[Paper Review] Interpretable Structured Learning with Sparse Gated Sequence Encoder for Protein-Protein Interaction Prediction

Kishan KC, Feng Cui+2 more|arXiv (Cornell University)|Oct 16, 2020

Machine Learning in Bioinformatics26 references4 citations

TL;DR

This paper proposes a sparse gated bidirectional GRU model for interpretable protein-protein interaction (PPI) prediction from amino acid sequences. By integrating structured sparsity and a Gaussian latent space, the model efficiently encodes sequences, selectively highlighting biologically relevant motifs while outperforming state-of-the-art methods on yeast and human PPI datasets with improved interpretability.

ABSTRACT

Predicting protein-protein interactions (PPIs) by learning informative representations from amino acid sequences is a challenging yet important problem in biology. Although various deep learning models in Siamese architecture have been proposed to model PPIs from sequences, these methods are computationally expensive for a large number of PPIs due to the pairwise encoding process. Furthermore, these methods are difficult to interpret because of non-intuitive mappings from protein sequences to their sequence representation. To address these challenges, we present a novel deep framework to model and predict PPIs from sequence alone. Our model incorporates a bidirectional gated recurrent unit to learn sequence representations by leveraging contextualized and sequential information from sequences. We further employ a sparse regularization to model long-range dependencies between amino acids and to select important amino acids (protein motifs), thus enhancing interpretability. Besides, the novel design of the encoding process makes our model computationally efficient and scalable to an increasing number of interactions. Experimental results on up-to-date interaction datasets demonstrate that our model achieves superior performance compared to other state-of-the-art methods. Literature-based case studies illustrate the ability of our model to provide biological insights to interpret the predictions.

Motivation & Objective

• To address the computational inefficiency of Siamese architecture in pairwise PPI modeling.
• To enhance model interpretability by identifying key amino acid residues involved in PPIs.
• To improve performance on large-scale PPI prediction using sequence-only data.
• To enable biological insight by aligning learned attention with known protein motifs.
• To develop a scalable, efficient framework that avoids redundant pairwise encoding.

Proposed method

• Employ a bidirectional gated recurrent unit (BiGRU) to learn contextual and sequential representations from variable-length protein sequences.
• Model sequence representations as latent Gaussian distributions to capture uncertainty and improve generalization.
• Introduce a structured sparse gating mechanism that assigns sparse, learnable weights to amino acid positions to highlight important residues.
• Use a contrastive loss function based on statistical distance between latent distributions to distinguish interacting from non-interacting protein pairs.
• Apply fusedmax and sparsemax gating strategies to control sparsity and improve alignment with biological motifs.
• Encode each unique protein only once per batch, reducing computational cost compared to pairwise encoding in Siamese models.

Experimental results

Research questions

• RQ1 Can a sequence-encoding model achieve superior PPI prediction performance while being computationally efficient at scale?
• RQ2 Can a sparse gating mechanism effectively identify biologically relevant protein motifs from sequences?
• RQ3 Does the model's attention mechanism align with known functional motifs from databases like Pfam?
• RQ4 Can the model maintain high performance while reducing parameter redundancy through structured sparsity?
• RQ5 How does the model's interpretability compare to existing state-of-the-art PPI predictors?

Key findings

• The model achieves state-of-the-art performance on both yeast and human PPI datasets from the BioGRID database.
• With fusedmax gating, 23.33% of selected amino acids on average align with Pfam motifs in the human dataset, indicating strong biological relevance.
• For the yeast dataset, 59.05% of amino acids selected by the fusedmax gate mechanism align with known motifs, demonstrating high interpretability.
• The model reduces training time and memory usage by encoding unique proteins only once per batch, unlike pairwise Siamese models.
• Visualization confirms that the model's activated regions correspond to known functional domains such as the LSM domain (PF01423) and AP4 complex (PF14807).
• Case studies show the model successfully identifies critical residues in proteins like LSM8, SMD2, and RPC11, matching known motif locations.