Sequence-based Machine Learning Research Articles

Sequence-Based Machine Learning Reveals 3D Genome Differences between Bonobos and Chimpanzees.

The 3D structure of the genome is an important mediator of gene expression. As phenotypic divergence is largely driven by gene regulatory variation, comparing genome 3D contacts across species can further understanding of the molecular basis of species differences. However, while experimental data on genome 3D contacts in humans are increasingly abundant, only a handful of 3D genome contact maps exist for other species. Here, we demonstrate that human experimental data can be used to close this data gap. We apply a machine learning model that predicts 3D genome contacts from DNA sequence to the genomes from 56 bonobos and chimpanzees and identify species-specific patterns of genome folding. We estimated 3D divergence between individuals from the resulting contact maps in 4,420 1 Mb genomic windows, of which ∼17% were substantially divergent in predicted genome contacts. Bonobos and chimpanzees diverged at 89 windows, overlapping genes associated with multiple traits implicated in Pan phenotypic divergence. We discovered 51 bonobo-specific variants that individually produce the observed bonobo contact pattern in bonobo-chimpanzee divergent windows. Our results demonstrate that machine learning methods can leverage human data to fill in data gaps across species, offering the first look at population-level 3D genome variation in nonhuman primates. We also identify loci where changes in 3D folding may contribute to phenotypic differences in our closest living relatives.

Knot or not? Identifying unknotted proteins in knotted families with sequence-based Machine Learning model.

Knotted proteins, although scarce, are crucial structural components of certain protein families, and their roles continue to be a topic of intense research. Capitalizing on the vast collection of protein structure predictions offered by AlphaFold (AF), this study computationally examines the entire UniProt database to create a robust dataset of knotted and unknotted proteins. Utilizing this dataset, we develop a machine learning (ML) model capable of accurately predicting the presence of knots in protein structures solely from their amino acid sequences. We tested the model's capabilities on 100 proteins whose structures had not yet been predicted by AF and found agreement with our local prediction in 92% cases. From the point of view of structural biology, we found that all potentially knotted proteins predicted by AF can be classified only into 17 families. This allows us to discover the presence of unknotted proteins in families with a highly conserved knot. We found only three new protein families: UCH, DUF4253, and DUF2254, that contain both knotted and unknotted proteins, and demonstrate that deletions within the knot core could potentially account for the observed unknotted (trivial) topology. Finally, we have shown that in the majority of knotted families (11 out of 15), the knotted topology is strictly conserved in functional proteins with very low sequence similarity. We have conclusively demonstrated that proteins AF predicts as unknotted are structurally accurate in their unknotted configurations. However, these proteins often represent nonfunctional fragments, lacking significant portions of the knot core (amino acid sequence).

SuPreMo: a computational tool for streamlining in silico perturbation using sequence-based predictive models.

The increasing development of sequence-based machine learning models has raised the demand for manipulating sequences for this application. However, existing approaches to edit and evaluate genome sequences using models have limitations, such as incompatibility with structural variants, challenges in identifying responsible sequence perturbations, and the need for vcf file inputs and phased data. To address these bottlenecks, we present Sequence Mutator for Predictive Models (SuPreMo), a scalable and comprehensive tool for performing and supporting in silico mutagenesis experiments. We then demonstrate how pairs of reference and perturbed sequences can be used with machine learning models to prioritize pathogenic variants or discover new functional sequences. SuPreMo was written in Python, and can be run using only one line of code to generate both sequences and 3D genome disruption scores. The codebase, instructions for installation and use, and tutorials are on the GitHub page: https://github.com/ketringjoni/SuPreMo.

