Multi-class protein fold recognition using support vector machines and neural networks

MOTIVATION: Protein fold recognition is an important approach to structure discovery without relying on sequence similarity. We study this approach with new multi-class classification methods and examined many issues important for a practical recognition system. RESULTS: Most current discriminative methods for protein fold prediction use the one-against-others method, which has the well-known 'False Positives' problem. We investigated two new methods: the unique one-against-others and the all-against-all methods. Both improve prediction accuracy by 14-110% on a dataset containing 27 SCOP folds. We used the Support Vector Machine (SVM) and the Neural Network (NN) learning methods as base classifiers. SVMs converges fast and leads to high accuracy. When scores of multiple parameter datasets are combined, majority voting reduces noise and increases recognition accuracy. We examined many issues involved with large number of classes, including dependencies of prediction accuracy on the number of folds and on the number of representatives in a fold. Overall, recognition systems achieve 56% fold prediction accuracy on a protein test dataset, where most of the proteins have below 25% sequence identity with the proteins used in training.     

SCGPred： A Score-based Method for Gene Structure Prediction by Combining Multiple Sources of Evidence

Predicting protein-coding genes still remains a significant challenge. Although a variety of computational programs that use commonly machine learning methods have emerged, the accuracy of predictions remains a low level when implementing in large genomic sequences. Moreover, computational gene finding in newly se- quenced genomes is especially a difficult task due to the absence of a training set of abundant validated genes. Here we present a new gene-finding program, SCGPred, to improve the accuracy of prediction by combining multiple sources of evidence. SCGPred can perform both supervised method in previously well-studied genomes and unsupervised one in novel genomes. By testing with datasets composed of large DNA sequences from human and a novel genome of Ustilago maydi, SCGPred gains a significant improvement in comparison to the popular ab initio gene predictors. We also demonstrate that SCGPred can significantly improve prediction in novel genomes by combining several foreign gene finders with similarity alignments, which is superior to other unsupervised methods. Therefore, SCGPred can serve as an alternative gene-finding tool for newly sequenced eukaryotic genomes. The program is freely available at http：//bio.scu.edu.cn/SCGPred/.     

Dragon PolyA Spotter: predictor of poly(A) motifs within human genomic DNA sequences

MOTIVATION: Recognition of poly(A) signals in mRNA is relatively straightforward due to the presence of easily recognizable polyadenylic acid tail. However, the task of identifying poly(A) motifs in the primary genomic DNA sequence that correspond to poly(A) signals in mRNA is a far more challenging problem. Recognition of poly(A) signals is important for better gene annotation and understanding of the gene regulation mechanisms. In this work, we present one such poly(A) motif prediction method based on properties of human genomic DNA sequence surrounding a poly(A) motif. These properties include thermodynamic, physico-chemical and statistical characteristics. For predictions, we developed Artificial Neural Network and Random Forest models. These models are trained to recognize 12 most common poly(A) motifs in human DNA. Our predictors are available as a free web-based tool accessible at http://cbrc.kaust.edu.sa/dps. Compared with other reported predictors, our models achieve higher sensitivity and specificity and furthermore provide a consistent level of accuracy for 12 poly(A) motif variants. CONTACT: vladimir.bajic@kaust.edu.sa SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.     

Analysis of protein features and machine learning algorithms for prediction of druggable proteins

Background: Computational tools have been widely used in drug discovery process since they reduce the time and cost. Prediction of whether a protein is druggable is fundamental and crucial for drug research pipeline. Sequence based protein function prediction plays vital roles in many research areas. Training data, protein features selection and machine learning algorithms are three indispensable elements that drive the successfulness of the models. Methods: In this study, we tested the performance of different combinations of protein features and machine learning algorithms, based on FDA-approved small molecules’ targets, in druggable proteins prediction. We also enlarged the dataset to include the targets of small molecules that were in experiment or clinical investigation. Results: We found that although the 146-d vector used by Li et al. with neuron network achieved the best training accuracy of 91.10%, overlapped 3-gram word2vec with logistic regression achieved best prediction accuracy on independent test set (89.55%) and on newly approved-targets. Enlarged dataset with targets of small molecules in experiment and clinical investigation were trained. Unfortunately, the best training accuracy was only 75.48%. In addition, we applied our models to predict potential targets for references in future study. Conclusions: Our study indicates the potential ability of word2vec in the prediction of druggable protein. And the training dataset of druggable protein should not be extended to targets that are lack of verification. The target prediction package could be found on https://github.com/pkumdl/target_prediction.     

