Extending the druggable kinome by combining machine learning and experimental target profiling

The IDG-DREAM Drug Kinase Binding Prediction Challenge benchmarked machine learning models for predicting kinase inhibitor potencies and identified novel compound-kinase interactions for guiding experimental evaluation. The first paper from the Challenge was just published in Nature Communications.

The IDG-DREAM Challenge was carried out between October 2018 and April 2019 as a crowdsourced community competition to evaluate the power of machine learning (ML) predictive models as systematic and cost-effective means for guiding experimental efforts to map the massive drug-target spaces. The Challenge focused on kinase inhibitors, due to their clinical importance, toward extending the druggability of the human kinome space.

The first Challenge paper was published in Nature Communications, and it describes the benchmarking results of the kinase activity predictions from 212 active Challenge participants. A total of 268 prediction sets were scored, covering a wide range of ML approaches, including linear regularized regression, deep and kernel learning algorithms, and gradient boosting decision trees, which performed the best.

“The IDG-DREAM Challenge attracted participants from all around the world to evaluate their predictive algorithms for kinase inhibitor discovery. Our study serves as a great demonstration of how machine learning models can guide experimental drug screening efforts and lead to novel kinase inhibitor activities”, says Dr. Anna Cichońska from University of Helsinki and Aalto University, the first author of the study.

The collection of the target profiling data for training of the prediction models was based on the open-data web-platform DrugTargetCommons platform housed at the Institute for Molecular Medicine Finland FIMM, University of Helsinki. The crowdsourcing platform enables the community to take part in the bioactivity data extraction, annotation and curation to provide harmonized drug bioactivity profiles and related information for predictive modelling.

“This challenge further demonstrates the operation and usability of Drug Target Commons as a community test-bench and drug activity resource by providing the training data for the participating teams from all around the world”, says Dr. Balaguru Ravikumar from FIMM, another lead author of the study.

The Challenge was implemented as part of a pre-competitive drug discovery project in collaboration with the NIH-funded Illuminating the Druggable Genome (IDG) consortium, using their unpublished kinase activity datasets to systematically evaluate the model predictions. All the models, new bioactivity data, and benchmarking results have been made publicly available in the publication and at DTC.

In the post-Challenge phase, new experimental assays were designed based on the best performing model predictions, supporting model-guided experimental mapping efforts. These experiments identified unexpected activities even for under-studied kinases. The overall objective is to extend the therapeutic application area of approved or abandoned agents (so-called drug repurposing).

“We hope that the IDG-DREAM Challenge will provide a continuously-updated resource for the chemical biology community to prioritize and experimentally test new target activities toward accelerating many drug discovery and repurposing applications”, says FIMM Group Leader Tero Aittokallio.

Original publication: Cichońska, A., Ravikumar, B., Allaway, R.J. et al. Crowdsourced mapping of unexplored target space of kinase inhibitors. Nat Commun 12, 3307 (2021). https://doi.org/10.1038/s41467-021-23165-1

Contact information:

Anna Cichonska, PhD, FIMM & HIIT, Aalto University, anna.cichonska@helsinki.fi

Balaguru Ravikumar, PhD, Aidian Oy, balaguru.ravikumar@helsinki.fi

Tero Aittokallio, PhD, Prof., FIMM Group Leader, tero.aittokallio@helsinki.fi