DRBpred: A sequence-based machine learning method to effectively predict DNA- and RNA-binding residues

DNA-binding and RNA-binding proteins are essential to an organism's normal life cycle. These proteins have diverse functions in various biological processes. DNA-binding proteins are crucial for DNA replication, transcription, repair, packaging, and gene expression. Likewise, RNA-binding proteins are essential for the post-transcriptional control of RNAs and RNA metabolism. Identifying DNA- and RNA-binding residue is essential for biological research and understanding the pathogenesis of many diseases. However, most DNA-binding and RNA-binding proteins still need to be discovered. This research explored various properties of the protein sequences, such as amino acid composition type, Position-Specific Scoring Matrix (PSSM) values of amino acids, Hidden Markov model (HMM) profiles, physiochemical properties, structural properties, torsion angles, and disorder regions. We utilized a sliding window technique to extract more information from a target residue's neighbors. We proposed an optimized Light Gradient Boosting Machine (LightGBM) method, named DRBpred, to predict DNA-binding and RNA-binding residues from the protein sequence. DRBpred shows an improvement of 112.00 %, 33.33 %, and 6.49 % for the DNA-binding test set compared to the state-of-the-art method. It shows an improvement of 112.50 %, 16.67 %, and 7.46 % for the RNA-binding test set regarding Sensitivity, Mathews Correlation Coefficient (MCC), and AUC metric.

A sequence-based machine learning model for predicting antigenic distance for H3N2 influenza virus.

Seasonal influenza A H3N2 viruses are constantly changing, reducing the effectiveness of existing vaccines. As a result, the World Health Organization (WHO) needs to frequently update the vaccine strains to match the antigenicity of emerged H3N2 variants. Traditional assessments of antigenicity rely on serological methods, which are both labor-intensive and time-consuming. Although numerous computational models aim to simplify antigenicity determination, they either lack a robust quantitative linkage between antigenicity and viral sequences or focus restrictively on selected features. Here, we propose a novel computational method to predict antigenic distances using multiple features, including not only viral sequence attributes but also integrating four distinct categories of features that significantly affect viral antigenicity in sequences. This method exhibits low error in virus antigenicity prediction and achieves superior accuracy in discerning antigenic drift. Utilizing this method, we investigated the evolution process of the H3N2 influenza viruses and identified a total of 21 major antigenic clusters from 1968 to 2022. Interestingly, our predicted antigenic map aligns closely with the antigenic map generated with serological data. Thus, our method is a promising tool for detecting antigenic variants and guiding the selection of vaccine candidates.

Machine Learning Sequence Modeling Identifies Gene Regulatory Responses to Bone Marrow Stromal Interactions in Multiple Myeloma

