Accelerating drug combination discovery with machine learning

Achieving scale through machine learning and sound waves

To achieve scale, Combocat integrates miniaturized drug dispensing with machine learning. For drug dispensing, sonic technology enables customized and efficient experimental layouts. “We incorporated acoustic liquid handlers that use sound waves to transfer tiny droplets of drugs very precisely,” Wright said. “They use the exact minimum of each liquid you need, allowing for the use of far less material than conventional pin or pipette-based techniques, and increasing the number of testable combinations.”

Machine learning helps fill in the picture, using one of Combocat’s two modes, “dense mode” to inform its “sparse mode.” In dense mode, the researchers measured every possible dose pairing for each drug combination. Alongside this dense mode, a sparse mode allows for even tighter resource management by predicting the full results from only a small fraction of the original data. The sparse mode model was trained on hundreds of drug combination experiments generated with the platform’s dense mode approach. When the researchers compared the machine learning predictions to measured values, they found them to be highly consistent, providing confidence in the approach.

“We created two screening ‘modes’, with something of a tradeoff between them,” Geeleher said. “We optimized the sparse mode approach for scale, but it trades detail for efficiency, while we optimized the dense mode to obtain ultra-reliable measurements, which can’t scale. However, we showed that Combocat can combine them to analyze more combinations and validate them more rapidly than traditional approaches.”

Combocat continues a St. Jude legacy of drug combination therapy innovation, providing a way to easily screen combinations not just for cancer researchers, but for any disease in need of new treatments.

“We’ve created a platform that’s free, open-source and highly usable that could become a strong standard in the drug combination discovery field,” Geeleher said. “Combocat can help expedite the identification of potentially safe and effective drug combinations, which could ultimately yield useful and potentially practice-changing new drug combinations in the clinic.”

Combocat can be accessed at combocat.stjude.org. 

Authors and funding

The study’s other authors are Min Pan, Gregory Phelps, Jonathan Low, Duane Currier, Ankita Sanjali, Marlon Trotter, Jihye Hwang, Richard Chapple, Xueying Liu, Declan Bennett, Yinwen Zhang, Richard Lee and Taosheng Chen, all of St. Jude.

The study was supported by grants from National Institute of General Medical Sciences (R35GM138293), National Cancer Institute (R01CA260060), National Human Genome Research Institute (R00HG009679) and ALSAC, the fundraising and awareness organization of St. Jude.