Tripleknock: predicting lethal effect of three-gene knockout in bacteria by deep learning.

Area of Science:

• Computational Biology
• Systems Biology
• Genomics

Background:

• Investigating multi-gene knockouts is vital for identifying novel antibiotic targets and advancing metabolic engineering.
• Experimental screening of three-gene combinations is time-consuming due to complex interactions.
• Existing computational methods like Flux Balance Analysis (FBA) require extensive model construction and are slow for large-scale screening.

Purpose of the Study:

• To develop a faster, Genome-scale metabolic model (GEM)-independent computational approach for predicting the lethal effects of three-gene knockouts.
• To introduce Tripleknock, a novel predictive model for facilitating genome-wide three-gene knockout screening.

Main Methods:

• Tripleknock was trained using genome-derived protein sequence features from *Escherichia coli* K-12 MG1655.
• Three-gene knockout lethality was simulated using FBA, defining lethal effects by a >=90% reduction in cell growth.
• Model performance was evaluated using cross-species F1 scores and benchmarked against essentiality rules and curated literature data.

Main Results:

• Tripleknock predicts three-gene knockout lethality approximately 20 times faster than FBA.
• The model achieved an average cross-species F1 score of 0.77 on six *Enterobacteriaceae* pathogens.
• External validation on *E. coli* triple perturbations showed no false positives among predicted lethal cases (FP=0).

Conclusions:

• Tripleknock provides a significantly accelerated and GEM-independent alternative for predicting bacterial triple-gene knockout lethality.
• The developed framework establishes a reproducible baseline for evaluating bacterial triple-knockout lethality prediction methods.
• This approach facilitates faster discovery of essential genes for antibiotic targets and metabolic engineering applications.