Introduction Multiple myeloma (MM) is a hematological malignancy of plasma cells which affects different organs including the bone marrow. The interaction between MM and the bone marrow stromal cells (BMSC) has been shown to affect the disease progression and response to treatment (Kumar et al. Multiple myeloma. Nat. Rev. Dis. Primers, 2017). An in vitro system to isolate the signaling mechanisms between BMSC and MM cells has been developed by (Dziadowicz et al., Cancers, 2022), who co-cultured MM cell lines in transwell media (TSW) exposed to BMSC secreted factors and compared their ATAC-seq profiles to MM cell lines (MONO) in isolation without the exposure to BMSC. While some of the gene expression changes induced by the BMSC-MM crosstalk were reported in (Dziadowicz et al., Cancers, 2022), here we model the entire chromatin accessibility profiles with a DNA-sequence based machine learning model to detect the altered Transcription Factor (TF) and enhancer activity responsible for driving the gene expression changes. We used sequence-based machine learning models (Beer et al., 2020, Annu. Rev. Genom. Hum. Genet.) and applied a systematic approach to model the chromatin accessibility changes induced by the interaction of BMSC and MM cells (MM-TSW) vs. MM cell lines cultured in isolation (MM-MONO). Methods We trained gapped-kmer SVM (gkm-SVM) machine learning models on MM1S and RPMI8226 MM cell line ATAC-seq data generated by Dziadowicz et al. as well as 1270 DNase-seq chromatin accessibility profiles of various human primary tissues and cell lines (including MM1S and RPMI8226) from the ENCODE consortium. We trained gkm-SVM on the top 2000 differentially-accessible distal ATAC peaks (>2000bp from transcription start site, e.g. enhancers) to find the differential TF activity between MM-TSW and MM-MONO (Figure 1) and found the TF binding site motifs explaining the gkmSVM output kmer weight distribution (Figure 2). We also identified the set of differentially-expressed TFs and performed gene set enrichment using MSigDB. Results We found that RUNX and Ebox-binding (5‘-CACCTG-3‘) TF family members are the most active TFs in MM1S-MONO and RPMI8226-MONO enhancers (gkmSVM AUROCs~0.92). The contribution of the Ebox TFs to distal chromatin accessibility in the MM cell lines is greater than in any other ENCODE sample (n=1270). Principal Component Analysis (PCA) on distal ATAC peaks compared to promoters showed that the MM transcriptional response to BMSC is more clearly explained by the differences in enhancer activity rather than promoters. This response is driven by dysregulated activity of a set of enhancer regions in MM-TSW compared to MM-MONO (Figure 1) and our machine learning method can identify these regulatory changes (AUROC~0.90). In addition to the altered AP1 and JAK/STAT activity reported by Dziadowicz et al., we discovered that RUNX and Ebox-binding transcriptional activity are downregulated in MM cells following the interaction with BMSC (Figure 2). Although our RNA-seq analysis found differential expression of some Ebox-binding TFs such as TCF3 and ZEB1 and RUNX members, the identification of the exact TF member(s) driving these RUNX and Ebox/E2A downregulation in MM-TSW is still challenging. Furthermore, PCA analysis of the gene expression profiles of MM1S and RPMI8226 cells showed that the targets of Bone Morphogenetic Protein 2 (BMP2) are upregulated in MM-TSW (FDR < 10 -28). Conclusion While some of the gene expression changes induced by the BMSC-MM interaction were reported in (Dziadowicz et al., Cancers, 2022), our modeling of the entire chromatin landscape discovered the previously unreported downregulation of RUNX and Ebox/E2A TFs in MM as the dominant response of MM interacting with BMSCs, accompanied by an upregulation of AP1 and JAK/STAT. RUNX and E2A-binding families of TFs include RUNX1/2/3, ZEB1/2, SNAI1/2 and TCF3, some of which we previously found to be strongly involved in epithelial to mesenchymal transition in gastric cancer and its invasiveness (Razavi-Mohseni et al., submitted, 2023). Bone marrow microenvironment interactions with MM are known to affect the disease progression and treatment, therefore, identifying the set of TFs driving the MM response to stroma may provide better understanding of the disease and better therapeutic opportunities.

Ordering taxa in image convolution networks improves microbiome-based machine learning accuracy

ABSTRACT The human gut microbiome is associated with a large number of disease etiologies. As such, it is a natural candidate for machine-learning-based biomarker development for multiple diseases and conditions. The microbiome is often analyzed using 16S rRNA gene sequencing or shotgun metagenomics. However, several properties of microbial sequence-based studies hinder machine learning (ML), including non-uniform representation, a small number of samples compared with the dimension of each sample, and sparsity of the data, with the majority of taxa present in a small subset of samples. We show here using a graph representation that the cladogram structure is as informative as the taxa frequency. We then suggest a novel method to combine information from different taxa and improve data representation for ML using microbial taxonomy. iMic (image microbiome) translates the microbiome to images through an iterative ordering scheme, and applies convolutional neural networks to the resulting image. We show that iMic has a higher precision in static microbiome gene sequence-based ML than state-of-the-art methods. iMic also facilitates the interpretation of the classifiers through an explainable artificial intelligence (AI) algorithm to iMic to detect taxa relevant to each condition. iMic is then extended to dynamic microbiome samples by translating them to movies.

