[Paper Review] FluxGAT: Integrating Flux Sampling with Graph Neural Networks for Unbiased Gene Essentiality Classification

Kieren Sharma, Lucia Marucci+1 more|arXiv (Cornell University)|Mar 27, 2024

Machine Learning in BioinformaticsBiochemistry, Genetics and Molecular Biology3 citations

TL;DR

FluxGAT introduces a graph neural network that predicts gene essentiality directly from flux sampling data, bypassing the need for predefined objective functions in flux balance analysis (FBA). By leveraging graph attention networks on metabolic reaction networks derived from flux sampling, FluxGAT achieves nearly double the sensitivity of FBA (0.576 vs. 0.304) in identifying essential genes in Chinese hamster ovary cells, reducing observer bias and enabling more generalizable predictions across diverse biological systems.

ABSTRACT

Gene essentiality, the necessity of a specific gene for the survival of an organism, is crucial to our understanding of cellular processes and identifying drug targets. Experimental determination of gene essentiality requires large growth screens that are time-consuming and expensive, motivating the development of in-silico approaches. Existing methods predominantly utilise flux balance analysis (FBA), a constraint-based optimisation algorithm; however, they are fundamentally limited by the necessity of a predefined cellular objective function. This requirement introduces an element of observer bias, as the objective function often reflects the researcher's assumptions rather than the cell's biological goals. Here, we present FluxGAT, a graph neural network (GNN) model capable of predicting gene essentiality directly from graphical representations of flux sampling data. Flux sampling removes the need for objective functions, thereby eliminating observer bias. FluxGAT leverages the unique strengths of GNNs in learning representations of complex relationships within metabolic reaction networks. The success of our approach in predicting experimentally determined gene essentiality, with almost double the sensitivity of FBA, explores the possibility of predicting cellular phenotypes in cases when objectives are less understood. Thus, we demonstrate a method for more general gene essentiality predictions across a broader spectrum of biological systems and environments.

Motivation & Objective

• To address the observer bias inherent in flux balance analysis (FBA), which relies on predefined cellular objectives that may not reflect true biological goals.
• To develop a machine learning approach that predicts gene essentiality using intrinsic network structure and flux distribution data, independent of assumed objectives.
• To improve sensitivity in essentiality prediction, especially for non-model organisms and complex systems where objective functions are poorly defined.
• To demonstrate that graph neural networks can effectively learn from flux sampling distributions to classify gene essentiality with higher accuracy than traditional FBA.
• To enable more generalizable and unbiased in-silico prediction of cellular phenotypes in metabolic networks.

Proposed method

• Flux sampling is used to generate a distribution of feasible metabolic flux states, eliminating the need for a predefined objective function.
• A weighted metabolic reaction network is constructed from flux sampling data, where nodes represent reactions and edges represent shared metabolites.
• Graph Attention Networks (GATs) process the network structure, using message-passing mechanisms to learn node-level embeddings based on topological and flux-based features.
• Node features include chemical properties of reactions, such as stoichiometry and thermodynamic constraints, integrated into the GNN input.
• The final node representations are passed through a dense layer with sigmoid activation to perform binary classification of essentiality (essential vs. non-essential).
• The model is trained end-to-end on known essentiality labels from CRISPR screens, with performance evaluated via cross-validation and comparison to FBA.

Experimental results

Research questions

• RQ1 Can gene essentiality be predicted with higher sensitivity by bypassing the need for a predefined objective function in flux balance analysis?
• RQ2 How does a graph neural network trained on flux sampling data compare to FBA in predicting experimentally determined gene essentiality?
• RQ3 Can FluxGAT generalize across diverse biological systems by learning from intrinsic network and flux properties rather than assumed objectives?
• RQ4 What is the contribution of topological and flux-based features in improving essentiality prediction beyond sequence or topology-only methods?
• RQ5 How do attention weights in the GNN reflect biologically meaningful relationships within the reaction network?

Key findings

• FluxGAT achieved a sensitivity of 0.576 in identifying essential genes in the iCHO2291 genome-scale metabolic model, nearly double the 0.304 sensitivity of FBA.
• FluxGAT correctly identified 32 of the 41 essential genes labeled by FBA and discovered an additional 44 essential genes missed by FBA, though with 14 false positives.
• The model maintained high specificity (0.977), comparable to FBA’s 0.985, indicating strong precision in classifying non-essential genes.
• FluxGAT reduced false negatives from 94 (FBA) to 56, significantly improving recall of essential genes.
• The model demonstrated improved generalization by learning from flux distributions rather than relying on observer-defined objectives, enabling broader applicability to non-model organisms.
• Attention mechanisms in FluxGAT revealed biologically relevant reaction relationships, suggesting interpretability potential for future systems biology applications.