Accurate prediction of solvent accessibility using neural networks-based regression

Accurate prediction of relative solvent accessibilities (RSAs) of amino acid residues in proteins may be used to facilitate protein structure prediction and functional annotation. Toward that goal we developed a novel method for improved prediction of RSAs. Contrary to other machine learning-based methods from the literature, we do not impose a classification problem with arbitrary boundaries between the classes. Instead, we seek a continuous approximation of the real-value RSA using nonlinear regression, with several feed forward and recurrent neural networks, which are then combined into a consensus predictor. A set of 860 protein structures derived from the PFAM database was used for training, whereas validation of the results was carefully performed on several nonredundant control sets comprising a total of 603 structures derived from new Protein Data Bank structures and had no homology to proteins included in the training. Two classes of alternative predictors were developed for comparison with the regression-based approach: one based on the standard classification approach and the other based on a semicontinuous approximation with the so-called thermometer encoding. Furthermore, a weighted approximation, with errors being scaled by the observed levels of variability in RSA for equivalent residues in families of homologous structures, was applied in order to improve the results. The effects of including evolutionary profiles and the growth of sequence databases were assessed. In accord with the observed levels of variability in RSA for different ranges of RSA values, the regression accuracy is higher for buried than for exposed residues, with overall 15.3-15.8% mean absolute errors and correlation coefficients between the predicted and experimental values of 0.64-0.67 on different control sets. The new method outperforms classification-based algorithms when the real value predictions are projected onto two-class classification problems with several commonly used thresholds to separate exposed and buried residues. For example, classification accuracy of about 77% is consistently achieved on all control sets with a threshold of 25% RSA. A web server that enables RSA prediction using the new method and provides customizable graphical representation of the results is available at http://sable.cchmc.org.     

Solvers for the cardiac bidomain equations

The bidomain equations are widely used for the simulation of electrical activity in cardiac tissue. They are especially important for accurately modeling extracellular stimulation, as evidenced by their prediction of virtual electrode polarization before experimental verification. However, solution of the equations is computationally expensive due to the fine spatial and temporal discretization needed. This limits the size and duration of the problem which can be modeled. Regardless of the specific form into which they are cast, the computational bottleneck becomes the repeated solution of a large, linear system. The purpose of this review is to give an overview of the equations and the methods by which they have been solved. Of particular note are recent developments in multigrid methods, which have proven to be the most efficient.     

Two-stage support vector regression approach for predicting accessible surface areas of amino acids

We address the problem of predicting solvent accessible surface area (ASA) of amino acid residues in protein sequences, without classifying them into buried and exposed types. A two-stage support vector regression (SVR) approach is proposed to predict real values of ASA from the position-specific scoring matrices generated from PSI-BLAST profiles. By adding SVR as the second stage to capture the influences on the ASA value of a residue by those of its neighbors, the two-stage SVR approach achieves improvements of mean absolute errors up to 3.3%, and correlation coefficients of 0.66, 0.68, and 0.67 on the Manesh dataset of 215 proteins, the Barton dataset of 502 nonhomologous proteins, and the Carugo dataset of 338 proteins, respectively, which are better than the scores published earlier on these datasets. A Web server for protein ASA prediction by using a two-stage SVR method has been developed and is available (http://birc.ntu.edu.sg/~ pas0186457/asa.html).     

Ranking predicted protein structures with support vector regression

Protein structure prediction is an important problem of both intellectual and practical interest. Most protein structure prediction approaches generate multiple candidate models first, and then use a scoring function to select the best model among these candidates. In this work, we develop a scoring function using support vector regression (SVR). Both consensus-based features and features from individual structures are extracted from a training data set containing native protein structures and predicted structural models submitted to CASP5 and CASP6. The SVR learns a scoring function that is a linear combination of these features. We test this scoring function on two data sets. First, when used to rank server models submitted to CASP7, the SVR score selects predictions that are comparable to the best performing server in CASP7, Zhang-Server, and significantly better than all the other servers. Even if the SVR score is not allowed to select Zhang-Server models, the SVR score still selects predictions that are significantly better than all the other servers. In addition, the SVR is able to select significantly better models and yield significantly better Pearson correlation coefficients than the two best Quality Assessment groups in CASP7, QA556 (LEE), and QA634 (Pcons). Second, this work aims to improve the ability of the Robetta server to select best models, and hence we evaluate the performance of the SVR score on ranking the Robetta server template-based models for the CASP7 targets. The SVR selects significantly better models than the Robetta K*Sync consensus alignment score.     

