Predicting antimicrobial mechanism-of-action from transcriptomes: A generalizable explainable artificial intelligence approach

To better combat the expansion of antibiotic resistance in pathogens, new compounds, particularly those with novel mechanisms-of-action [MOA], represent a major research priority in biomedical science. However, rediscovery of known antibiotics demonstrates a need for approaches that accurately identify potential novelty with higher throughput and reduced labor. Here we describe an explainable artificial intelligence classification methodology that emphasizes prediction performance and human interpretability by using a Hierarchical Ensemble of Classifiers model optimized with a novel feature selection algorithm called Clairvoyance; collectively referred to as a CoHEC model. We evaluated our methods using whole transcriptome responses from Escherichia coli challenged with 41 FDA-approved antibiotics and 9 crude extracts while depositing 306 transcriptomes. Our CoHEC model can properly predict the primary MOA of previously unobserved compounds in both purified forms and crude extracts at an accuracy above 99%, while also correctly identifying darobactin, a newly discovered antibiotic, as having a novel MOA. In addition, we deploy our methods on a recent E. coli transcriptomics dataset in a different strain and a Mycobacterium smegmatis metabolomics timeseries dataset and showcase exceptionally high performance; improving upon the performance metrics of the original publications. We not only provide insight into the biological interpretation of our model but also that the concept of MOA is a non-discrete heuristic with diverse effects for different compounds within the same MOA, suggesting substantial antibiotic diversity awaiting discovery within existing MOA.Author summary: As antimicrobial resistance is on the rise, the need for compounds with novel targets or mechanisms-of-action [MOA] are of the utmost importance from the standpoint of public health. A major bottleneck in drug discovery is the ability to rapidly screen candidate compounds for precise MOA activity as current approaches are expensive, time consuming, and are difficult to implement in high-throughput. To alleviate this bottleneck in drug discovery, we developed a human interpretable artificial intelligence classification framework that can be used to build highly accurate and flexible predictive models. In this study, we investigated antimicrobial MOA through the transcriptional responses of Escherichia coli challenged with 41 FDA-approved antibiotic compounds, 9 crude extracts, and a recently discovered (circa 2019) compound, darobactin, with novel MOA activity. We implemented a highly stringent Leave Compound Out Cross-Validation procedure to stress-test our predictive models by simulating the scenario of observing novel compounds. Furthermore, we developed a versatile feature selection algorithm, Clairvoyance, that we apply to our hierarchical ensemble of classifiers framework to build high performance explainable machine-learning models. Although the methods in this study were developed and stress-tested to predict the primary MOA from transcriptomic responses in E. coli, we designed these methods for general application to any classification problem and open-sourced the implementations in our Soothsayer Python package. We further demonstrate the versatility of these methods by deploying them on recent Mycobacterium smegmatis metabolomic and E. coli transcriptomics datasets to predict MOA with high accuracy.