Sequence-based machine learning method for predicting the effects of phosphorylation on protein-protein interactions

Protein phosphorylation, catalyzed by kinases, is an important biochemical process, which plays an essential role in multiple cell signaling pathways. Meanwhile, protein-protein interactions (PPI) constitute the signaling pathways. Abnormal phosphorylation status on protein can regulate protein functions through PPI to evoke severe diseases, such as Cancer and Alzheimer's disease. Due to the limited experimental evidence and high costs to experimentally identify novel evidence of phosphorylation regulation on PPI, it is necessary to develop a high-accuracy and user-friendly artificial intelligence method to predict phosphorylation effect on PPI. Here, we proposed a novel sequence-based machine learning method named PhosPPI, which achieved better identification performance (Accuracy and AUC) than other competing predictive methods of Betts, HawkDock and FoldX. PhosPPI is now freely available in web server (https://phosppi.sjtu.edu.cn/). This tool can help the user to identify functional phosphorylation sites affecting PPI and explore phosphorylation-associated disease mechanism and drug development.

IASMP: An interpretable in-silico predictive tool focusing on species-specific antimicrobial peptides.

Antimicrobial peptides (AMPs), a crucial part of the innate immune system, have been exploited as promising candidates for antibacterial agents. Many researchers have been devoting their efforts to develop novel AMPs in recent decades. In this term, many computational approaches have been developed to identify potential AMPs accurately. However, finding peptides specific to a particular bacterial species is challenging. Streptococcus mutans is a pathogen with an apparent cariogenic effect, and it is of great significance to study AMP that inhibit S.mutans for the prevention and treatment of caries. In this study, we proposed a sequence-based machine learning model, namely iASMP, to exactly identify potential anti-S.mutans peptides (ASMPs). After collecting ASMPs, the performances of models were compared by utilizing multiple feature descriptors and different classification algorithms. Among the baseline predictors, the model integrating the extra trees (ET) algorithm and the hybrid features exhibited optimal results. The feature selection method was utilized to remove redundant feature information to improve the model performance further. Finally, the proposed model achieved the maximum accuracy (ACC) of 0.962 on the training dataset and performed on the testing dataset with an ACC of 0.750. The results demonstrated that iASMP had an excellent predictive performance and was suitable for identifying potential ASMP. Furthermore, we also visualized the selected features and rationally explained the impact of individual features on the model output.

Using random forest to predict antimicrobial minimum inhibitory concentrations of nontyphoidal Salmonella in Taiwan

Antimicrobial resistance (AMR) is a global health issue and surveillance of AMR can be useful for understanding AMR trends and planning intervention strategies. Salmonella, widely distributed in food-producing animals, has been considered the first priority for inclusion in the AMR surveillance program by the World Health Organization (WHO). Recent advances in rapid and affordable whole-genome sequencing (WGS) techniques lead to the emergence of WGS as a one-stop test to predict the antimicrobial susceptibility. Since the variation of sequencing and minimum inhibitory concentration (MIC) measurement methods could result in different results, this study aimed to develop WGS-based random forest models for predicting MIC values of 24 drugs using data generated from the same laboratories in Taiwan. The WGS data have been transformed as a feature vector of 10-mers for machine learning. Based on rigorous validation and independent tests, a good performance was obtained with an average mean absolute error (MAE) less than 1 for both validation and independent test. Feature selection was then applied to identify top-ranked 10-mers that can further improve the prediction performance. For surveillance purposes, the genome sequence-based machine learning methods could be utilized to monitor the difference between predicted and experimental MIC, where a large difference might be worthy of investigation on the emerging genomic determinants.

Development of machine learning models for prediction of antibody non-specificity.