Protein-protein interaction site predictions with three-dimensional probability distributions of interacting atoms on protein surfaces

Protein-protein interactions are key to many biological processes. Computational methodologies devised to predict protein-protein interaction (PPI) sites on protein surfaces are important tools in providing insights into the biological functions of proteins and in developing therapeutics targeting the protein-protein interaction sites. One of the general features of PPI sites is that the core regions from the two interacting protein surfaces are complementary to each other, similar to the interior of proteins in packing density and in the physicochemical nature of the amino acid composition. In this work, we simulated the physicochemical complementarities by constructing three-dimensional probability density maps of non-covalent interacting atoms on the protein surfaces. The interacting probabilities were derived from the interior of known structures. Machine learning algorithms were applied to learn the characteristic patterns of the probability density maps specific to the PPI sites. The trained predictors for PPI sites were cross-validated with the training cases (consisting of 432 proteins) and were tested on an independent dataset (consisting of 142 proteins). The residue-based Matthews correlation coefficient for the independent test set was 0.423; the accuracy, precision, sensitivity, specificity were 0.753, 0.519, 0.677, and 0.779 respectively. The benchmark results indicate that the optimized machine learning models are among the best predictors in identifying PPI sites on protein surfaces. In particular, the PPI site prediction accuracy increases with increasing size of the PPI site and with increasing hydrophobicity in amino acid composition of the PPI interface; the core interface regions are more likely to be recognized with high prediction confidence. The results indicate that the physicochemical complementarity patterns on protein surfaces are important determinants in PPIs, and a substantial portion of the PPI sites can be predicted correctly with the physicochemical complementarity features based on the non-covalent interaction data derived from protein interiors.     

Explainable deep transfer learning model for disease risk prediction using high-dimensional genomic data

Building an accurate disease risk prediction model is an essential step in the modern quest for precision medicine. While high-dimensional genomic data provides valuable data resources for the investigations of disease risk, their huge amount of noise and complex relationships between predictors and outcomes have brought tremendous analytical challenges. Deep learning model is the state-of-the-art methods for many prediction tasks, and it is a promising framework for the analysis of genomic data. However, deep learning models generally suffer from the curse of dimensionality and the lack of biological interpretability, both of which have greatly limited their applications. In this work, we have developed a deep neural network (DNN) based prediction modeling framework. We first proposed a group-wise feature importance score for feature selection, where genes harboring genetic variants with both linear and non-linear effects are efficiently detected. We then designed an explainable transfer-learning based DNN method, which can directly incorporate information from feature selection and accurately capture complex predictive effects. The proposed DNN-framework is biologically interpretable, as it is built based on the selected predictive genes. It is also computationally efficient and can be applied to genome-wide data. Through extensive simulations and real data analyses, we have demonstrated that our proposed method can not only efficiently detect predictive features, but also accurately predict disease risk, as compared to many existing methods.     

CONTRAfold: RNA secondary structure prediction without physics-based models

MOTIVATION: For several decades, free energy minimization methods have been the dominant strategy for single sequence RNA secondary structure prediction. More recently, stochastic context-free grammars (SCFGs) have emerged as an alternative probabilistic methodology for modeling RNA structure. Unlike physics-based methods, which rely on thousands of experimentally-measured thermodynamic parameters, SCFGs use fully-automated statistical learning algorithms to derive model parameters. Despite this advantage, however, probabilistic methods have not replaced free energy minimization methods as the tool of choice for secondary structure prediction, as the accuracies of the best current SCFGs have yet to match those of the best physics-based models. RESULTS: In this paper, we present CONTRAfold, a novel secondary structure prediction method based on conditional log-linear models (CLLMs), a flexible class of probabilistic models which generalize upon SCFGs by using discriminative training and feature-rich scoring. In a series of cross-validation experiments, we show that grammar-based secondary structure prediction methods formulated as CLLMs consistently outperform their SCFG analogs. Furthermore, CONTRAfold, a CLLM incorporating most of the features found in typical thermodynamic models, achieves the highest single sequence prediction accuracies to date, outperforming currently available probabilistic and physics-based techniques. Our result thus closes the gap between probabilistic and thermodynamic models, demonstrating that statistical learning procedures provide an effective alternative to empirical measurement of thermodynamic parameters for RNA secondary structure prediction. AVAILABILITY: Source code for CONTRAfold is available at http://contra.stanford.edu/contrafold/.     

