SBIR-STTR Award

Direct to Phase IINew pathogens, both naturally occurring and adversary-engineered, are increasingly likely to emerge and represent a significant and growing risk to global health and security. These new threats often have limited genetic similarity to prior known pathogens and cannot be identified through standard genetic tests. The application of machine learning algorithms to phenotypic tests to predict pathogenic potential will face challenges in the integration of heterogeneous data sources, and the application of machine learning algorithms to sparse, inconsistent datasets. We propose to build an advanced computational platform called PathEngine that will rapidly ingest and integrate measurements of phenotypic tests from conventional microtiter plate assays, as well as single-cell resolution microfluidics systems. It will use a tailored semi-supervised learning algorithm to predict the pathogenic potential of bacterial strains from limited, sparse, inconsistent training datasets. PathEngine will ingest, integrate, and analyze phenotypic tests of three different categories (harming a host, niche finding, and self-preservation) and be capable of identifying the pathogenic potential of bacteria at >90% accuracy.

Netrias, the Texas A&M Health Science Center (TAMHSC) and the Texas A&M Engineering Experiment Station (TEES), and expert consultants will expand and extend the capabilities of PathEngine, an advanced computational platform that ingests and integrates a corpus of bacterial phenotype measurements suitable for the training of a pathogenicity machine learning algorithm. We will enhance the data integration technology and advanced machine learning techniques of PathEngine to rapidly combine and prioritize relevant host and microorganism phenotypic factors to create an accurate, robust, and comprehensive model of pathogenicity for users with little to no computational experience. We will train the platform on 50 additional strains, adding to the 20 strains that were used in Phase I. We will automatically identify and ingest subsets of genomic data from PATRIC, such as virulence factors, and automate the execution of algorithms that are capable of robustly identifying the pathogenic potential of bacteria at >90% accuracy. State-of-the-art methods currently only achieve this accuracy by predicting self-preservation traits using whole genome sequence data.