Large Protein Datasets Research Articles

Application of artificial intelligence and machine learning techniques to the analysis of dynamic protein sequences.

We apply methods of Artificial Intelligence and Machine Learning to protein dynamic bioinformatics. We rewrite the sequences of a large protein data set, containing both folded and intrinsically disordered molecules, using a representation developed previously, which encodes the intrinsic dynamic properties of the naturally occurring amino acids. We Fourier analyze the resulting sequences. It is demonstrated that classification models built using several different supervised learning methods are able to successfully distinguish folded from intrinsically disordered proteins from sequence alone. It is further shown that the most important sequence property for this discrimination is the sequence mobility, which is the sequence averaged value of the residue-specific average alpha carbon B factor. This is in agreement with previous work, in which we have demonstrated the central role played by the sequence mobility in protein dynamic bioinformatics and biophysics. This finding opens a path to the application of dynamic bioinformatics, in combination with machine learning algorithms, to a range of significant biomedical problems.

INGNN-DTI: prediction of drug-target interaction with interpretable nested graph neural network and pretrained molecule models.

Drug-target interaction (DTI) prediction aims to identify interactions between drugs and protein targets. Deep learning can automatically learn discriminative features from drug and protein target representations for DTI prediction, but challenges remain, making it an open question. Existing approaches encode drugs and targets into features using deep learning models, but they often lack explanations for underlying interactions. Moreover, limited labeled DTIs in the chemical space can hinder model generalization. We propose an interpretable nested graph neural network for DTI prediction (iNGNN-DTI) using pre-trained molecule and protein models. The analysis is conducted on graph data representing drugs and targets by using a specific type of nested graph neural network, in which the target graphs are created based on 3D structures using Alphafold2. This architecture is highly expressive in capturing substructures of the graph data. We use a cross-attention module to capture interaction information between the substructures of drugs and targets. To improve feature representations, we integrate features learned by models that are pre-trained on large unlabeled small molecule and protein datasets, respectively. We evaluate our model on three benchmark datasets, and it shows a consistent improvement on all baseline models in all datasets. We also run an experiment with previously unseen drugs or targets in the test set, and our model outperforms all of the baselines. Furthermore, the iNGNN-DTI can provide more insights into the interaction by visualizing the weights learned by the cross-attention module. The source code of the algorithm is available at https://github.com/syan1992/iNGNN-DTI.

RedRibbon: A new rank-rank hypergeometric overlap for gene and transcript expression signatures.

High-throughput omics technologies have generated a wealth of large protein, gene, and transcript datasets that have exacerbated the need for new methods to analyse and compare big datasets. Rank-rank hypergeometric overlap is an important threshold-free method to combine and visualize two ranked lists of P-values or fold-changes, usually from differential gene expression analyses. Here, we introduce a new rank-rank hypergeometric overlap-based method aimed at gene level and alternative splicing analyses at transcript or exon level, hitherto unreachable as transcript numbers are an order of magnitude larger than gene numbers. We tested the tool on synthetic and real datasets at gene and transcript levels to detect correlation and anticorrelation patterns and found it to be fast and accurate, even on very large datasets thanks to an evolutionary algorithm-based minimal P-value search. The tool comes with a ready-to-use permutation scheme allowing the computation of adjusted P-values at low time cost. The package compatibility mode is a drop-in replacement to previous packages. RedRibbon holds the promise to accurately extricate detailed information from large comparative analyses.

