A Combined Bayesian and Similarity-Based Approach for Predicting E. coli Biofilm Inhibition by Phenolic Natural Compounds

Screening for biofilm inhibition by purified natural compounds is difficult due to compounds' chemical diversity and limited commercial availability, combined with time- and cost-intensiveness of the laboratory process. In silico prediction of chemical and biological properties of molecules is a widely used technique when experimental data availability is of concern. At the same time, the performance of predictive models directly depends on the amount and quality of experimental data. Driven by the interest in developing a model for prediction of the antibiofilm effect of phenolic natural compounds such as flavonoids, we performed experimental assessment of antibiofilm activity of 320 compounds from this subset of chemicals. The assay was performed once on two Escherichia coli strains on agar in 24-well microtiter plates. The inhibition was assessed visually by detecting morphological changes in macrocolonies. Using the data obtained, we subsequently trained a Bayesian logistic regression model for prediction of biofilm inhibition, which was combined with a similarity-based method in order to increase the overall sensitivity (at the cost of accuracy). The quality of the predictions was subsequently validated by experimental assessment in three independent experiments with two resistant E. coli strains of 23 compounds absent in the initial data set. The validation demonstrated that the model may successfully predict the targeted effect as compared to the baseline accuracy. Using a randomly selected database of commercially available natural phenolics, we obtained approximately 6.0% of active compounds, whereas using our prediction-based substance selection, the percentage of phenolics found to be active increased to 34.8%.

Automated summary

Screening for biofilm inhibition by purified natural compounds is challenging due to their chemical diversity and limited availability. In this study, researchers performed experimental assessment of antibiofilm activity of 320 phenolic natural compounds and trained a predictive model for biofilm inhibition. The model achieved a high accuracy in predicting the targeted effect and significantly improved the discovery rate of active phenolic compounds when compared to random selection.