Using Genetic Distance to Infer the Accuracy of Genomic Prediction

The prediction of phenotypic traits using high-density genomic data has many applications such as the selection of plants and animals of commercial interest; and it is expected to play an increasing role in medical diagnostics. Statistical models used for this task are usually tested using cross-validation, which implicitly assumes that new individuals (whose phenotypes we would like to predict) originate from the same population the genomic prediction model is trained on. In this paper we propose an approach based on clustering and resampling to investigate the effect of increasing genetic distance between training and target populations when predicting quantitative traits. This is important for plant and animal genetics, where genomic selection programs rely on the precision of predictions in future rounds of breeding. Therefore, estimating how quickly predictive accuracy decays is important in deciding which training population to use and how often the model has to be recalibrated. We find that the correlation between true and predicted values decays approximately linearly with respect to either F_ST or mean kinship between the training and the target populations. We illustrate this relationship using simulations and a collection of data sets from mice, wheat and human genetics.     

Protein tertiary structure modeling driven by deep learning and contact distance prediction in CASP13

Predicting residue-residue distance relationships (eg, contacts) has become the key direction to advance protein structure prediction since 2014 CASP11 experiment, while deep learning has revolutionized the technology for contact and distance distribution prediction since its debut in 2012 CASP10 experiment. During 2018 CASP13 experiment, we enhanced our MULTICOM protein structure prediction system with three major components: contact distance prediction based on deep convolutional neural networks, distance-driven template-free (ab initio) modeling, and protein model ranking empowered by deep learning and contact prediction. Our experiment demonstrates that contact distance prediction and deep learning methods are the key reasons that MULTICOM was ranked 3rd out of all 98 predictors in both template-free and template-based structure modeling in CASP13. Deep convolutional neural network can utilize global information in pairwise residue-residue features such as coevolution scores to substantially improve contact distance prediction, which played a decisive role in correctly folding some free modeling and hard template-based modeling targets. Deep learning also successfully integrated one-dimensional structural features, two-dimensional contact information, and three-dimensional structural quality scores to improve protein model quality assessment, where the contact prediction was demonstrated to consistently enhance ranking of protein models for the first time. The success of MULTICOM system clearly shows that protein contact distance prediction and model selection driven by deep learning holds the key of solving protein structure prediction problem. However, there are still challenges in accurately predicting protein contact distance when there are few homologous sequences, folding proteins from noisy contact distances, and ranking models of hard targets.     

Prediction of carbohydrate binding sites on protein surfaces with 3-dimensional probability density distributions of interacting atoms

Non-covalent protein-carbohydrate interactions mediate molecular targeting in many biological processes. Prediction of non-covalent carbohydrate binding sites on protein surfaces not only provides insights into the functions of the query proteins; information on key carbohydrate-binding residues could suggest site-directed mutagenesis experiments, design therapeutics targeting carbohydrate-binding proteins, and provide guidance in engineering protein-carbohydrate interactions. In this work, we show that non-covalent carbohydrate binding sites on protein surfaces can be predicted with relatively high accuracy when the query protein structures are known. The prediction capabilities were based on a novel encoding scheme of the three-dimensional probability density maps describing the distributions of 36 non-covalent interacting atom types around protein surfaces. One machine learning model was trained for each of the 30 protein atom types. The machine learning algorithms predicted tentative carbohydrate binding sites on query proteins by recognizing the characteristic interacting atom distribution patterns specific for carbohydrate binding sites from known protein structures. The prediction results for all protein atom types were integrated into surface patches as tentative carbohydrate binding sites based on normalized prediction confidence level. The prediction capabilities of the predictors were benchmarked by a 10-fold cross validation on 497 non-redundant proteins with known carbohydrate binding sites. The predictors were further tested on an independent test set with 108 proteins. The residue-based Matthews correlation coefficient (MCC) for the independent test was 0.45, with prediction precision and sensitivity (or recall) of 0.45 and 0.49 respectively. In addition, 111 unbound carbohydrate-binding protein structures for which the structures were determined in the absence of the carbohydrate ligands were predicted with the trained predictors. The overall prediction MCC was 0.49. Independent tests on anti-carbohydrate antibodies showed that the carbohydrate antigen binding sites were predicted with comparable accuracy. These results demonstrate that the predictors are among the best in carbohydrate binding site predictions to date.     

The importance of peptide detectability for protein identification, quantification, and experiment design in MS/MS proteomics

Peptide detectability is defined as the probability that a peptide is identified in an LC-MS/MS experiment and has been useful in providing solutions to protein inference and label-free quantification. Previously, predictors for peptide detectability trained on standard or complex samples were proposed. Although the models trained on complex samples may benefit from the large training data sets, it is unclear to what extent they are affected by the unequal abundances of identified proteins. To address this challenge and improve detectability prediction, we present a new algorithm for the iterative learning of peptide detectability from complex mixtures. We provide evidence that the new method approximates detectability with useful accuracy and, based on its design, can be used to interpret the outcome of other learning strategies. We studied the properties of peptides from the bacterium Deinococcus radiodurans and found that at standard quantities, its tryptic peptides can be roughly classified as either detectable or undetectable, with a relatively small fraction having medium detectability. We extend the concept of detectability from peptides to proteins and apply the model to predict the behavior of a replicate LC-MS/MS experiment from a single analysis. Finally, our study summarizes a theoretical framework for peptide/protein identification and label-free quantification.     