Protein language models can capture protein quaternary state

BackgroundDetermining a protein’s quaternary state, i.e. the number of monomers in a functional unit, is a critical step in protein characterization. Many proteins form multimers for their activity, and over 50% are estimated to naturally form homomultimers. Experimental quaternary state determination can be challenging and require extensive work. To complement these efforts, a number of computational tools have been developed for quaternary state prediction, often utilizing experimentally validated structural information. Recently, dramatic advances have been made in the field of deep learning for predicting protein structure and other characteristics. Protein language models, such as ESM-2, that apply computational natural-language models to proteins successfully capture secondary structure, protein cell localization and other characteristics, from a single sequence. Here we hypothesize that information about the protein quaternary state may be contained within protein sequences as well, allowing us to benefit from these novel approaches in the context of quaternary state prediction.ResultsWe generated ESM-2 embeddings for a large dataset of proteins with quaternary state labels from the curated QSbio dataset. We trained a model for quaternary state classification and assessed it on a non-overlapping set of distinct folds (ECOD family level). Our model, named QUEEN (QUaternary state prediction using dEEp learNing), performs worse than approaches that include information from solved crystal structures. However, it successfully learned to distinguish multimers from monomers, and predicts the specific quaternary state with moderate success, better than simple sequence similarity-based annotation transfer. Our results demonstrate that complex, quaternary state related information is included in such embeddings.ConclusionsQUEEN is the first to investigate the power of embeddings for the prediction of the quaternary state of proteins. As such, it lays out strengths as well as limitations of a sequence-based protein language model approach, compared to structure-based approaches. Since it does not require any structural information and is fast, we anticipate that it will be of wide use both for in-depth investigation of specific systems, as well as for studies of large sets of protein sequences. A simple colab implementation is available at: https://colab.research.google.com/github/Furman-Lab/QUEEN/blob/main/QUEEN_prediction_notebook.ipynb.

Deep transfer learning for inter-chain contact predictions of transmembrane protein complexes

Membrane proteins are encoded by approximately a quarter of human genes. Inter-chain residue-residue contact information is important for structure prediction of membrane protein complexes and valuable for understanding their molecular mechanism. Although many deep learning methods have been proposed to predict the intra-protein contacts or helix-helix interactions in membrane proteins, it is still challenging to accurately predict their inter-chain contacts due to the limited number of transmembrane proteins. Addressing the challenge, here we develop a deep transfer learning method for predicting inter-chain contacts of transmembrane protein complexes, named DeepTMP, by taking advantage of the knowledge pre-trained from a large data set of non-transmembrane proteins. DeepTMP utilizes a geometric triangle-aware module to capture the correct inter-chain interaction from the coevolution information generated by protein language models. DeepTMP is extensively evaluated on a test set of 52 self-associated transmembrane protein complexes, and compared with state-of-the-art methods including DeepHomo2.0, CDPred, GLINTER, DeepHomo, and DNCON2_Inter. It is shown that DeepTMP considerably improves the precision of inter-chain contact prediction and outperforms the existing approaches in both accuracy and robustness.

SHEPHARD: a modular and extensible software architecture for analyzing and annotating large protein datasets.

The emergence of high-throughput experiments and high-resolution computational predictions has led to an explosion in the quality and volume of protein sequence annotations at proteomic scales. Unfortunately, sanity checking, integrating, and analyzing complex sequence annotations remains logistically challenging and introduces a major barrier to entry for even superficial integrative bioinformatics. To address this technical burden, we have developed SHEPHARD, a Python framework that trivializes large-scale integrative protein bioinformatics. SHEPHARD combines an object-oriented hierarchical data structure with database-like features, enabling programmatic annotation, integration, and analysis of complex datatypes. Importantly SHEPHARD is easy to use and enables a Pythonic interrogation of largescale protein datasets with millions of unique annotations. We use SHEPHARD to examine three orthogonal proteome-wide questions relating protein sequence to molecular function, illustrating its ability to uncover novel biology. We provided SHEPHARD as both a stand-alone software package (https://github.com/holehouse-lab/shephard), and as a Google Colab notebook with a collection of precomputed proteome-wide annotations (https://github.com/holehouse-lab/shephard-colab). Supplementary data are available at Bioinformatics online.

COLLAPSE: A representation learning framework for identification and characterization of protein structural sites.

