Harvard Medical School investigators have developed a machine learning approach using high-quality, large-scale animal model data that sheds new light on the biology of the liver and kidneys after toxin exposure.

The findings were recently published in Molecular Systems Biology, and reveal new mechanisms of toxin vulnerability and tolerance that may be broadly relevant to studies of human disease, the authors said.

Researchers found nine distinct patterns of response to chemical exposure that the authors termed "disease states."

These states shed light on the dynamics of toxin-induced liver and kidney injury, including defense mechanisms and novel biomarkers, and provide insights into molecular signals that cause toxin-induced appetite suppression and weight loss.

"We used machine learning to ask a simple question: What can we learn from this rich data set about what happens to the liver and kidneys after exposure to different chemicals?,” lead study author Kenichi Shimada, HMS research fellow in therapeutic science in the Laboratory of Systems Pharmacology, said in a release.

Researchers focused on the Open TG-GATEs database, the result of a 10-year effort by a Japanese public-private consortium to assess 170 different compounds with the aim of improving and enhancing drug safety. These compounds represent a wide range of chemicals and medications, including common ones such as ibuprofen and acetaminophen, known for their toxic effects on the liver and kidneys at high doses.

Each compound was administered at multiple dosages and time points to rats, as well as to rat and human liver cells grown in culture. For each of these treatment conditions, a variety of measures were collected, including blood chemistry, physiological measures such as body and tissue weight, histology and gene expression data.

To identify commonalities and patterns in how the liver and kidneys respond to different drugs, the research team developed an unsupervised machine-learning approach in which a computational algorithm - without relying on predefined questions, labels or categorizations in order to avoid researcher-introduced bias - analyzed data on 160 compounds administered in rats, representing more than 3,500 treatment conditions.