TooT-PLM-P2S: Incorporating secondary structure information into protein language models

Abstract

In bioinformatics, modeling the protein space to better predict function and structure has benefitted from Protein Language Models (PLMs). Their basis is the proteins amino acid sequence and self-supervised learning. Ankh is a prime example of such a PLM. While there has been some recent work on integrating structure with a PLM to enhance predictive performance, to date there has been no work on integrating secondary structure rather than three-dimensional structure. Here we present TooT-PLM-P2S that begins with the Ankh model pre-trained on 45 million proteins using self-supervised learning. TooT-PLM-P2S builds upon the Ankh model by initially using its pre-trained encoder and decoder. It then undergoes an additional training phase with approximately 10,000 proteins and their corresponding secondary structures. This retraining process modifies the encoder and decoder, resulting in the creation of TooT-PLM-P2S. We then assess the impact of integrating secondary structure information into the Ankh model by comparing Ankh and TooT-PLM-P2S on eight downstream tasks including fluorescence and solubility prediction, sub-cellular localization, and membrane protein classification. For both Ankh and TooT-PLM-P2S the downstream tasks required task-specific training. Few of the results showed statistically significant differences. Ankh outperformed on three of the eight tasks, TooT-PLM-P2S did not outperform on any task for the primary metric. TooT-PLM-P2S did outperform for the precision metric for the task of discriminating membrane proteins from non-membrane proteins. This study requires future work with expanded datasets and refined integration methods.