The identification and characterization of the structural sites which contribute to protein function are crucial for understanding biological mechanisms, evaluating disease risk, and developing targeted therapies. However, the quantity of known protein structures is rapidly outpacing our ability to functionally annotate them. Existing methods for function prediction either do not operate on local sites, suffer from high false positive or false negative rates, or require large site-specific training datasets, necessitating the development of new computational methods for annotating functional sites at scale. We present COLLAPSE (Compressed Latents Learned from Aligned Protein Structural Environments), a framework for learning deep representations of protein sites. COLLAPSE operates directly on the 3D positions of atoms surrounding a site and uses evolutionary relationships between homologous proteins as a self-supervision signal, enabling learned embeddings to implicitly capture structure-function relationships within each site. Our representations generalize across disparate tasks in a transfer learning context, achieving state-of-the-art performance on standardized benchmarks (protein-protein interactions and mutation stability) and on the prediction of functional sites from the Prosite database. We use COLLAPSE to search for similar sites across large protein datasets and to annotate proteins based on a database of known functional sites. These methods demonstrate that COLLAPSE is computationally efficient, tunable, and interpretable, providing a general-purpose platform for computational protein analysis.

Estimating amino acid substitution models for metazoan evolutionary studies.

Amino acid substitution models represent the substitution rates among amino acids during the evolution of protein sequences. The models are a prerequisite for maximum likelihood or Bayesian methods to analyse the phylogenetic relationships among species based on their protein sequences. Estimating amino acid substitution models requires large protein datasets and intensive computation. In this paper, we presented the estimation of both time-reversible model (Q.met) and time non-reversible model (NQ.met) for multicellular animals (Metazoa). Analyses showed that the Q.met and NQ.met models were significantly better than existing models in analysing metazoan protein sequences. Moreover, the time non-reversible model NQ.met enables us to reconstruct the rooted phylogenetic tree for Metazoa. We recommend researchers to employ the Q.met and NQ.met models in analysing metazoan protein sequences.

Evaluating Mineral Lattices as Evolutionary Proxies for Metalloprotein Evolution.

Protein coordinated iron-sulfur clusters drive electron flow within metabolic pathways for organisms throughout the tree of life. It is not known how iron-sulfur clusters were first incorporated into proteins. Structural analogies to iron-sulfide minerals present on early Earth, suggest a connection in the evolution of both proteins and minerals. The availability of large protein and mineral crystallographic structure data sets, provides an opportunity to explore co-evolution of proteins and minerals on a large-scale using informatics approaches. However, quantitative comparisons are confounded by the infinite, repeating nature of the mineral lattice, in contrast to metal clusters in proteins, which are finite in size. We address this problem using the Niggli reduction to transform a mineral lattice to a finite, unique structure that when translated reproduces the crystal lattice. Protein and reduced mineral structures were represented as quotient graphs with the edges and nodes corresponding to bonds and atoms, respectively. We developed a graph theory-based method to calculate the maximum common connected edge subgraph (MCCES) between mineral and protein quotient graphs. MCCES can accommodate differences in structural volumes and easily allows additional chemical criteria to be considered when calculating similarity. To account for graph size differences, we use the Tversky similarity index. Using consistent criteria, we found little similarity between putative ancient iron-sulfur protein clusters and iron-sulfur mineral lattices, suggesting these metal sites are not as evolutionarily connected as once thought. We discuss possible evolutionary implications of these findings in addition to suggesting an alternative proxy, mineral surfaces, for better understanding the coevolution of the geosphere and biosphere.

Pathogenic variation types in human genes relate to diseases through Pfam and InterPro mapping.

