Sequence-based prediction of protein-protein interactions: a structure-aware interpretable deep learning model

Samuel Sledzieski, Rohit Singh, Lenore Cowen, Bonnie BergerbioRxiv 2021 · DOI

TL;DR

This paper introduces D-SCRIPT, a deep learning model that predicts protein-protein interactions (PPIs) using only protein sequences. D-SCRIPT excels at generalizing to new species and is more robust to limited training data by incorporating a "contact map" prediction step, offering a more interpretable and structurally-informed approach to PPI prediction.

What Problem Does This Paper Address?

Understanding how proteins interact (PPIs) is crucial for deciphering biological processes. However, experimentally determining these interactions is costly and time-consuming, leading to sparse data, especially in non-model organisms. Existing computational methods often struggle to generalize to new species or require specific types of training data. This paper aims to develop a method that can accurately predict PPIs from sequences alone, with improved cross-species generalization and robustness to data scarcity.

Methods

D-SCRIPT employs a two-stage deep learning architecture:

  1. Sequence-based Protein Embedding: It leverages a pre-trained language model (from Bepler & Berger) that encodes structural information within protein sequences. This embedding is then projected into a lower-dimensional space.
  2. Structure-Aware Interaction Prediction: This is the novel part. Instead of directly predicting interaction probability, D-SCRIPT first predicts a contact map between residues of the two proteins. This map estimates which amino acids are likely to be in close proximity if the proteins interact. This intermediate step is inspired by the intuition that physical interaction requires direct contact between specific residues. The contact map is then used to predict the final interaction probability.

The model is trained end-to-end, optimizing for both accurate interaction prediction and a sparse, realistic contact map.

Datasets

  • STRING database (version 11): Used for training and evaluating PPIs in various model organisms. High-confidence binding interactions with experimental evidence were selected.
  • BioGRID: Used for estimating the recency of PPI data in Figure 1.
  • Protein Data Bank (PDB): Used to evaluate the model's ability to predict self-contacts within individual proteins.
  • Hwang et al.'s benchmark data set of docked protein structures: Used to assess the physical accuracy of the predicted inter-protein contact maps.
  • SARS-CoV-2 and human protein sequences: Used for a preliminary study on predicting viral-host interactions.

Key Insights and Ideas

  • Contact Map as an Intermediate Step: Predicting an inter-protein contact map before the final interaction score is a key innovation. This forces the model to learn about the physical basis of interaction, not just correlations.
  • Structure-Aware Embeddings: Leveraging pre-trained models that already encode structural information in protein sequence embeddings is crucial for D-SCRIPT's performance.
  • Cross-Species Generalization: D-SCRIPT's structural focus allows it to generalize much better to new species than methods that rely heavily on within-species network structure or frequency.
  • Interpretability: The contact map output provides a degree of interpretability, showing where on the proteins the interaction might occur.
  • Robustness to Data Sparsity: D-SCRIPT performs better than other methods when training data is limited, especially for proteins that are less frequently studied.

Key Results

  • Improved Cross-Species Prediction: D-SCRIPT significantly outperforms the state-of-the-art method PIPR in predicting PPIs across different species, achieving higher AUPR scores. For example, in D. melanogaster, D-SCRIPT achieved an AUPR of 0.552 compared to PIPR's 0.231.
  • Enhanced Functional Characterization: A D-SCRIPT-predicted PPI network for Drosophila led to more functionally coherent clusters of genes compared to a PIPR-predicted network.
  • Physically Plausible Contact Maps: The predicted contact maps showed significant overlap with ground-truth contacts from known protein complexes, especially when D-SCRIPT predicted an interaction.
  • Better Performance on Infrequent Proteins: D-SCRIPT shows a relative advantage over PIPR when predicting interactions involving proteins that appear less frequently in the training data.
  • Computational Efficiency: D-SCRIPT trains in approximately 3 days for 10 epochs on a single 32GB GPU and prediction is very fast (0.13 seconds/pair).

Limitations and Future Work

  • Within-Species Performance: While excelling cross-species, D-SCRIPT's within-species performance on human PPIs is slightly lower than PIPR, which may be due to its regularization choices that prioritize generalizability over specific network structure.
  • Hybrid Models: Hybrid approaches combining D-SCRIPT with other methods (like PIPR) often yield the best overall performance, suggesting complementary strengths.
  • Transfer Learning: The authors suggest future work could involve fine-tuning a pre-trained D-SCRIPT model for specific target species.
  • Integration with Network-Based Methods: Combining D-SCRIPT with graph-theoretic PPI prediction methods could be a fruitful direction.
  • Incorporating Structural Data: Future models could potentially integrate explicit 3D structural information more directly.

Summary updated

Abstract

Abstract Protein-protein interaction (PPI) networks have proven to be a valuable tool in systems biology to facilitate the discovery and understanding of protein function. Unfortunately, experimental PPI data remains sparse in most model organisms and even more so in other species. Existing methods for computational prediction of PPIs seek to address this limitation, and while they perform well when sufficient within-species training data is available, they generalize poorly to new species or often require specific types and sizes of training data that may not be available in the species of interest. We therefore present D-SCRIPT, a deep learning method for predicting a physical interaction between two proteins given just their sequences. Compared to existing methods, D-SCRIPT generalizes better to new species and is robust to limitations in training data size. Our approach encodes the intuition that for two proteins to physically interact, a subset of amino acids from each protein should be in contact with the other. The intermediate stages of D-SCRIPT directly implement this intuition; the penultimate stage in D-SCRIPT is a rough estimate of the inter-protein contact map of the protein dimer. This structurally-motivated design enables interpretability of our model and, since structure is more conserved evolutionarily than sequence, improves generalizability across species. We show that a D-SCRIPT model trained on 38,345 human PPIs enables significantly improved functional characterization of fly proteins compared to the state-of-the-art approach. Evaluating the same D-SCRIPT model on protein complexes with known 3-D structure, we find that the inter-protein contact map output by D-SCRIPT has significant overlap with the ground truth. Our work suggests that recent advances in deep learning language modeling of protein structure can be leveraged for protein interaction prediction from sequence. D-SCRIPT is available at http://dscript.csail.mit.edu .
BibTeX
@misc{sledzieski2021sequence,
  author = {Samuel Sledzieski and Rohit Singh and Lenore Cowen and Bonnie Berger},
  title = {{Sequence-based prediction of protein-protein interactions: a structure-aware interpretable deep learning model}},
  year = {2021},
  doi = {10.1101/2021.01.22.427866},
  howpublished = {bioRxiv},
}
Figure