Pathogen Classification Tool (PACT)

Stottler Henke proposes PACT to address the threat posed by unknown/novel bacteria. Stottler Henke’s solution leverages AI/ML technologies to assess the pathogenic potential of unknown/novel bacteria for DARPA’s Biological Technologies Office. Threat assessment is inferred from phenotype as characterized by a series of assays developed by Harvard University as part of DARPA’s Friend or Foe program. These assays are designed specifically to target various aspects of a bacteria’s phenotype that are heavily correlated with pathogenicity. Examples of the phenotypic features captured in the data include: cellular growth, viability in the presence of different media, niche finding, immune avoidance, & cytotoxicity. We expect, and have observed it to be the case in previous work, that our model will be able to learn a better decision boundary from this carefully curated feature set. The PACT system features a novel semi-supervised neural architecture that is capable of learning from both labeled and unlabeled data. This concept of using an unsupervised task to improve neural network performance has driven an entire thread of neural network research leading to state-of-the-art performance on various tasks. The ultimate goal for this technology is to improve national security by furthering the state-of-the-art in biosurveillance/biodefense technology.