Grouping residue variations in a protein according to their physicochemical properties allows a dimensionality reduction of all the possible substitutions in a variant with respect to the wild type. Here, by using a large dataset of proteins with disease-related and benign variations, as derived by merging Humsavar and ClinVar data, we investigate to which extent our physicochemical grouping procedure can help in determining whether patterns of variation types are related to specific groups of diseases and whether they occur in Pfam and/or InterPro gene domains. Here, we download 75,145 germline disease-related and benign variations of 3,605 genes, group them according to physicochemical categories and map them into Pfam and InterPro gene domains. Statistically validated analysis indicates that each cluster of genes associated to Mondo anatomical system categorizations is characterized by a specific variation pattern. Patterns identify specific Pfam and InterPro domain–Mondo category associations. Our data suggest that the association of variation patterns to Mondo categories is unique and may help in associating gene variants to genetic diseases. This work corroborates in a much larger data set previous observations from our group.

GRaSP-web: a machine learning strategy to predict binding sites based on residue neighborhood graphs.

Proteins are essential macromolecules for the maintenance of living systems. Many of them perform their function by interacting with other molecules in regions called binding sites. The identification and characterization of these regions are of fundamental importance to determine protein function, being a fundamental step in processes such as drug design and discovery. However, identifying such binding regions is not trivial due to the drawbacks of experimental methods, which are costly and time-consuming. Here we propose GRaSP-web, a web server that uses GRaSP (Graph-based Residue neighborhood Strategy to Predict binding sites), a residue-centric method based on graphs that uses machine learning to predict putative ligand binding site residues. The method outperformed 6 state-of-the-art residue-centric methods (MCC of 0.61). Also, GRaSP-web is scalable as it takes 10-20 seconds to predict binding sites for a protein complex (the state-of-the-art residue-centric method takes 2-5h on the average). It proved to be consistent in predicting binding sites for bound/unbound structures (MCC 0.61 for both) and for a large dataset of multi-chain proteins (4500 entries, MCC 0.61). GRaSPWeb is freely available at https://grasp.ufv.br.

PolyX2: Fast Detection of Homorepeats in Large Protein Datasets.

Homorepeat sequences, consecutive runs of identical amino acids, are prevalent in eukaryotic proteins. It has become necessary to annotate and evaluate this feature in entire proteomes. The definition of what constitutes a homorepeat is not fixed, and different research approaches may require different definitions; therefore, flexible approaches to analyze homorepeats in complete proteomes are needed. Here, we present polyX2, a fast, simple but tunable script to scan protein datasets for all possible homorepeats. The user can modify the length of the window to scan, the minimum number of identical residues that must be found in the window, and the types of homorepeats to be found.

Inter-paralog amino acid inversion events in large phylogenies of duplicated proteins.

Connecting protein sequence to function is becoming increasingly relevant since high-throughput sequencing studies accumulate large amounts of genomic data. In order to go beyond the existing database annotation, it is fundamental to understand the mechanisms underlying functional inheritance and divergence. If the homology relationship between proteins is known, can we determine whether the function diverged? In this work, we analyze different possibilities of protein sequence evolution after gene duplication and identify "inter-paralog inversions", i.e., sites where the relationship between the ancestry and the functional signal is decoupled. The amino acids in these sites are masked from being recognized by other prediction tools. Still, they play a role in functional divergence and could indicate a shift in protein function. We develop a method to specifically recognize inter-paralog amino acid inversions in a phylogeny and test it on real and simulated datasets. In a dataset built from the Epidermal Growth Factor Receptor (EGFR) sequences found in 88 fish species, we identify 19 amino acid sites that went through inversion after gene duplication, mostly located at the ligand-binding extracellular domain. Our work uncovers an outcome of protein duplications with direct implications in protein functional annotation and sequence evolution. The developed method is optimized to work with large protein datasets and can be readily included in a targeted protein analysis pipeline.