For the development of therapeutic antibodies, it is acknowledged that optimal target binding affinity should be accompanied by high specificity. In recent years, there has been an increased focus on lowering non-specific binding during early-stage drug development. In particular, non-specificity has been recognised as a root-cause for failure in numerous antibody-based drug programs due to unexpected pharmacokinetics and elevated toxicity. Hence, there is a need for predictive tools for flagging and deselecting non-specific antibodies at an early stage to mitigate failure at clinical late-stage development. From a computational design perspective, prediction has remained a challenge as non-specificity is currently a poorly understood biophysical phenomena, typically involving weak interactions to various, unrelated targets and surfaces. In this work, we describe the development of sequence-based machine learning (ML) models for prediction of antibody non-specificity. A dataset of >1000 antibody sequences paired with their respective non-specificity flag, defined by ELISA using a panel of common antigens, were retrieved from public sources (Jain et al., 2017; Boughter et al., 2020). The sequences were embedded using state-of-the-art protein language models for representation of their physiochemical properties, whereafter the vectorized embeddings were served as features for training of binary classifiers for non-specificity flags. A total of 12 different antibody domain-specific models were generated. Highest predictability of non-specificity was obtained for ML algorithms trained on the heavy-chain variable domain (VH), and the complementarity determining regions (CDRs), and resulted in a mean accuracy of 73% across 3-, 5-, and 10-fold cross-validations. The classification probability of non-specificity followed an uptrend with the experimental data (non-specificity flags), thereby mimicking regression behaviour. Altogether, our work demonstrates feasibility for prediction of non-specific binding from the antibody sequence, and that the non-specificity primarily stems from the VH domain, with main contributions from the heavy-chain CDRs.

Identification of Phase Separating Proteins With Distributed Reduced Alphabet Representations of Sequences.

Phase separation of proteins play key roles in cellular physiology including bacterial division, tumorigenesis etc. Consequently, understanding the molecular forces that drive phase separation has gained considerable attention and several factors including hydrophobicity, protein dynamics, etc., have been implicated in phase separation. Data-driven identification of new phase separating proteins can enable in-depth understanding of cellular physiology and may pave way towards developing novel methods of tackling disease progression. In this work, we exploit the existing wealth of data on phase separating proteins to develop sequence-based machine learning method for prediction of phase separating proteins. We use reduced alphabet schemes based on hydrophobicity and conformational similarity along with distributed representation of protein sequences and biochemical properties as input features to Support Vector Machine (SVM) and Random Forest (RF) machine learning algorithms. We used both curated and balanced dataset for building the models. RF trained on balanced dataset with hydropathy, conformational similarity embeddings and biochemical properties achieved accuracy of 97%. Our work highlights the use of conformational similarity, a feature that reflects amino acid flexibility, and hydrophobicity for predicting phase separating proteins. Use of such "interpretable" features obtained from the ever-growing knowledgebase of phase separation is likely to improve prediction performances further.

Quality assessment and refinement of chromatin accessibility data using a sequence-based predictive model

Chromatin accessibility assays are central to the genome-wide identification of gene regulatory elements associated with transcriptional regulation. However, the data have highly variable quality arising from several biological and technical factors. To surmount this problem, we developed a sequence-based machine learning method to evaluate and refine chromatin accessibility data. Our framework, gapped k-mer SVM quality check (gkmQC), provides the quality metrics for a sample based on the prediction accuracy of the trained models. We tested 886 DNase-seq samples from the ENCODE/Roadmap projects to demonstrate that gkmQC can effectively identify "high-quality" (HQ) samples with low conventional quality scores owing to marginal read depths. Peaks identified in HQ samples are more accurately aligned at functional regulatory elements, show greater enrichment of regulatory elements harboring functional variants, and explain greater heritability of phenotypes from their relevant tissues. Moreover, gkmQC can optimize the peak-calling threshold to identify additional peaks, especially for rare cell types in single-cell chromatin accessibility data.

DENVIS: Scalable and High-Throughput Virtual Screening Using Graph Neural Networks with Atomic and Surface Protein Pocket Features.

Computational methods for virtual screening can dramatically accelerate early-stage drug discovery by identifying potential hits for a specified target. Docking algorithms traditionally use physics-based simulations to address this challenge by estimating the binding orientation of a query protein-ligand pair and a corresponding binding affinity score. Over the recent years, classical and modern machine learning architectures have shown potential for outperforming traditional docking algorithms. Nevertheless, most learning-based algorithms still rely on the availability of the protein-ligand complex binding pose, typically estimated via docking simulations, which leads to a severe slowdown of the overall virtual screening process. A family of algorithms processing target information at the amino acid sequence level avoid this requirement, however, at the cost of processing protein data at a higher representation level. We introduce deep neural virtual screening (DENVIS), an end-to-end pipeline for virtual screening using graph neural networks (GNNs). By performing experiments on two benchmark databases, we show that our method performs competitively to several docking-based, machine learning-based, and hybrid docking/machine learning-based algorithms. By avoiding the intermediate docking step, DENVIS exhibits several orders of magnitude faster screening times (i.e., higher throughput) than both docking-based and hybrid models. When compared to an amino acid sequence-based machine learning model with comparable screening times, DENVIS achieves dramatically better performance. Some key elements of our approach include protein pocket modeling using a combination of atomic and surface features, the use of model ensembles, and data augmentation via artificial negative sampling during model training. In summary, DENVIS achieves competitive to state-of-the-art virtual screening performance, while offering the potential to scale to billions of molecules using minimal computational resources.

