De novo protein structure prediction by incremental inter-residue geometries prediction and model quality assessment using deep learning

Motivation: The successful application of deep learning has promoted progress in protein model quality assessment. How to use model quality assessment to further improve the accuracy of protein structure prediction, especially not reliant on the existing templates, is helpful for unraveling the folding mechanism. Here, we investigate whether model quality assessment can be introduced into structure prediction to form a closed-loop feedback, and iteratively improve the accuracy of de novo protein structure prediction. Results: In this study, we propose a de novo protein structure prediction method called RocketX. In RocketX, a feedback mechanism is constructed through the geometric constraint prediction network GeomNet, the structural simulation module, and the model quality evaluation network EmaNet. In GeomNet, the co-evolutionary features extracted from MSA that search from the sequence databases are sent to an improved residual neural network to predict the inter-residue geometric constraints. The structure model is folded based on the predicted geometric constraints. In EmaNet, the 1D and 2D features are extracted from the folded model and sent to the deep residual neural network to estimate the inter-residue distance deviation and per-residue lDDT of the model, which will be fed back to GeomNet as dynamic features to correct the geometries prediction and progressively improve model accuracy. RocketX is tested on 483 benchmark proteins and 20 FM targets of CASP14. Experimental results show that the closed-loop feedback mechanism significantly contributes to the performance of RocketX, and the prediction accuracy of RocketX outperforms that of the state-of-the-art methods trRosetta (without templates) and RaptorX. In addition, the blind test results on CAMEO show that although no template is used, the prediction accuracy of RocketX on medium and hard targets is comparable to the advanced methods that integrate templates.

Construct a variable-length fragment library for de novo protein structure prediction

Although remarkable achievements, such as AlphaFold2, have been made in end-to-end structure prediction, fragment libraries remain essential for de novo protein structure prediction, which can help explore and understand the protein-folding mechanism. In this work, we developed a variable-length fragment library (VFlib). In VFlib, a master structure database was first constructed from the Protein Data Bank through sequence clustering. The Hidden Markov Model (HMM) profile of each protein in the master structure database was generated by HHsuite, and the secondary structure of each protein was calculated by DSSP. For the query sequence, the HMM-profile was first constructed. Then, variable-length fragments were retrieved from the master structure database through dynamically variable-length profile-profile comparison. A complete method for chopping the query HMM-profile during this process was proposed to obtain fragments with increased diversity. Finally, secondary structure information was used to further screen the retrieved fragments to generate the final fragment library of specific query sequence. The experimental results obtained with a set of 120 nonredundant proteins showed that the global precision and coverage of the fragment library generated by VFlib were 55.04% and 94.95% at the RMSD cutoff of 1.5 Å, respectively. Compared to the benchmark method of NNMake, the global precision of our fragment library had increased by 62.89% with equivalent coverage. Furthermore, the fragments generated by VFlib and NNMake were used to predict structure models through fragment assembly. Controlled experimental results demonstrated that the average TM-score of VFlib was 16.00% higher than that of NNMake.