Application of support vector regression to genome-assisted prediction of quantitative traits

A byproduct of genome-wide association studies is the possibility of carrying out genome-enabled prediction of disease risk or of quantitative traits. This study is concerned with predicting two quantitative traits, milk yield in dairy cattle and grain yield in wheat, using dense molecular markers as predictors. Two support vector regression (SVR) models, ε-SVR and least-squares SVR, were explored and compared to a widely applied linear regression model, the Bayesian Lasso, the latter assuming additive marker effects. Predictive performance was measured using predictive correlation and mean squared error of prediction. Depending on the kernel function chosen, SVR can model either linear or nonlinear relationships between phenotypes and marker genotypes. For milk yield, where phenotypes were estimated breeding values of bulls (a linear combination of the data), SVR with a Gaussian radial basis function (RBF) kernel had a slightly better performance than with a linear kernel, and was similar to the Bayesian Lasso. For the wheat data, where phenotype was raw grain yield, the RBF kernel provided clear advantages over the linear kernel, e.g., a 17.5% increase in correlation when using the ε-SVR. SVR with a RBF kernel also compared favorably to the Bayesian Lasso in this case. It is concluded that a nonlinear RBF kernel may be an optimal choice for SVR, especially when phenotypes to be predicted have a nonlinear dependency on genotypes, as it might have been the case in the wheat data.     

On the relation between residue flexibility and local solvent accessibility in proteins

We investigate the relationship between the flexibility, expressed with B‐factor, and the relative solvent accessibility (RSA) in the context of local, with respect to the sequence, neighborhood and related concepts such as residue depth. We observe that the flexibility of a given residue is strongly influenced by the solvent accessibility of the adjacent neighbors. The mean normalized B‐factor of the exposed residues with two buried neighbors is smaller than that of the buried residues with two exposed neighbors. Inclusion of RSA of the neighboring residues (local RSA) significantly increases correlation with the B‐factor. Correlation between the local RSA and B‐factor is shown to be stronger than the correlation that considers local distance‐ or volume‐based residue depth. We also found that the correlation coefficients between B‐factor and RSA for the 20 amino acids, called flexibility‐exposure correlation index, are strongly correlated with the stability scale that characterizes the average contributions of each amino acid to the folding stability. Our results reveal that the predicted RSA could be used to distinguish between the disordered and ordered residues and that the inclusion of local predicted RSA values helps providing a better contrast between these two types of residues. Prediction models developed based on local actual RSA and local predicted RSA show similar or better results in the context of B‐factor and disorder predictions when compared with several existing approaches. We validate our models using three case studies, which show that this work provides useful clues for deciphering the structure–flexibility–function relation. Proteins 2009. © 2009 Wiley‐Liss, Inc.     

Combining prediction of secondary structure and solvent accessibility in proteins

Owing to the use of evolutionary information and advanced machine learning protocols, secondary structures of amino acid residues in proteins can be predicted from the primary sequence with more than 75% per-residue accuracy for the 3-state (i.e., helix, beta-strand, and coil) classification problem. In this work we investigate whether further progress may be achieved by incorporating the relative solvent accessibility (RSA) of an amino acid residue as a fingerprint of the overall topology of the protein. Toward that goal, we developed a novel method for secondary structure prediction that uses predicted RSA in addition to attributes derived from evolutionary profiles. Our general approach follows the 2-stage protocol of Rost and Sander, with a number of Elman-type recurrent neural networks (NNs) combined into a consensus predictor. The RSA is predicted using our recently developed regression-based method that provides real-valued RSA, with the overall correlation coefficients between the actual and predicted RSA of about 0.66 in rigorous tests on independent control sets. Using the predicted RSA, we were able to improve the performance of our secondary structure prediction by up to 1.4% and achieved the overall per-residue accuracy between 77.0% and 78.4% for the 3-state classification problem on different control sets comprising, together, 603 proteins without homology to proteins included in the training. The effects of including solvent accessibility depend on the quality of RSA prediction. In the limit of perfect prediction (i.e., when using the actual RSA values derived from known protein structures), the accuracy of secondary structure prediction increases by up to 4%. We also observed that projecting real-valued RSA into 2 discrete classes with the commonly used threshold of 25% RSA decreases the classification accuracy for secondary structure prediction. While the level of improvement of secondary structure prediction may be different for prediction protocols that implicitly account for RSA in other ways, we conclude that an increase in the 3-state classification accuracy may be achieved when combining RSA with a state-of-the-art protocol utilizing evolutionary profiles. The new method is available through a Web server at http://sable.cchmc.org.     

