PRIEST - Predicting viral mutations with immune escape capability of SARS-CoV-2 using temporal evolutionary information

Gourab Saha, Shashata Sawmya, Md. Ajwad Akil, Arpita Saha, Sadia Tasnim, Md. Saifur RahmanbioRxiv 2023 · DOI

TL;DR

This paper introduces PRIEST, a novel deep-learning framework that predicts SARS-CoV-2 mutations and their immune escape capabilities by analyzing temporal evolutionary information from viral sequences. PRIEST demonstrates high accuracy in predicting future variants, offering a crucial tool for proactive pandemic response and vaccine development.

What Problem Does This Paper Address?

The rapid evolution of viruses like SARS-CoV-2, driven by mutations, leads to the emergence of variants that can evade host immune responses and vaccines. Accurately predicting these mutations and their potential for immune escape is critical for effective pandemic mitigation, vaccine design, and therapeutic development. Existing methods often struggle to capture the dynamic, temporal nature of viral evolution and its impact on immune evasion.

Methods

The core of this work is the PRIEST (Predicting viral mutations with immune escape capability) framework. PRIEST utilizes a deep-learning architecture that combines:

  • 1D Convolutional Neural Networks (CNNs): These are used for initial feature extraction from time-series protein sequences. The architecture employs multiple "channels" with varying kernel sizes and operations to capture diverse patterns.
  • Transformer Attention Mechanism: This component is crucial for capturing long-range temporal dependencies within the viral sequences. It uses an attention mask to prevent the model from "looking into the future," ensuring predictions are based on past and present data.
  • ProtVec Embeddings: Protein sequences are transformed into fixed-size vector representations (embeddings) using the ProtVec module, which captures contextual information about amino acid triplets (trigrams).
  • Time Series Construction: Viral sequences are organized into time series based on quarters, allowing the model to learn from the evolutionary trajectory of the virus.
  • Labeling: Data points are labeled as "mutation" (1) or "no mutation" (0) based on residue changes between consecutive time steps.

Datasets

The study utilizes SARS-CoV-2 protein sequences obtained from the GISAID database. Sequences were collected up to the first quarter of 2022 and categorized into three classes: benign variants, variants being monitored, and variants of concern. The data was divided into time steps based on quarters, creating a time-series dataset for training and evaluation.

Key Insights and Ideas

  • Temporal Information is Key: The paper strongly emphasizes that analyzing the evolutionary history (time-series) of viral sequences is crucial for predicting future mutations and their properties, rather than just looking at static sequences.
  • Transformer's Advantage: The transformer's self-attention mechanism proves highly effective in capturing long-term dependencies in viral evolution, outperforming traditional recurrent models like LSTMs for this task.
  • Multi-Channel CNNs for Feature Richness: Employing multiple CNN channels with different kernel sizes allows PRIEST to extract a richer set of features from the sequence data, contributing to its predictive power.
  • Proactive Prediction: PRIEST's ability to predict real-world variant emergence before they become widespread is a significant advancement, enabling earlier intervention and vaccine adaptation.
  • Interpretable Attention: While not explicitly detailed in the main text, the authors mention extracting attention scores, suggesting potential for understanding which parts of the sequence are most influential in predictions.

Key Results

  • High Predictive Accuracy: PRIEST achieved higher Matthews Correlation Coefficient (MCC) scores than several baseline models (RNN, LSTM, GRU, Tempel, Tempo, CNN) for predicting mutations at time steps k=3 and k=6. For k=9, it performed comparably.
  • Superior AUROC and AUPRC: PRIEST consistently outperformed other models in terms of Area Under the Receiver Operating Characteristic (AUROC) and Area Under the Precision-Recall (AUPRC) curves across all tested timesteps.
  • Quality of Generated Mutations: The generated sequences by PRIEST showed high similarity to actual Q2 2022 sequences, as evidenced by low Euclidean distance (0.368) and cosine dissimilarity (0.024).
  • Immune Escape Potential: PRIEST's generated mutations had a higher Immunescape score (-51.049) compared to random mutations, suggesting a greater potential to escape host antibodies.
  • Identification of High-Risk Sites: The model successfully identified known high-probability mutation sites and also highlighted potential future mutation hotspots, some of which are in critical regions like the RBD of the S-glycoprotein.

Limitations and Future Work

  • Need for Experimental Validation: The paper acknowledges that the predicted immune-evading mutations require further (wet lab) experimental validation.
  • Limited Data for Some Variants: While the methodology is robust, the availability of sufficient evolutionary information for certain virus sequences is a prerequisite for PRIEST's effectiveness.
  • Polygenic Risk Scoring: The authors did not evaluate predictions in the context of Polygenic Risk Scoring (PRS) for associating disease at the individual level.
  • Future Work: The authors suggest expanding PRIEST to handle other viruses and pandemics, and to integrate it with PRS for more comprehensive disease analysis. They also aim to present PRIEST as an all-purpose prediction and analysis framework for pandemic preparedness.

Summary updated

Abstract

Abstract The dynamic evolution of the SARS-CoV-2 virus is largely driven by mutations in its genetic sequence, culminating in the emergence of variants with increased capability to evade host immune responses. Accurate prediction of such mutations is fundamental in mitigating pandemic spread and developing effective control measures. In this study, we introduce a robust and interpretable deep-learning approach called PRIEST. This innovative model leverages time-series viral sequences to foresee potential viral mutations. Our comprehensive experimental evaluations underscore PRIEST’s proficiency in accurately predicting immune-evading mutations. Our work represents a substantial step forward in the utilization of deep-learning methodologies for anticipatory viral mutation analysis and pandemic response.
BibTeX
@misc{saha2023priest,
  author = {Gourab Saha and Shashata Sawmya and Md. Ajwad Akil and Arpita Saha and Sadia Tasnim and Md. Saifur Rahman and M. Sohel Rahman},
  title = {{PRIEST - Predicting viral mutations with immune escape capability of SARS-CoV-2 using temporal evolutionary information}},
  year = {2023},
  doi = {10.1101/2023.08.11.552988},
  howpublished = {bioRxiv},
}
Figure