A deep learning approach to discover cyclin-dependent kinases 12 (CDK12) inhibitors in breast cancer.

Abstract

e15086 Background: Although anti-HER2 tyrosine kinase inhibitors (TKIs) have significantly prolonged survival and improved prognosis in HER2-positive breast cancer patients, resistance is a constant obstacle leading to TKIs treatment failure and tumor progression. Our previous research shows dual inhibition of HER2/CDK12 will prominently benefit the outcomes of patients with HER2-positive breast cancer by sensitizing the tumors to anti-HER2 TKIs treatment. Unfortunately, there is no high selective CDK12 inhibitor in clinical. This study sought to accelerate the CDK12 inhibitor discovery by a deep learning approach. Methods: We developed a large-scale self-supervised graph neural network to learn 11 million unlabelled compounds for molecular representation. And the model was trained with 233,823 kinase domain sequences by drug-target interaction datasets from the BindingDB database. Then, this transformer architecture was used to predict potential CDK12 inhibitors. The candidate hits compounds with diverse skeletons were divided into 50 clusters by clustering analysis. The representative hits with high transform scores and rational binding mode in each cluster were detected in the homogeneous time-resolved fluorescence CDK12 kinase assay. Candidate clusters with half-maximal inhibitory concentrations (IC50) lower than 200 nM were selected into expanded validation and kinase selective panel detection. xCELLigence RTCA system was adopted to monitor cell growth and sensitivity of breast cancer cells to CDK12 inhibitors in a real-time manner. Results: Our predictive transfer learning model yielded satisfactory predictions on various targets including CDK12. We screened a total of 4,527,236 compounds and recommended 50 clusters of potential CDK12 inhibitors for further enzyme activity detection. The IC50 of seven cluster representative compounds is lower than 10 uM. Expanded screening of 500 compounds in these clusters discovered a lot of hits with highly CDK12 selective inhibition. Half of the new compounds are showed promising proliferation inhibition in breast cancer cell lines. Conclusions: Our study provides a highly efficient and end-to-end deep learning approach to discover highly selective CDK12 inhibitors in breast cancer.