A novel support vector regression model to estimate the phycocyanin concentration in turbid inland waters from hyperspectral reflectance

This study develops a novel support vector regression (SVR) model for retrieving the specific cyanobacterial pigment C-phycocyanin (C-PC) concentrations in cyanobacteria-dominated large turbid lakes of China. Lake Taihu, Lake Chaohu, and Lake Dianchi in China were our study areas. Five field cruises were carried out to collect data sets of optical and water quality parameters. To retrieve the C-PC, three types of reflectance forms, including single band, band ratio, and three-band-combination, were compared. The band ratio was the best candidate to serve for algorithm development. On this basis, two types of models, including linear models and a SVR model, were originally established. The previous typical algorithms were also examined. The obtained results showed that the best-performing model was the SVR model. By our validation data set, the proposed SVR model also presented accurate prediction results, with the lowest errors among all methods. The novelty of the SVR model compared to the previous ones lies in the inclusion of band ratios that are located outside of the main pigment absorption peaks but hold information on inflection points, curvature, etc., into empirical optimization. The implications of these findings indicates the potential applicability of the SVR models in lakes of the similar type.     

Systematic Artifacts in Support Vector Regression-Based Compound Potency Prediction Revealed by Statistical and Activity Landscape Analysis

Support vector machines are a popular machine learning method for many classification tasks in biology and chemistry. In addition, the support vector regression (SVR) variant is widely used for numerical property predictions. In chemoinformatics and pharmaceutical research, SVR has become the probably most popular approach for modeling of non-linear structure-activity relationships (SARs) and predicting compound potency values. Herein, we have systematically generated and analyzed SVR prediction models for a variety of compound data sets with different SAR characteristics. Although these SVR models were accurate on the basis of global prediction statistics and not prone to overfitting, they were found to consistently mispredict highly potent compounds. Hence, in regions of local SAR discontinuity, SVR prediction models displayed clear limitations. Compared to observed activity landscapes of compound data sets, landscapes generated on the basis of SVR potency predictions were partly flattened and activity cliff information was lost. Taken together, these findings have implications for practical SVR applications. In particular, prospective SVR-based potency predictions should be considered with caution because artificially low predictions are very likely for highly potent candidate compounds, the most important prediction targets.     

Screening methods for linear errors-in-variables models in high dimensions

Microarray studies, in order to identify genes associated with an outcome of interest, usually produce noisy measurements for a large number of gene expression features from a small number of subjects. One common approach to analyzing such high-dimensional data is to use linear errors-in-variables (EIV) models; however, current methods for fitting such models are computationally expensive. In this paper, we present two efficient screening procedures, namely, corrected penalized marginal screening (PMSc) and corrected sure independence screening (SISc), to reduce the number of variables for final model building. Both screening procedures are based on fitting corrected marginal regression models relating the outcome to each contaminated covariate separately, which can be computed efficiently even with a large number of features. Under mild conditions, we show that these procedures achieve screening consistency and reduce the number of features substantially, even when the number of covariates grows exponentially with sample size. In addition, if the true covariates are weakly correlated, we show that PMSc can achieve full variable selection consistency. Through a simulation study and an analysis of gene expression data for bone mineral density of Norwegian women, we demonstrate that the two new screening procedures make estimation of linear EIV models computationally scalable in high-dimensional settings, and improve finite sample estimation and selection performance compared with estimators that do not employ a screening stage.     