Complementarity of the residue-level protein function and structure predictions in human proteins

Sequence-based predictors of the residue-level protein function and structure cover a broad spectrum of characteristics including intrinsic disorder, secondary structure, solvent accessibility and binding to nucleic acids. They were catalogued and evaluated in numerous surveys and assessments. However, methods focusing on a given characteristic are studied separately from predictors of other characteristics, while they are typically used on the same proteins. We fill this void by studying complementarity of a representative collection of methods that target different predictions using a large, taxonomically consistent, and low similarity dataset of human proteins. First, we bridge the gap between the communities that develop structure-trained vs. disorder-trained predictors of binding residues. Motivated by a recent study of the protein-binding residue predictions, we empirically find that combining the structure-trained and disorder-trained predictors of the DNA-binding and RNA-binding residues leads to substantial improvements in predictive quality. Second, we investigate whether diverse predictors generate results that accurately reproduce relations between secondary structure, solvent accessibility, interaction sites, and intrinsic disorder that are present in the experimental data. Our empirical analysis concludes that predictions accurately reflect all combinations of these relations. Altogether, this study provides unique insights that support combining results produced by diverse residue-level predictors of protein function and structure.

DispHScan: A Multi-Sequence Web Tool for Predicting Protein Disorder as a Function of pH.

Proteins are exposed to fluctuating environmental conditions in their cellular context and during their biotechnological production. Disordered regions are susceptible to these fluctuations and may experience solvent-dependent conformational switches that affect their local dynamism and activity. In a recent study, we modeled the influence of pH in the conformational state of IDPs by exploiting a charge–hydrophobicity diagram that considered the effect of solution pH on both variables. However, it was not possible to predict context-dependent transitions for multiple sequences, precluding proteome-wide analysis or the screening of collections of mutants. In this article, we present DispHScan, the first computational tool dedicated to predicting pH-induced disorder–order transitions in large protein datasets. The DispHScan web server allows the users to run pH-dependent disorder predictions of multiple sequences and identify context-dependent conformational transitions. It might provide new insights on the role of pH-modulated conditional disorder in the physiology and pathology of different organisms. The DispHScan web server is freely available for academic users, it is platform-independent and does not require previous registration.

PARROT is a flexible recurrent neural network framework for analysis of large protein datasets.

The rise of high-throughput experiments has transformed how scientists approach biological questions. The ubiquity of large-scale assays that can test thousands of samples in a day has necessitated the development of new computational approaches to interpret this data. Among these tools, machine learning approaches are increasingly being utilized due to their ability to infer complex nonlinear patterns from high-dimensional data. Despite their effectiveness, machine learning (and in particular deep learning) approaches are not always accessible or easy to implement for those with limited computational expertise. Here we present PARROT, a general framework for training and applying deep learning-based predictors on large protein datasets. Using an internal recurrent neural network architecture, PARROT is capable of tackling both classification and regression tasks while only requiring raw protein sequences as input. We showcase the potential uses of PARROT on three diverse machine learning tasks: predicting phosphorylation sites, predicting transcriptional activation function of peptides generated by high-throughput reporter assays, and predicting the fibrillization propensity of amyloid beta with data generated by deep mutational scanning. Through these examples, we demonstrate that PARROT is easy to use, performs comparably to state-of-the-art computational tools, and is applicable for a wide array of biological problems.

Dynamics-Evolution Correspondence in Protein Structures.

The genotype-phenotype mapping of proteins is a fundamental question in structural biology. In this Letter, with the analysis of a large dataset of proteins from hundreds of protein families, we quantitatively demonstrate the correlations between the noise-induced protein dynamics and mutation-induced variations of native structures, indicating the dynamics-evolution correspondence of proteins. Based on the investigations of the linear responses of native proteins, the origin of such a correspondence is elucidated. It is essential that the noise- and mutation-induced deformations of the proteins are restricted on a common low-dimensional subspace, as confirmed from the data. These results suggest an evolutionary mechanism of the proteins gaining both dynamical flexibility and evolutionary structural variability.