Investigating Autoregressive and Machine Learning-based Time Series Modeling with Dielectric Spectroscopy for Predicting Quality of Biofabricated Constructs

Advances in biofabrication processes need to be complemented with appropriate nondestructive quality engineering techniques that can be integrated into scalable engineered tissue manufacturing systems. Previous studies have demonstrated the feasibility of dielectric spectroscopy (DS) as a inline, real time biological quality monitoring alternative. Time series modeling can help improve the efficiency and accuracy of quality prediction by analyzing trends in DS data as the biofabricated constructs mature over time. These models can help forecast potential future deviations in quality attributes and provide opportunities to take preemptive, corrective actions, leading to better yields and higher quality of final products. In this study, we investigated time series modeling of DS data to characterize the effects of two critical biofabrication parameters on constructs of gelatin methacryloyl (GelMA) hydrogel containing human adipose-derived stem cells (hASC) over 11 days of in vitro culture. The performance of standard autoregressive time series models (Exponential Smoothing, ARMA, ARIMA, SARIMA) and conventional sequence-based machine learning (ML) models (SVM, ANN, CNN and LSTM) were analyzed to forecast trends in Δɛ, a key DS metric that directly correlates to the volume of viable cells in constructs. The ML-based time series models, in general, showed superior performance in predicting future trends in Δɛ, with LSTM providing the lowest least mean square errors (MSE) in Δɛ forecasts. The outcomes of this study highlight the benefits of using DS and time series modeling synergistically for efficient quality monitoring in biofabrication.

SeBPPI: A Sequence-Based Protein–Protein Binding Predictor

Protein–protein interaction (PPI) plays an important role in biological processes such as signal transduction, immune response and membrane transport, etc. In this work, a protein sequence-based machine learning model, SeBPPI, to predict protein–protein binding was proposed. In this model, the descriptors were generated from three pre-trained models: Unirep, ESM and TAPE. The performance of SeBPPI with these pre-trained models was evaluated on several different test datasets. The accuracy of our binary prediction model shows improvement over the existing methods. We also compared the performance of two classification heads: The Recurrent convolution neural network (RCNN) and the fully connected neural network (FNN) and found that the use of RCNN is beneficial for the overall improvement in the accuracy of the model. This study helps to improve the accuracy in sequence-based protein–protein binding predictions. The model used in this work is integrated in the web server http://www.icdrug.com/ICDrug/SeBPPI .

Support vector machine-based prediction of pore-forming toxins (PFT) using distributed representation of reduced alphabets

Bacterial virulence can be attributed to a wide variety of factors including toxins that harm the host. Pore-forming toxins are one class of toxins that confer virulence to the bacteria and are one of the promising targets for therapeutic intervention. In this work, we develop a sequence-based machine learning framework for the prediction of pore-forming toxins. For this, we have used distributed representation of the protein sequence encoded by reduced alphabet schemes based on conformational similarity and hydropathy index as input features to Support Vector Machines (SVMs). The choice of conformational similarity and hydropathy indices is based on the functional mechanism of pore-forming toxins. Our methodology achieves about 81% accuracy indicating that conformational similarity, an indicator of the flexibility of amino acids, along with hydrophobic index can capture the intrinsic features of pore-forming toxins that distinguish it from other types of transporter proteins. Increased understanding of the mechanisms of pore-forming toxins can further contribute to the use of such "mechanism-informed" features that may increase the prediction accuracy further.