Hierarchical spatial modeling of additive and dominance genetic variance for large spatial trial datasets

Summary .  This article expands upon recent interest in Bayesian hierarchical models in quantitative genetics by developing spatial process models for inference on additive and dominance genetic variance within the context of large spatially referenced trial datasets. Direct application of such models to large spatial datasets are, however, computationally infeasible because of cubic-order matrix algorithms involved in estimation. The situation is even worse in Markov chain Monte Carlo (MCMC) contexts where such computations are performed for several iterations. Here, we discuss approaches that help obviate these hurdles without sacrificing the richness in modeling. For genetic effects, we demonstrate how an initial spectral decomposition of the relationship matrices negate the expensive matrix inversions required in previously proposed MCMC methods. For spatial effects, we outline two approaches for circumventing the prohibitively expensive matrix decompositions: the first leverages analytical results from Ornstein–Uhlenbeck processes that yield computationally efficient tridiagonal structures, whereas the second derives a modified predictive process model from the original model by projecting its realizations to a lower-dimensional subspace, thereby reducing the computational burden. We illustrate the proposed methods using a synthetic dataset with additive, dominance, genetic effects and anisotropic spatial residuals, and a large dataset from a Scots pine ( Pinus sylvestris L.) progeny study conducted in northern Sweden. Our approaches enable us to provide a comprehensive analysis of this large trial, which amply demonstrates that, in addition to violating basic assumptions of the linear model, ignoring spatial effects can result in downwardly biased measures of heritability.     

A comparison of five methods to predict genomic breeding values of dairy bulls from genome-wide SNP markers

Background

Genomic selection (GS) uses molecular breeding values (MBV) derived from dense markers across the entire genome for selection of young animals. The accuracy of MBV prediction is important for a successful application of GS. Recently, several methods have been proposed to estimate MBV. Initial simulation studies have shown that these methods can accurately predict MBV. In this study we compared the accuracies and possible bias of five different regression methods in an empirical application in dairy cattle.

Methods

Genotypes of 7,372 SNP and highly accurate EBV of 1,945 dairy bulls were used to predict MBV for protein percentage (PPT) and a profit index (Australian Selection Index, ASI). Marker effects were estimated by least squares regression (FR-LS), Bayesian regression (Bayes-R), random regression best linear unbiased prediction (RR-BLUP), partial least squares regression (PLSR) and nonparametric support vector regression (SVR) in a training set of 1,239 bulls. Accuracy and bias of MBV prediction were calculated from cross-validation of the training set and tested against a test team of 706 young bulls.

Results

For both traits, FR-LS using a subset of SNP was significantly less accurate than all other methods which used all SNP. Accuracies obtained by Bayes-R, RR-BLUP, PLSR and SVR were very similar for ASI (0.39-0.45) and for PPT (0.55-0.61). Overall, SVR gave the highest accuracy.All methods resulted in biased MBV predictions for ASI, for PPT only RR-BLUP and SVR predictions were unbiased. A significant decrease in accuracy of prediction of ASI was seen in young test cohorts of bulls compared to the accuracy derived from cross-validation of the training set. This reduction was not apparent for PPT. Combining MBV predictions with pedigree based predictions gave 1.05 - 1.34 times higher accuracies compared to predictions based on pedigree alone. Some methods have largely different computational requirements, with PLSR and RR-BLUP requiring the least computing time.

Conclusions

The four methods which use information from all SNP namely RR-BLUP, Bayes-R, PLSR and SVR generate similar accuracies of MBV prediction for genomic selection, and their use in the selection of immediate future generations in dairy cattle will be comparable. The use of FR-LS in genomic selection is not recommended.     

A probabilistic model for secondary structure prediction from protein chemical shifts

Protein chemical shifts encode detailed structural information that is difficult and computationally costly to describe at a fundamental level. Statistical and machine learning approaches have been used to infer correlations between chemical shifts and secondary structure from experimental chemical shifts. These methods range from simple statistics such as the chemical shift index to complex methods using neural networks. Notwithstanding their higher accuracy, more complex approaches tend to obscure the relationship between secondary structure and chemical shift and often involve many parameters that need to be trained. We present hidden Markov models (HMMs) with Gaussian emission probabilities to model the dependence between protein chemical shifts and secondary structure. The continuous emission probabilities are modeled as conditional probabilities for a given amino acid and secondary structure type. Using these distributions as outputs of first‐ and second‐order HMMs, we achieve a prediction accuracy of 82.3%, which is competitive with existing methods for predicting secondary structure from protein chemical shifts. Incorporation of sequence‐based secondary structure prediction into our HMM improves the prediction accuracy to 84.0%. Our findings suggest that an HMM with correlated Gaussian distributions conditioned on the secondary structure provides an adequate generative model of chemical shifts. Proteins 2013; © 2012 Wiley Periodicals, Inc.     

