Explanatory AI for Phenotype Prediction

In recent years our ability to generate multi-omics data has greatly outpaced our ability to interpret it- While machine learning approaches can process massive quantities of data, they frequently make mistakes that can be costly in applications like microbial enhanced oil recovery where “checking the AI’s work” is difficult- We seek to mitigate the impact of these errors by enabling the AI to collaborate with human subject matter experts- To enable AI systems to collaborate with a human, our approach will learn a generative model for a subset of the multi-omics pathway based on formal grammars- Unlike neural networks, grammar-based approaches produce a hierarchical structure that can be interpreted as descriptions of the output at various levels of fidelity- We will leverage this structure to provide subject matter experts with both explanations (the hierarchy) and predictions (the lowest level symbols)-