Prediction of gene essentiality using machine learning and genome-scale metabolic models

Lilli Freischem, Mauricio Barahona, Diego A. Oyarzún

TL;DR

This paper introduces a novel machine learning approach to predict gene essentiality by analyzing the connectivity of metabolic reactions in a "mass flow graph" derived from wild-type cell simulations. This method bypasses the need to assume that gene-deleted cells optimize the same objectives as wild-type cells, offering a more flexible and potentially more accurate way to identify essential genes.

What Problem Does This Paper Address?

Identifying essential genes – those critical for cell survival – is vital for fields like medicine and biotechnology. Traditional methods involve costly and time-consuming experimental screens. While computational methods like Flux Balance Analysis (FBA) exist, they often rely on the assumption that cells with deleted genes will still try to maximize their growth rate. This assumption can fail when gene deletions drastically alter cellular physiology, forcing cells to pursue different survival objectives. This paper tackles the challenge of predicting gene essentiality without making these potentially flawed assumptions about the behavior of gene-deleted strains.

Methods

The core of the proposed method involves two main steps:

  1. Mass Flow Graph (MFG) Construction: The authors first generate a "mass flow graph" from the wild-type cell's metabolic network. In this graph, each metabolic reaction is a node, and edges represent the flow of chemical mass between reactions. The weight of these edges is determined by the flux distribution calculated from FBA of the wild-type cell. This graph captures how reactions are interconnected in terms of material transfer.
  2. Machine Learning Classification: They then extract features from the connectivity of these nodes in the MFG. Specifically, they use the incoming and outgoing edge weights for each reaction node as features. These features, along with experimentally determined essentiality labels (essential or non-essential), are used to train binary classifiers (e.g., Random Forest, Logistic Regression). The trained classifiers can then predict the essentiality of genes based on their corresponding reaction's connectivity in the MFG.

What's novel here is using the wild-type flux distribution to build the graph and extract features, rather than simulating FBA for each gene deletion. This avoids the problematic assumption about deletion strain optimization.

Datasets

The study uses the most complete metabolic model of Escherichia coli (iML1515) for constructing the MFG. For essentiality labels, they utilize growth assay data from E. coli strain BW25113. This data quantifies the impact of single gene knockouts on cellular growth. The authors map these gene-level essentiality labels to reaction-level labels within the iML1515 model, creating a dataset of 255 reactions with assigned essentiality.

Key Insights and Ideas

  • Wild-type FBA holds predictive power: The study demonstrates that the metabolic flux distribution of a wild-type cell, even without assuming optimality of deletion strains, contains sufficient information to predict gene essentiality.
  • Graph connectivity as a feature: Using the structure and connectivity of a mass flow graph derived from metabolic fluxes proves to be a powerful way to represent metabolic reactions for machine learning.
  • Flexibility beyond optimality assumptions: By decoupling the prediction from the assumption of growth maximization in deletion strains, this method can potentially handle more complex cellular responses to gene knockouts.

Key Results

The Random Forest classifier, trained using features derived from the MFG, achieved a test set accuracy of 76% with a precision of 82.5% and a recall of 86.8%. While slightly lower than FBA predictions (84.3% accuracy, 89.5% precision/recall), these results are considered "near state-of-the-art" and are particularly promising given the method's reduced assumptions. The classifier showed particularly strong performance in predicting essential genes.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Limited environmental conditions: The study was conducted under a single growth condition (aerobic growth on glucose). Future work should explore performance across various conditions to identify conditionally essential genes.
  • Imbalanced datasets: The method shows comparatively poor accuracy for predicting non-essential genes, likely due to their underrepresentation in FBA solutions. Strategies like under/over-sampling or adjusted loss functions are suggested to address this.
  • Generalizability: Further testing is needed to establish the method's validity across different organisms and environmental conditions.
  • Gene-reaction mapping: The current approach relies on gene-protein-reaction (GPR) rules, and a more comprehensive mapping between genes and reactions would improve coverage.

Summary updated

Abstract

The identification of essential genes, i.e. those that impair cell survival when deleted, requires large growth assays of knock-out strains. The complexity and cost of such experiments has triggered a growing interest in computational methods for prediction of gene essentiality. In the case of metabolic genes, Flux Balance Analysis (FBA) is widely employed to predict essentiality under the assumption that cells maximize their growth rate. However, this approach assumes that knock-out strains optimize the same objectives as the wild-type, which excludes cases in which deletions cause large physiological changes to meet other objectives for survival. Here, we resolve this limitation with a novel machine learning approach that predicts essentiality directly from wild-type flux distributions. We first project the wild-type FBA solution onto a mass flow graph, a digraph with reactions as nodes and edge weights proportional to the mass transfer between reactions, and then train binary classifiers on the connectivity of graph nodes. We demonstrate the efficacy of this approach using the most complete metabolic model of Escherichia coli, achieving near state-of-the art prediction accuracy for essential genes. Our approach suggests that wild-type FBA solutions contain enough information to predict essentiality, without the need to assume optimality of deletion strains.
BibTeX
@inproceedings{freischem2022prediction,
  author = {Lilli Freischem and Mauricio Barahona and Diego A. Oyarzún},
  title = {{Prediction of gene essentiality using machine learning and genome-scale metabolic models}},
  booktitle = {IFAC-PapersOnLine},
  year = {2022},
  doi = {10.1016/j.ifacol.2023.01.006},
}

Loading related content...

Figure