Genome‐enabled prediction of reproductive traits in Nellore cattle using parametric models and machine learning methods

This study aimed to assess the predictive ability of different machine learning (ML) methods for genomic prediction of reproductive traits in Nellore cattle. The studied traits were age at first calving (AFC), scrotal circumference (SC), early pregnancy (EP) and stayability (STAY). The numbers of genotyped animals and SNP markers available were 2342 and 321 419 (AFC), 4671 and 309 486 (SC), 2681 and 319 619 (STAY) and 3356 and 319 108 (EP). Predictive ability of support vector regression (SVR), Bayesian regularized artificial neural network (BRANN) and random forest (RF) were compared with results obtained using parametric models (genomic best linear unbiased predictor, GBLUP, and Bayesian least absolute shrinkage and selection operator, BLASSO). A 5‐fold cross‐validation strategy was performed and the average prediction accuracy (ACC) and mean squared errors (MSE) were computed. The ACC was defined as the linear correlation between predicted and observed breeding values for categorical traits (EP and STAY) and as the correlation between predicted and observed adjusted phenotypes divided by the square root of the estimated heritability for continuous traits (AFC and SC). The average ACC varied from low to moderate depending on the trait and model under consideration, ranging between 0.56 and 0.63 (AFC), 0.27 and 0.36 (SC), 0.57 and 0.67 (EP), and 0.52 and 0.62 (STAY). SVR provided slightly better accuracies than the parametric models for all traits, increasing the prediction accuracy for AFC to around 6.3 and 4.8% compared with GBLUP and BLASSO respectively. Likewise, there was an increase of 8.3% for SC, 4.5% for EP and 4.8% for STAY, comparing SVR with both GBLUP and BLASSO. In contrast, the RF and BRANN did not present competitive predictive ability compared with the parametric models. The results indicate that SVR is a suitable method for genome‐enabled prediction of reproductive traits in Nellore cattle. Further, the optimal kernel bandwidth parameter in the SVR model was trait‐dependent, thus, a fine‐tuning for this hyper‐parameter in the training phase is crucial.     

A guide for kernel generalized regression methods for genomic-enabled prediction

The primary objective of this paper is to provide a guide on implementing Bayesian generalized kernel regression methods for genomic prediction in the statistical software R. Such methods are quite efficient for capturing complex non-linear patterns that conventional linear regression models cannot. Furthermore, these methods are also powerful for leveraging environmental covariates, such as genotype × environment (G×E) prediction, among others. In this study we provide the building process of seven kernel methods: linear, polynomial, sigmoid, Gaussian, Exponential, Arc-cosine 1 and Arc-cosine L. Additionally, we highlight illustrative examples for implementing exact kernel methods for genomic prediction under a single-environment, a multi-environment and multi-trait framework, as well as for the implementation of sparse kernel methods under a multi-environment framework. These examples are followed by a discussion on the strengths and limitations of kernel methods and, subsequently by conclusions about the main contributions of this paper.Subject terms: Genomics, Plant sciences     

cnnAlpha: Protein disordered regions prediction by reduced amino acid alphabets and convolutional neural networks

Intrinsically disordered regions (IDR) play an important role in key biological processes and are closely related to human diseases. IDRs have great potential to serve as targets for drug discovery, most notably in disordered binding regions. Accurate prediction of IDRs is challenging because their genome wide occurrence and a low ratio of disordered residues make them difficult targets for traditional classification techniques. Existing computational methods mostly rely on sequence profiles to improve accuracy which is time consuming and computationally expensive. This article describes an ab initio sequence-only prediction method—which tries to overcome the challenge of accurate prediction posed by IDRs—based on reduced amino acid alphabets and convolutional neural networks (CNNs). We experiment with six different 3-letter reduced alphabets. We argue that the dimensional reduction in the input alphabet facilitates the detection of complex patterns within the sequence by the convolutional step. Experimental results show that our proposed IDR predictor performs at the same level or outperforms other state-of-the-art methods in the same class, achieving accuracy levels of 0.76 and AUC of 0.85 on the publicly available Critical Assessment of protein Structure Prediction dataset (CASP10). Therefore, our method is suitable for proteome-wide disorder prediction yielding similar or better accuracy than existing approaches at a faster speed.     

