ACIS: A Machine Learning Platform for Chemical Signature and Toxicity Prediction

The rapid detection of highly toxic compounds, such as chemical warfare agents, toxic industrial compounds, pharmaceutical-based agents, and non-traditional agents, is paramount to industrial and national security. While multiple chemical detection platforms exist, the ability to detect toxic compounds is fundamentally limited by the available reference database of known chemical signatures. Recently, machine learning techniques have been developed to predict both spectroscopic signatures and toxicity, but these nascent capabilities have yet to be matured into a useable product meeting the needs of critical applications such as at the Department of Homeland Security. In Phase I, we will collect and process large datasets of such chemicals, and mature the capabilities for simulating theoretical spectra and toxicity using machine learning tools. Infrared spectroscopy will be the initial platform of interest in Phase I and II. In Phase II, the scope will be further expanded to include other platforms such as mass spectroscopy. The recently developed Chemprop-IR Python library will be used to predict infrared spectra of relevant compounds. Encoded molecular features will also be used in the prediction of the toxicity. As an end product, a prototype app-based software consisting of a user interface and a cloud-based back-end will be demonstrated by Phase II end. Our commercialization strategy involves: i) the development of a standalone software capable of interfacing with a variety of data formats, and ii) interfacing with spectroscopic platform manufacturers to integrate the software into their existing systems. These are complementary approaches which should serve to improve commercialization potential.