The liquid-liquid phase separation of intrinsically disordered proteins (IDPs) is a potential response mechanism triggered by cellular environment variations. However, molecular-level understanding of IDPs is limited due to experimental characterization difficulties. Computational understanding of IDPs also requires fundamental advances, as the leading tools for predicting protein folding (e.g., Alphafold), typically fail to describe the structure of IDPs due to their data-driven approaches focusing on folded proteins. The focus of this work is to 1) develop new representations for intrinsically disordered proteins, 2) pair these representations with classical machine learning and deep learning models to predict the radius of gyration (R_g) of IDPs, and 3) develop protocols to build robust models compatible with the fundamental physics of IDPs and design new IDP sequences.
The current state-of-the-art featurization techniques for peptide-based sequences are one-hot encoding schemes adapted from natural language processing. These schemes have problems with the high-dimensionality of features for longer sequences. Moreover, the longest sequence of interest must be known a priori. Here, we build a new physically-motivated feature (motivated by the importance of neighbor-neighbor interactions) called the bag-of-amino-acid-interactions representation. This feature essentially counts and weights all possible non-bonded interactions in a sequence, and thus is, in principle, compatible with arbitrary sequence lengths. Here, we assess the performance of this new feature, and compare its performance in predicting disordered polymer R_g to existing categorical and physically-motivated featurization techniques. Our test computational dataset consists of 10,000 sequences simulated at the coarse-grained level, with the polypeptides ranging from 30 to 200 amino acids. Current results indicate that this new feature outperforms the others and possess solid extrapolation capabilities. To improve the model and understand the physical insight of IDPs, new sequence design using generative model could give a more diverse and real dataset. For future use, this feature can not only potentially provide physical insight of amino acid interactions but also be applied to the other protein spaces.

Tzu Hsuan Chao

Appearances