基于无人机的冬小麦拔节期表层土壤有机质含量遥感反演

快速监测大面积分布的盐渍化麦田土壤有机质含量,可为推进盐渍土改良和促进碳循环研究提供数据支撑。通过野外采样与获取无人机遥感影像,分别基于裸土和植被情况,采用多元线性回归(MLR)、偏最小二乘回归(PLSR)和支持向量机回归(SVR)3种方法,建立区域有机质含量遥感模型,并进行检验和对比,确定最优的土壤有机质含量反演模型;最后基于最优模型进行研究区表层土壤有机质的反演,并与插值结果进行比较。结果表明: 经5×5的中值滤波处理后的光谱与土壤表层有机质对应最优;3种模型中,SVR模型的预测精度最高,PLSR次之,MLR效果最差。对比两种变量的建模效果,基于植被的SVR建模效果最好,其建模决定系数(R²)、均方根误差(RMSE)分别为0.89、0.20,验证R²、RMSE分别为0.82、0.24;基于裸土的建模效果不理想,最优的也是SVR模型,其建模R²、RMSE分别为0.63、0.26,验证R²、RMSE分别为0.61、0.25。根据最优模型反演得到该区域有机质含量为17.51～22.53 g·kg^-1,平均值为19.51 g·kg^-1,与实地调查结果较为一致;插值结果与反演结果相比,精度受到限制。综上,基于无人机多光谱可以对盐渍土冬小麦拔节期土壤有机质含量进行快速、大范围精准估测。     

Automating human intuition for protein design

In the design of new enzymes and binding proteins, human intuition is often used to modify computationally designed amino acid sequences prior to experimental characterization. The manual sequence changes involve both reversions of amino acid mutations back to the identity present in the parent scaffold and the introduction of residues making additional interactions with the binding partner or backing up first shell interactions. Automation of this manual sequence refinement process would allow more systematic evaluation and considerably reduce the amount of human designer effort involved. Here we introduce a benchmark for evaluating the ability of automated methods to recapitulate the sequence changes made to computer‐generated models by human designers, and use it to assess alternative computational methods. We find the best performance for a greedy one‐position‐at‐a‐time optimization protocol that utilizes metrics (such as shape complementarity) and local refinement methods too computationally expensive for global Monte Carlo (MC) sequence optimization. This protocol should be broadly useful for improving the stability and function of designed binding proteins. Proteins 2014; 82:858–866. © 2013 Wiley Periodicals, Inc.     

Remote homology detection based on oligomer distances

MOTIVATION: Remote homology detection is among the most intensively researched problems in bioinformatics. Currently discriminative approaches, especially kernel-based methods, provide the most accurate results. However, kernel methods also show several drawbacks: in many cases prediction of new sequences is computationally expensive, often kernels lack an interpretable model for analysis of characteristic sequence features, and finally most approaches make use of so-called hyperparameters which complicate the application of methods across different datasets. RESULTS: We introduce a feature vector representation for protein sequences based on distances between short oligomers. The corresponding feature space arises from distance histograms for any possible pair of K-mers. Our distance-based approach shows important advantages in terms of computational speed while on common test data the prediction performance is highly competitive with state-of-the-art methods for protein remote homology detection. Furthermore the learnt model can easily be analyzed in terms of discriminative features and in contrast to other methods our representation does not require any tuning of kernel hyperparameters. AVAILABILITY: Normalized kernel matrices for the experimental setup can be downloaded at www.gobics.de/thomas. Matlab code for computing the kernel matrices is available upon request. CONTACT: thomas@gobics.de, peter@gobics.de.     

PaleAle 5.0: prediction of protein relative solvent accessibility by deep learning

Predicting the three-dimensional structure of proteins is a long-standing challenge of computational biology, as the structure (or lack of a rigid structure) is well known to determine a protein’s function. Predicting relative solvent accessibility (RSA) of amino acids within a protein is a significant step towards resolving the protein structure prediction challenge especially in cases in which structural information about a protein is not available by homology transfer. Today, arguably the core of the most powerful prediction methods for predicting RSA and other structural features of proteins is some form of deep learning, and all the state-of-the-art protein structure prediction tools rely on some machine learning algorithm. In this article we present a deep neural network architecture composed of stacks of bidirectional recurrent neural networks and convolutional layers which is capable of mining information from long-range interactions within a protein sequence and apply it to the prediction of protein RSA using a novel encoding method that we shall call “clipped”. The final system we present, PaleAle 5.0, which is available as a public server, predicts RSA into two, three and four classes at an accuracy exceeding 80% in two classes, surpassing the performances of all the other predictors we have benchmarked.