Analyzing hCov Genome Sequences: Predicting Virulence and Mutation

Shashata Sawmya, Arpita Saha, Sadia Tasnim, Md. Toufikuzzaman, Naser A. Anjum, Ali Haisam Muhammad Rafid, Mohammad Saifur RahmanbioRxiv 2020 · DOI

TL;DR

This paper introduces a novel pipeline for analyzing human coronavirus (hCoV) genomes, specifically SARS-CoV-2. It leverages machine learning to classify viral strains as "mild" or "severe" and then uses this information to predict mutations at key genetic sites. This work offers a powerful tool for understanding viral evolution and potential virulence.

What Problem Does This Paper Address?

The COVID-19 pandemic, caused by SARS-CoV-2, has had a devastating global impact. Understanding the virus's genetic makeup is crucial for predicting its behavior, including its virulence and propensity for mutation. While previous research has focused on individual aspects, this paper aims to create a comprehensive pipeline that integrates genome classification with mutation prediction, providing a more holistic view of viral genetic characteristics. The challenge lies in efficiently processing vast amounts of genomic data and identifying biologically significant patterns that can inform these predictions.

Methods

The authors developed a two-phase analysis pipeline.

Phase 1: Virulence Classification: * Traditional Machine Learning: They employed an "Extremely Randomized Trees" classifier to rank features based on their Gini value. Features with a Gini value greater than the mean were then used to train a "LightGBM" classifier. * Deep Learning: Several deep learning models, including vanilla CNN, AlexNet, and InceptionNet, were utilized. These models processed viral genome sequences transformed into either K-mers spectral representations or one-hot vectors.

Phase 2: Mutation Prediction: * Feature Identification: Important features identified during the classification phase were designated as "Sites of Interest" (SoIs). SHAP values and univariate feature selection (SelectKBest) were used to pinpoint these significant genetic regions. * Time-Series Analysis: A CNN-RNN (specifically CNN-LSTM and CNN-bidirectional LSTM) model was developed to predict mutations. This model treats the evolutionary path of the virus across different geographical locations and time steps as a time-series problem. Genome sequences were concatenated from various time steps to represent this evolutionary trajectory.

Datasets

  • Dataset A: 10,179 hCoV genome sequences collected from the GISAID initiative database up to April 24, 2020. This dataset was used for training and testing the classification models.
  • Dataset B: Published genome sequences from several South-Asian countries collected up to June 27, 2020. This dataset was used for analyzing the mutation pipeline.
  • Country-wise Death Statistics: Obtained from the World Health Organization, used to label genome sequences as "mild" or "severe" based on a death threshold.

Key Insights and Ideas

  • LightGBM Outperforms Deep Learning in Classification: Surprisingly, the traditional LightGBM classifier achieved higher accuracy (97%) in classifying viral genomes as mild or severe compared to the deep learning models (84-89%), and did so much faster. This suggests that for this specific classification task, simpler models can be highly effective.
  • Biologically Significant Sites of Interest: The identified "Sites of Interest" (SoIs) were not just statistically significant but also corresponded to known biologically important regions of the SARS-CoV-2 genome, such as those encoding for non-structural proteins (Nsp3) and the Spike protein. This adds biological validation to the machine learning findings.
  • Time-Series Approach for Mutation Prediction: Framing mutation prediction as a time-series problem, where the model learns from the evolutionary progression of the virus across different regions, is a novel and promising approach.
  • Potential Ancestry of South Asian Strains: The mutation prediction pipeline suggested that some US states (California, Massachusetts, Texas, New Jersey, and Maryland) might be ancestral to some South Asian SARS-CoV-2 strains, potentially explaining surges in those regions.

Key Results

  • Classification Accuracy: LightGBM achieved 97% accuracy in classifying genome sequences as mild or severe. Deep learning models achieved 84-89% accuracy.
  • Mutation Prediction Accuracy: The CNN-RNN models showed promising accuracy in predicting mutations at specific Sites of Interest. For example, for positions 24067, 24068, and 24069, they achieved accuracies of 92.33%, 87.43%, and 92.48%, respectively. For the Spike protein region (pos_24065_24069) in Bangladeshi sequences, accuracy ranged from 80-86%.
  • Improved Prediction with More Context: Providing the model with preceding positions in the sequence significantly improved mutation prediction accuracy for the initial sites of interest.

Limitations and Future Work

Limitations: * Environmental and Geographical Factors Ignored: The study did not incorporate environmental or geographical factors that might influence viral spread and mutation. * Site-Specific Prediction Bottleneck: Whole genome mutation prediction is computationally infeasible; the current approach is limited to pre-identified "sites of interest." * Limited Computational Resources: The sample size for deep learning models could not be increased due to computational constraints, potentially limiting their full potential. * Lack of Health Data: Access to detailed health data (recovery rates, healthcare infrastructure, governmental strategies) was limited, which could have improved labeling and analysis.

Future Work: * Detailed Latent Factor Analysis and Improved Labeling: Exploring more sophisticated labeling schemes and performing detailed latent factor analysis could enhance classification and identify more biologically significant SoIs. * New Mutation Prediction Architectures: Experimenting with novel architectures, such as those incorporating transition matrix concepts, could further improve mutation prediction accuracy. * Increased Data for Deep Learning: Feeding more data into the deep learning models is crucial to unlock their full potential, especially given current computational constraints.

Summary updated

Abstract

Abstract Background Covid-19 pandemic, caused by the SARS-CoV-2 genome sequence of coronavirus, has affected millions of people all over the world and taken thousands of lives. It is of utmost importance that the character of this deadly virus be studied and its nature be analyzed. Methods We present here an analysis pipeline comprising a classification exercise to identify the virulence of the genome sequences and extraction of important features from its genetic material that are used subsequently to predict mutation at those interesting sites using deep learning techniques. Results We have classified the SARS-CoV-2 genome sequences with high accuracy and predicted the mutations in the sites of Interest. Conclusions In a nutshell, we have prepared an analysis pipeline for hCov genome sequences leveraging the power of machine intelligence and uncovered what remained apparently shrouded by raw data.
BibTeX
@misc{sawmya2020analyzing,
  author = {Shashata Sawmya and Arpita Saha and Sadia Tasnim and Md. Toufikuzzaman and Naser A. Anjum and Ali Haisam Muhammad Rafid and Mohammad Saifur Rahman and M. Sohel Rahman},
  title = {{Analyzing hCov Genome Sequences: Predicting Virulence and Mutation}},
  year = {2020},
  doi = {10.1101/2020.06.03.131987},
  howpublished = {bioRxiv},
}

Loading related content...

Figure