Entropy of a bacterial stress response is a generalizable predictor for fitness and antibiotic sensitivity.

Genes implicated in bacterial stress responses have been used to construct models that infer the growth outcome of a bacterium in the presence of antibiotics with the objective to develop novel diagnostic methods in the clinic. Current models are trained on data specific to a species or type of stress, making them potentially limited in their application. It is unclear if a generalizable response-signature exists that can predict bacterial fitness independent of strain, species or type of stress. Here we present a substantial RNA-Seq and experimental evolution dataset for 9 strains and species, under multiple antibiotic and non-antibiotic stress conditions. We show that gene panel-based models can accurately predict antibiotic mechanism of action, as well as the fitness outcome of Streptococcus pneumoniae in the presence of antibiotics or under nutrient depletion. However, these models quickly become species-specific as gene homology is limited. Instead, we define a new concept, transcriptomic entropy, which we use to quantify the amount of transcriptional disruption that occurs in a bacterium when responding to the environment. With entropy at the center, we train a suite of predictive (machine learning) models enabling generalizable fitness and antibiotic sensitivity predictions. These entropy-based models that predict bacterial fitness are validated for 7 Gram-positive and -negative species under antibiotic and non-antibiotic conditions indicating that transcriptomic entropy can be used as a generalizable stress signature. Moreover, rather than being a binary indicator of fitness, an entropy-based model was developed and validated to predict the minimum inhibitory concentration of an antibiotic. Lastly, we show that the inclusion of a varied-set of multi-omics features of a bacterial stress response further enhances fitness predictions by reducing ambiguity. By demonstrating the feasibility of generalizable predictions of bacterial fitness, this work establishes the fundamentals for potentially new approaches in infectious disease diagnostics, including antibiotic susceptibility testing.\n\nSignificance statementAccurate predictions of bacterial fitness outcome could potentially have clinical diagnostic value, such as predicting optimum antibiotic choice and dosage for treating infectious diseases. Existing models of fitness predictions rely mainly on gene panel approaches, which may be species- and stress-specific due to a lack of gene and response conservation. In order to overcome this limitation, we generated a substantial experimental dataset and identified entropy as a universal stress response signature that quantifies the level of transcriptional disruption that is indicative of fitness outcome under a stressful condition. We present and validate for Gram-positive and negative species a suite of entropy-based models that enable accurate predictions of fitness outcome and the level of antibiotic sensitivity in a species and stress-type independent manner.