PNNL

Abstract

Predicting the condition-dependent fitness effects of gene inactivation remains a central challenge in microbial phenomics. We introduce GIMME (Graph Inference for Microbial Dark Matter Exploration), a knowledge graph integrating 7.94 million RB-TnSeq fitness measurements spanning 55,421 genes, 1,770 experiments, and 10 Pseudomonas strains with experimental metadata, metabolic annotations, and pretrained text and protein-sequence representations. We train GFM1 (GIMME Fitness Model 1), a heterogeneous graph attention network (GAT) to predict gene-inactivation fitness effects under three held-out evaluation settings: experiments, genes, and ortholog clusters. Message passing over a subset of the knowledge graph increases performance: In the held-out-gene setting, message passing reduced mean squared error (MSE) by 10.60% and increased Pearson correlation by 12.44% relative to a multi-layer perceptron (MLP) using the same multimodal node-encoder design and decoder template. Performance gains were concentrated among highly connected, conditionally essential genes, where GFM1 predicts %12.5 more true positives, but this advantage disappears under the more restrictive cluster-aware split, where it loses accuracy. Within the subset of test observations whose genes had level-1 EC annotations, the gene-split GFM1 had lower mean absolute error (MAE) than the MLP in all seven classes and showed greater predictive accuracy than a genome-scale metabolic model in a matched Pseudomonas putida benchmark. Together, these results indicate that message passing can leverage knowledge-graph structure to improve gene-fitness prediction when informative relationships to training genes are available. The GIMME knowledge graph and GFM1 represent a step forward in predictive inference of microbial metabolism and a new tool for in silico phenotype prediction.