OPUS-TASS: a protein backbone torsion angles and secondary structure predictor based on ensemble neural networks

Motivation Predictions of protein backbone torsion angles (ϕ and ψ) and secondary structure from sequence are crucial subproblems in protein structure prediction. With the development of deep learning approaches, their accuracies have been significantly improved. To capture the long-range interactions, most studies integrate bidirectional recurrent neural networks into their models. In this study, we introduce and modify a recently proposed architecture named Transformer to capture the interactions between the two residues theoretically with arbitrary distance. Moreover, we take advantage of multitask learning to improve the generalization of neural network by introducing related tasks into the training process. Similar to many previous studies, OPUS-TASS uses an ensemble of models and achieves better results. Results OPUS-TASS uses the same training and validation sets as SPOT-1D. We compare the performance of OPUS-TASS and SPOT-1D on TEST2016 (1213 proteins) and TEST2018 (250 proteins) proposed in the SPOT-1D paper, CASP12 (55 proteins), CASP13 (32 proteins) and CASP-FM (56 proteins) proposed in the SAINT paper, and a recently released PDB structure collection from CAMEO (93 proteins) named as CAMEO93. On these six test sets, OPUS-TASS achieves consistent improvements in both backbone torsion angles prediction and secondary structure prediction. On CAMEO93, SPOT-1D achieves the mean absolute errors of 16.89 and 23.02 for ϕ and ψ predictions, respectively, and the accuracies for 3- and 8-state secondary structure predictions are 87.72 and 77.15%, respectively. In comparison, OPUS-TASS achieves 16.56 and 22.56 for ϕ and ψ predictions, and 89.06 and 78.87% for 3- and 8-state secondary structure predictions, respectively. In particular, after using our torsion angles refinement method OPUS-Refine as the post-processing procedure for OPUS-TASS, the mean absolute errors for final ϕ and ψ predictions are further decreased to 16.28 and 21.98, respectively. Availability and implementation The training and the inference codes of OPUS-TASS and its data are available at https://github.com/thuxugang/opus_tass. Supplementary information Supplementary data are available at Bioinformatics online.

In silico identification of conserved cis-acting RNA elements in the SARS-CoV-2 genome

Aim: The aim of this study was to computationally predict conserved RNA sequences and structures known as cis-acting RNA elements (CREs) in the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) genome. Materials & methods: Bioinformatics tools were used to analyze and predict CREs by obtaining viral sequences from available databases. Results: Computational analysis revealed the presence of RNA stem-loop structures within the 3′ end of the ORF1ab region analogous to previously identified SARS-CoV genomic packaging signals. Alignment-based RNA secondary structure predictions of the 5′ end of the SARS-CoV-2 genome also identified conserved CREs. Conclusion: These CREs may be potential vaccine and/or antiviral therapeutic targets; however, further studies are warranted to confirm their roles in the SARS-CoV-2 life cycle.

Novel Polybasic Cleavage Site in SARS-CoV-2 Genome Is Likely to Induce a Major Change in the RNA Secondary Structure

Severe Acute Respiratory Syndrome CoronaVirus 2 (SARS-CoV-2) has claimed nearly 180,000 lives and continues to spread. There are currently no approved medications or vaccines for this new coronavirus. Studies have shown that the positive RNA genome of SARS-CoV-2 contains unique features, including a 12-base sequence inserted between the two subunits of viral receptor protein Spike. This inserted sequence facilitates the cleavage of Spike by the cellular proteases Furin and TMPRSS2, leading to the fusion of virus and host cell membranes. Current studies are mostly focused on the SARS-CoV-2 Spike protein and its interacting cellular proteins ACE2, Furin, and TMPRSS2. RNA structural studies are limited and little is known about the potential impact of the 12-base sequence insert on the secondary structure of SARS-CoV-2 genomic RNA and/or its transcripts. Here, by using local and global RNA secondary structure predictions, we show that the novel 12-base insert of SARS-CoV-2 genome likely induces a major RNA secondary structure change.

