Mitigating Data Scarcity in Protein Binding Prediction Using Meta-Learning

Abstract

A plethora of biological functions are performed through various types of protein-peptide binding. Prime examples include the protein kinase phosphorylation on peptide substrates and the binding of major histocompatibility complex to neoantigens in the immune system. Understanding the specificity of protein-peptide interactions is critical for unraveling the architectures of functional pathways and the mechanisms of cellular processes in human cells. Despite massspectrometric techniques were developed for the identification of protein-peptide interactions, our understanding of the preferences of proteins on their binding peptides is still rudimentary. As a complementary direction, a line of computational prediction methods has been recently proposed to predict protein-peptide bindings which efficiently provide rich functional annotations on a large scale. To achieve a high prediction accuracy, these computational methods require a sufficient amount of data to build the prediction model. However, the number of experimentally verified protein-peptide bindings is often limited in real cases. For example, a majority of protein kinases have very few experimentally verified phosphorylation sites (e.g., less than 30 sites) in existing databases. These methods are thus limited to building accurate prediction models for only well-characterized proteins with a large volume of known binding peptides and cannot be extended to predict new binding peptides for less-studied proteins. In this paper, we introduce a generic framework to address this issue of data scarcity in protein binding prediction. We demonstrate the applicability of our framework in predicting kinase-specific phosphorylation sites. Our method uses an effective training strategy to build a prediction model with robust transferability. The model is able to predict the phosphorylation sites of a less-studied kinase, even if there is only a small number of phosphorylation sites known for this kinase. To achieve this, we train the model via a meta-learning phase followed by a few-shot learning phase. We demonstrate our framework has better transferability than state-of-the-art methods and is effective in utilizing limited data to accurately predict phosphorylation sites for less-characterized kinases. The implementation of our framework is available at https://github.com/luoyunan/MetaKinase.

Integrating thermodynamic and sequence contexts improves protein-RNA binding prediction

“…Deep learning continues to proliferate as a powerful set of tools to solve an increasingly diverse range of problems, including many in structural and systems biology [23–29]. We use the convolutional neural network (CNN) as the base model f θ to predict protein-RNA binding, where θ represents the network weights.…”

Global Convergence of MAML and Theory-Inspired Neural Architecture Search for Few-Shot Learning

“…Meta-learning, or learning-to-learn (LTL) [59], has received much attention due to its applicability in few-shot image classification [23,69], meta reinforcement learning [18,23,62], and other domains such as natural language processing [8,78] and computational biology [43]. The primary motivation for meta-learning is to fast learn a new task from a small amount of data, with prior experience on similar but different tasks.…”

Bridging Multi-Task Learning and Meta-Learning: Towards Efficient Training and Effective Adaptation

“…On the other hand, in addition to the joint training scheme, modern meta-learning can leverage the shared representation to fast adapt to unseen tasks with only minimum limited data during the test phase (Hospedales et al, 2020). As a result, meta-learning has drawn increasing attention and been applied to a wide range of learning tasks, including few-shot learning (Snell et al, 2017;Vinyals et al, 2016;Lee et al, 2019b), meta reinforcement learning (Finn et al, 2017), speech recognition (Hsu et al, 2020) and bioinformatics (Luo et al, 2019).…”