Tumor cell intrinsic and extrinsic features predict prognosis in estrogen receptor positive breast cancer

Although estrogen-receptor-positive (ER+) breast cancer is generally associated with favorable prognosis, clinical outcome varies substantially among patients. Genomic assays have been developed and applied to predict patient prognosis for personalized treatment. We hypothesize that the recurrence risk of ER+ breast cancer patients is determined by both genomic mutations intrinsic to tumor cells and extrinsic immunological features in the tumor microenvironment. Based on the Cancer Genome Atlas (TCGA) breast cancer data, we identified the 72 most common genomic aberrations (including gene mutations and indels) in ER+ breast cancer and defined sample-specific scores that systematically characterized the deregulated pathways intrinsic to tumor cells. To further consider tumor cell extrinsic features, we calculated immune infiltration scores for six major immune cell types. Many individual intrinsic features are predictive of patient prognosis in ER+ breast cancer, and some of them achieved comparable accuracy with the Oncotype DX assay. In addition, statistical learning models that integrated these features predicts the recurrence risk of patients with significantly better performance than the Oncotype DX assay (our optimized random forest model AUC = 0.841, Oncotype DX model AUC = 0.792, p = 0.04). As a proof-of-concept, our study indicates the great potential of genomic and immunological features in prognostic prediction for improving breast cancer precision medicine. The framework introduced in this work can be readily applied to other cancers.     

Selective prediction-set models with coverage rate guarantees

The current approach to using machine learning (ML) algorithms in healthcare is to either require clinician oversight for every use case or use their predictions without any human oversight. We explore a middle ground that lets ML algorithms abstain from making a prediction to simultaneously improve their reliability and reduce the burden placed on human experts. To this end, we present a general penalized loss minimization framework for training selective prediction-set (SPS) models, which choose to either output a prediction set or abstain. The resulting models abstain when the outcome is difficult to predict accurately, such as on subjects who are too different from the training data, and achieve higher accuracy on those they do give predictions for. We then introduce a model-agnostic, statistical inference procedure for the coverage rate of an SPS model that ensembles individual models trained using K-fold cross-validation. We find that SPS ensembles attain prediction-set coverage rates closer to the nominal level and have narrower confidence intervals for its marginal coverage rate. We apply our method to train neural networks that abstain more for out-of-sample images on the MNIST digit prediction task and achieve higher predictive accuracy for ICU patients compared to existing approaches.     

Small angle X-ray scattering-assisted protein structure prediction in CASP13 and emergence of solution structure differences

Small angle X-ray scattering (SAXS) measures comprehensive distance information on a protein's structure, which can constrain and guide computational structure prediction algorithms. Here, we evaluate structure predictions of 11 monomeric and oligomeric proteins for which SAXS data were collected and provided to predictors in the 13th round of the Critical Assessment of protein Structure Prediction (CASP13). The category for SAXS-assisted predictions made gains in certain areas for CASP13 compared to CASP12. Improvements included higher quality data with size exclusion chromatography-SAXS (SEC-SAXS) and better selection of targets and communication of results by CASP organizers. In several cases, we can track improvements in model accuracy with use of SAXS data. For hard multimeric targets where regular folding algorithms were unsuccessful, SAXS data helped predictors to build models better resembling the global shape of the target. For most models, however, no significant improvement in model accuracy at the domain level was registered from use of SAXS data, when rigorously comparing SAXS-assisted models to the best regular server predictions. To promote future progress in this category, we identify successes, challenges, and opportunities for improved strategies in prediction, assessment, and communication of SAXS data to predictors. An important observation is that, for many targets, SAXS data were inconsistent with crystal structures, suggesting that these proteins adopt different conformation(s) in solution. This CASP13 result, if representative of PDB structures and future CASP targets, may have substantive implications for the structure training databases used for machine learning, CASP, and use of prediction models for biology.     

