Protein secondary structure and backbone torsion angle prediction using multitask learning

Abstract

Protein structures provide valuable insights into their roles and functions inside living organisms. However, experimental approaches to determining protein structures are time-consuming and expensive, resulting in the development of computational methods to predict them. In the post- AlphaFold2 era, single-sequence-based protein structure prediction is a new challenge, allowing reliable estimation of protein 3D structures solely based on their primary sequences and without depending on multiple- sequence-alignments (MSA) of their sequence homologs. Accurate single- sequence-based prediction of protein structural properties, such as 8- state (Q8) secondary structure as well as backbone torsion ϕ and ψ angles, will pave the way for highly precise sequence-based prediction of protein structures. We present two multitask learning-based methods: (i) evolutionary-feature-based SAINT-Evolve and (ii) single-sequence-based SAINT-Single to accurately predict protein Q8 secondary structure (SS) and backbone torsion angles (ϕ and ψ). We developed them based on the previously proposed single-task learning-based prediction methods SAINT and SAINT-Angle for Q8-SS and backbone torsion angle, respectively. We attempted to leverage simultaneous learning of Q8-SS and backbone torsion angle prediction to boost predictive performance. Besides, we took advantage of extracted sequence embeddings from state-of-the-art protein language models to obtain better prediction results, particularly for single- sequence-based models. We compared the predictions from our methods extensively with respect to other competing protein structural property predictors on a wide range of benchmark datasets. The experimental results indicate that our proposed methods produce reliable predictions for proteins regardless of whether they have few sequence homologs or abundant homologous sequences.