Protein secondary structure prediction (PSSP) using different machine algorithms

BackgroundThe computational biology approach has advanced exponentially in protein secondary structure prediction (PSSP), which is vital for the pharmaceutical industry. Extracting protein structure from the laboratory has insufficient information for PSSP that is used in bioinformatics studies. In this paper, the support vector machine (SVM) model and decision tree are presented on the RS126 dataset to address the problem of PSSP. A decision tree is applied for the SVM outcome to obtain the relevant guidelines possible for PSSP. Furthermore, the number of produced rules was fairly small, and they show a greater degree of comprehensibility compared to other rules. Several of the proposed principles have compelling and relevant biological clarification.ResultsThe results confirmed that the existence of a particular amino acid in a protein sequence increases the stability for the forecast of protein secondary structure. The suggested algorithm achieved 85% accuracy for the E|~E classifier.ConclusionsThe proposed rules can be very important in managing wet laboratory experiments intended at determining protein secondary structure. Lastly, future work will focus mainly on large protein datasets without overfitting and expand the amount of extracted regulations for PSSP.

Biostimulant impacts of Glutacetine® and derived formulations (VNT1 and VNT4) on the bread wheat grain proteome

Nitrogen (N) fertilizer is essential to ensure grain yield and quality in bread wheat. Improving N use efficiency is therefore crucial for wheat grain protein quality. In the present work, we analysed the effects on the winter wheat grain proteome of biostimulants containing Glutacetine® or two derived formulations (VNT1 and 4) when mixed with urea-ammonium-nitrate fertilizer. A large-scale quantitative proteomics analysis of two wheat flour fractions produced a dataset of 4369 identified proteins. Quantitative analysis revealed 9, 39 and 96 proteins with a significant change in abundance after Glutacetine®, VNT1 and VNT4 treatments, respectively, with a common set of 11 proteins that were affected by two different biostimulants. The major effects impacted proteins involved in (i) protein synthesis regulation (mainly ribosomal and binding proteins), (ii) defence and responses to stresses (including chitin-binding protein, heat shock 70 kDa protein 1 and glutathione S-transferase proteins), (iii) storage functions related to gluten protein alpha-gliadins and starch synthase and (iv) seed development with proteins implicated in protease activity, energy machinery, and the C and N metabolism pathways. Altogether, our study showed that Glutacetine®, VNT1 and VNT4 biostimulants positively affected protein composition related to grain quality.Data are available via ProteomeXchange with identifier PXD021513. SignificanceWe performed a large-scale quantitative proteomics study of the total protein extracts from flour samples to determine the effect of Glutacetine®-based biostimulants treatment on the protein composition of bread wheat grain. To our knowledge, only a few studies in the literature have applied proteomic approaches to study bread wheat grains and in particular to investigate the effect of biostimulants on the grain proteome of this cereal crop. In addition, most approaches used fractional extraction of proteins to target reserve proteins followed electrophoresis which leads to low identification rate of proteins. We identified and quantified a large protein dataset of 4369 proteins and determined ontological class of proteins affected by biostimulants treatments. Our proteomics investigation revealed the important role of these new biostimulants in achieving significant changes in protein synthesis regulation, storage functions, protease activity, energy machinery, C and N metabolism pathways and responses to biotic and abiotic stresses in grain.

Avoided motifs: short amino acid strings missing from protein datasets.

According to the amino acid composition of natural proteins, it could be expected that all possible sequences of three or four amino acids will occur at least once in large protein datasets purely by chance. However, in some species or cellular context, specific short amino acid motifs are missing due to unknown reasons. We describe these as Avoided Motifs, short amino acid combinations missing from biological sequences. Here we identify 209 human and 154 bacterial Avoided Motifs of length four amino acids, and discuss their possible functionality according to their presence in other species. Furthermore, we determine two Avoided Motifs of length three amino acids in human proteins specifically located in the cytoplasm, and two more in secreted proteins. Our results support the hypothesis that the characterization of Avoided Motifs in particular contexts can provide us with information about functional motifs, pointing to a new approach in the use of molecular sequences for the discovery of protein function.