Discriminative learning for protein conformation sampling

Protein structure prediction without using templates (i.e., ab initio folding) is one of the most challenging problems in structural biology. In particular, conformation sampling poses as a major bottleneck of ab initio folding. This article presents CRFSampler, an extensible protein conformation sampler, built on a probabilistic graphical model Conditional Random Fields (CRFs). Using a discriminative learning method, CRFSampler can automatically learn more than ten thousand parameters quantifying the relationship among primary sequence, secondary structure, and (pseudo) backbone angles. Using only compactness and self-avoiding constraints, CRFSampler can efficiently generate protein-like conformations from primary sequence and predicted secondary structure. CRFSampler is also very flexible in that a variety of model topologies and feature sets can be defined to model the sequence-structure relationship without worrying about parameter estimation. Our experimental results demonstrate that using a simple set of features, CRFSampler can generate decoys with much higher quality than the most recent HMM model.     

A range of complex probabilistic models for RNA secondary structure prediction that includes the nearest-neighbor model and more

The standard approach for single-sequence RNA secondary structure prediction uses a nearest-neighbor thermodynamic model with several thousand experimentally determined energy parameters. An attractive alternative is to use statistical approaches with parameters estimated from growing databases of structural RNAs. Good results have been reported for discriminative statistical methods using complex nearest-neighbor models, including CONTRAfold, Simfold, and ContextFold. Little work has been reported on generative probabilistic models (stochastic context-free grammars [SCFGs]) of comparable complexity, although probabilistic models are generally easier to train and to use. To explore a range of probabilistic models of increasing complexity, and to directly compare probabilistic, thermodynamic, and discriminative approaches, we created TORNADO, a computational tool that can parse a wide spectrum of RNA grammar architectures (including the standard nearest-neighbor model and more) using a generalized super-grammar that can be parameterized with probabilities, energies, or arbitrary scores. By using TORNADO, we find that probabilistic nearest-neighbor models perform comparably to (but not significantly better than) discriminative methods. We find that complex statistical models are prone to overfitting RNA structure and that evaluations should use structurally nonhomologous training and test data sets. Overfitting has affected at least one published method (ContextFold). The most important barrier to improving statistical approaches for RNA secondary structure prediction is the lack of diversity of well-curated single-sequence RNA secondary structures in current RNA databases.     

Genomic prediction based on data from three layer lines using non-linear regression models

Background

Most studies on genomic prediction with reference populations that include multiple lines or breeds have used linear models. Data heterogeneity due to using multiple populations may conflict with model assumptions used in linear regression methods.

Methods

In an attempt to alleviate potential discrepancies between assumptions of linear models and multi-population data, two types of alternative models were used: (1) a multi-trait genomic best linear unbiased prediction (GBLUP) model that modelled trait by line combinations as separate but correlated traits and (2) non-linear models based on kernel learning. These models were compared to conventional linear models for genomic prediction for two lines of brown layer hens (B1 and B2) and one line of white hens (W1). The three lines each had 1004 to 1023 training and 238 to 240 validation animals. Prediction accuracy was evaluated by estimating the correlation between observed phenotypes and predicted breeding values.

Results

When the training dataset included only data from the evaluated line, non-linear models yielded at best a similar accuracy as linear models. In some cases, when adding a distantly related line, the linear models showed a slight decrease in performance, while non-linear models generally showed no change in accuracy. When only information from a closely related line was used for training, linear models and non-linear radial basis function (RBF) kernel models performed similarly. The multi-trait GBLUP model took advantage of the estimated genetic correlations between the lines. Combining linear and non-linear models improved the accuracy of multi-line genomic prediction.

Conclusions

Linear models and non-linear RBF models performed very similarly for genomic prediction, despite the expectation that non-linear models could deal better with the heterogeneous multi-population data. This heterogeneity of the data can be overcome by modelling trait by line combinations as separate but correlated traits, which avoids the occasional occurrence of large negative accuracies when the evaluated line was not included in the training dataset. Furthermore, when using a multi-line training dataset, non-linear models provided information on the genotype data that was complementary to the linear models, which indicates that the underlying data distributions of the three studied lines were indeed heterogeneous.

Electronic supplementary material

The online version of this article (doi:10.1186/s12711-014-0075-3) contains supplementary material, which is available to authorized users.     

Gene recognition based on DAG shortest paths

We describe DAGGER, an ab initio gene recognition program which combines the output of high dimensional signal sensors in an intuitive gene model based on directed acyclic graphs. In the first stage, candidate start, donor, acceptor, and stop sites are scored using the SNoW learning architecture. These sites are then used to generate a directed acyclic graph in which each source-sink path represents a possible gene structure. Training sequences are used to optimize an edge weighting function so that the shortest source-sink path maximizes exon-level prediction accuracy. Experimental evaluation of prediction accuracy on two benchmark data sets demonstrates that DAGGERis competitive with ab initio gene finding programs based on Hidden Markov Models.