divergent sequences. However, recent works suggest that different features may have different importance in diverse protein classes and it would be an advantage to have more customizable approaches, capable to deal with different alignment definitions. RESULTS: Here we present Rigapollo, a highly flexible pairwise alignment method based on a pairwise HMM-SVM that can use any type of information to build alignments. Rigapollo lets the user decide the optimal features to align their protein class of interest. It outperforms current state of the art methods on two well-known benchmark datasets when aligning highly divergent sequences. AVAILABILITY AND IMPLEMENTATION: A Python implementation of the algorithm is available at http://ibsquare.be/rigapollo. CONTACT: wim.vranken@vub.be. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.