A method for predicting evolved fold switchers exclusively from their sequences

Although most proteins with known structures conform to the longstanding rule-of-thumb that high levels of aligned sequence identity tend to indicate similar folds and functions, an increasing number of exceptions is emerging. In spite of having highly similar sequences, these “evolved fold switchers” (1) can adopt radically different folds with disparate biological functions. Predictive methods for identifying evolved fold switchers are desirable because some of them are associated with disease and/or can perform different functions in cells. Previously, we showed that inconsistencies between predicted and experimentally determined secondary structures can be used to predict fold switching proteins (2). The usefulness of this approach is limited, however, because it requires experimentally determined protein structures, whose magnitude is dwarfed by the number of genomic proteins. Here, we use secondary structure predictions to identify evolved fold switchers from their amino acid sequences alone. To do this, we looked for inconsistencies between the secondary structure predictions of the alternative conformations of evolved fold switchers. We used three different predictors in this study: JPred4, PSIPRED, and SPIDER3. We find that overall inconsistencies are not a significant predictor of evolved fold switchers for any of the three predictors. Inconsistencies between α-helix and β-strand predictions made by JPred4, however, can discriminate between the different conformations of evolved fold switchers with statistical significance (p < 1.7*10−13). In light of this observation, we used these inconsistencies as a classifier and found that it could robustly discriminate between evolved fold switchers and evolved non-fold-switchers, as evidenced by a Matthews correlation coefficient of 0.90. These results indicate that inconsistencies between secondary structure predictions can indeed be used to identify evolved fold switchers from their genomic sequences alone. Our findings have implications for genomics, structural biology, and human health.

2dSS: a web server for protein secondary structure visualization

2dSS is a web-server for visualising and comparing secondary structure predictions. It provides two main functionalities: 2D-alignment and compare predictions. The “2D-alignment” has been designed to visualise conserved secondary structure elements in a multiple sequence alignment (MSA). From this we can study the secondary structure content of homologous proteins (a protein family) and highlight its structural patterns. The “compare predictions” has been designed to compare the output of several secondary structure prediction tools, and check their accuracy when compared with real secondary structure elements extracted from 3D-structure. 2dSS provides a comprehensive representation of protein secondary structure elements, and it can be used to visualise and compare secondary structures of any prediction tool.Availabilityhttp://genome.lcqb.upmc.fr/2dss/

A sequence-based approach for identifying protein fold switchers

Although most proteins conform to the classical one-structure/one-function paradigm, an increasing number of proteins with dual structures and functions are emerging. These fold-switching proteins remodel their secondary structures in response to cellular stimuli, fostering multi-functionality and tight cellular control. Accurate predictions of fold-switching proteins could both suggest underlying mechanisms for uncharacterized biological processes and reveal potential drug targets. Previously, we developed a prediction method for fold-switching proteins based on secondary structure predictions and structure-based thermodynamic calculations. Given the large number of genomic sequences without homologous experimentally characterized structures, however, we sought to predict fold-switching proteins from their sequences alone. To do this, we leveraged state-of-the-art secondary structure predictions, which require only amino acid sequences but are not currently designed to identify structural duality in proteins. Thus, we hypothesized that incorrect and inconsistent secondary structure predictions could be good initial predictors of fold-switching proteins. We found that secondary structure predictions of fold-switching proteins with solved structures are indeed less accurate than secondary structure predictions of non-fold-switching proteins with solved structures. These inaccuracies result largely from the conformations of fold-switching proteins that are underrepresented in the Protein Data Bank (PDB), and, consequently, the training sets of secondary structure predictors. Given that secondary structure predictions are homology-based, we hypothesized that decontextualizing the inaccurately-predicted regions of fold-switching proteins could weaken the homology relationships between these regions and their overpopulated structural representatives. Thus, we reran secondary structure predictions on these regions in isolation and found that they were significantly more inconsistent than in regions of non-fold-switching proteins. Thus, inconsistent secondary structure predictions can serve as a preliminary marker of fold switching. These findings have implications for genomics and the future development of secondary structure predictors.

Extant fold-switching proteins are widespread

A central tenet of biology is that globular proteins have a unique 3D structure under physiological conditions. Recent work has challenged this notion by demonstrating that some proteins switch folds, a process that involves remodeling of secondary structure in response to a few mutations (evolved fold switchers) or cellular stimuli (extant fold switchers). To date, extant fold switchers have been viewed as rare byproducts of evolution, but their frequency has been neither quantified nor estimated. By systematically and exhaustively searching the Protein Data Bank (PDB), we found ∼100 extant fold-switching proteins. Furthermore, we gathered multiple lines of evidence suggesting that these proteins are widespread in nature. Based on these lines of evidence, we hypothesized that the frequency of extant fold-switching proteins may be underrepresented by the structures in the PDB. Thus, we sought to identify other putative extant fold switchers with only one solved conformation. To do this, we identified two characteristic features of our ∼100 extant fold-switching proteins, incorrect secondary structure predictions and likely independent folding cooperativity, and searched the PDB for other proteins with similar features. Reassuringly, this method identified dozens of other proteins in the literature with indication of a structural change but only one solved conformation in the PDB. Thus, we used it to estimate that 0.5–4% of PDB proteins switch folds. These results demonstrate that extant fold-switching proteins are likely more common than the PDB reflects, which has implications for cell biology, genomics, and human health.