Systematic Evaluation of DNA Sequence Variations on in vivo Transcription Factor Binding Affinity.

The majority of the single nucleotide variants (SNVs) identified by genome-wide association studies (GWAS) fall outside of the protein-coding regions. Elucidating the functional implications of these variants has been a major challenge. A possible mechanism for functional non-coding variants is that they disrupted the canonical transcription factor (TF) binding sites that affect the in vivo binding of the TF. However, their impact varies since many positions within a TF binding motif are not well conserved. Therefore, simply annotating all variants located in putative TF binding sites may overestimate the functional impact of these SNVs. We conducted a comprehensive survey to study the effect of SNVs on the TF binding affinity. A sequence-based machine learning method was used to estimate the change in binding affinity for each SNV located inside a putative motif site. From the results obtained on 18 TF binding motifs, we found that there is a substantial variation in terms of a SNV’s impact on TF binding affinity. We found that only about 20% of SNVs located inside putative TF binding sites would likely to have significant impact on the TF-DNA binding.

LSTM-PHV: prediction of human-virus protein-protein interactions by LSTM with word2vec.

Viral infection involves a large number of protein–protein interactions (PPIs) between human and virus. The PPIs range from the initial binding of viral coat proteins to host membrane receptors to the hijacking of host transcription machinery. However, few interspecies PPIs have been identified, because experimental methods including mass spectrometry are time-consuming and expensive, and molecular dynamic simulation is limited only to the proteins whose 3D structures are solved. Sequence-based machine learning methods are expected to overcome these problems. We have first developed the LSTM model with word2vec to predict PPIs between human and virus, named LSTM-PHV, by using amino acid sequences alone. The LSTM-PHV effectively learnt the training data with a highly imbalanced ratio of positive to negative samples and achieved AUCs of 0.976 and 0.973 and accuracies of 0.984 and 0.985 on the training and independent datasets, respectively. In predicting PPIs between human and unknown or new virus, the LSTM-PHV learned greatly outperformed the existing state-of-the-art PPI predictors. Interestingly, learning of only sequence contexts as words is sufficient for PPI prediction. Use of uniform manifold approximation and projection demonstrated that the LSTM-PHV clearly distinguished the positive PPI samples from the negative ones. We presented the LSTM-PHV online web server and support data that are freely available at http://kurata35.bio.kyutech.ac.jp/LSTM-PHV.

Sequence-Based Prediction of Transmembrane Protein Crystallization Propensity.

Transmembrane proteins play a vital role in cell life activities. There are several techniques to determine transmembrane protein structures and X-ray crystallography is the primary methodology. However, due to the special properties of transmembrane proteins, it is still hard to determine their structures by X-ray crystallography technique. To reduce experimental consumption and improve experimental efficiency, it is of great significance to develop computational methods for predicting the crystallization propensity of transmembrane proteins. In this work, we proposed a sequence-based machine learning method, namely Prediction of TransMembrane protein Crystallization propensity (PTMC), to predict the propensity of transmembrane protein crystallization. First, we obtained several general sequence features and the specific encoded features of relative solvent accessibility and hydrophobicity. Second, feature selection was employed to filter out redundant and irrelevant features, and the optimal feature subset is composed of hydrophobicity, amino acid composition and relative solvent accessibility. Finally, we chose extreme gradient boosting by comparing with other several machine learning methods. Comparative results on the independent test set indicate that PTMC outperforms state-of-the-art sequence-based methods in terms of sensitivity, specificity, accuracy, Matthew's Correlation Coefficient (MCC) and Area Under the receiver operating characteristic Curve (AUC). In comparison with two competitors, Bcrystal and TMCrys, PTMC achieves an improvement by 0.132 and 0.179 for sensitivity, 0.014 and 0.127 for specificity, 0.037 and 0.192 for accuracy, 0.128 and 0.362 for MCC, and 0.027 and 0.125 for AUC, respectively.