Predicting kinase inhibitors using bioactivity matrix derived informer sets

Huikun Zhang, Spencer S. Ericksen, Ching-pei Lee, Gene E. Ananiev, Nathan Wlodarchak, Julie C. Mitchell, Anthony Gitter, Stephen J. Wright, F. Michael Hoffmann, Scott A. Wildman, Michael A. Newton

Abstract

Prediction of compounds active against a biological target is a critical step in drug discovery. When data for a novel target is scarce, initial screening must provide the necessary training data for supervised learning models. Traditional methods select initial libraries based on chemical diversity or random selection, which often fail to yield sufficient active compounds. We propose Informer-Based-Ranking (IBR), an approach that utilizes chemogenomic data from previous screenings to select a small "informer set" of compounds. By testing these informers on a novel target and combining the results with historical bioactivity data, we can accurately predict the activity of the remaining library. Prospective tests on two kinase targets (PknB and BGLF4) demonstrate that bioactivity fingerprints outperform traditional chemical fingerprints, enabling the discovery of novel chemical classes through "scaffold hopping."

Introduction: The Cold-Start Problem in Drug Discovery

In the early stages of drug discovery, researchers use virtual screening to prioritize compounds for experimental testing. However, effective computational models typically require existing data on the specific target of interest. For a brand-new target, this creates a "cold-start" problem: how do you choose the first compounds to test when you have no prior information?

Standard approaches involve screening a chemically diverse subset or a random selection of plates. These methods often fail to provide the high-quality data needed to train predictive models for the rest of the library. The IBR strategy addresses this by asking: can we strategically select a tiny fraction of the library—just 16 compounds—to jumpstart highly accurate predictions for the entire set?

Methodology: Informer-Based-Ranking (IBR)

The core innovation of IBR is the shift from chemical fingerprints (what a molecule looks like) to bioactivity fingerprints (how a molecule behaves across hundreds of other targets). By leveraging a "bioactivity matrix" of historical screening data, the authors developed three strategies to select informer sets:

  1. Regression Selection (RS): Uses statistical techniques to identify compounds that best predict a target's cluster membership.
  2. Coding Selection (CS): A non-parametric approach treating activity patterns as "code words" to partition the target space.
  3. Adaptive Selection (AS): Iteratively selects compounds that maximize coverage of the known bioactivity space.
The Informer-Based-Ranking (IBR) workflow: historical bioactivity data is used to select an informer set, which is then tested against a novel target to generate predictions for the rest of the library.

Once the informer set is tested, the new data is integrated with the chemogenomic matrix to rank the remaining compounds.

Key Results and Experimental Validation

The researchers validated IBR using the Published Kinase Inhibitor Sets (PKIS1 & PKIS2), which contain activity data for hundreds of compounds against hundreds of human kinases. To test the model's real-world utility, they conducted prospective laboratory experiments on two kinases not present in the original datasets: PknB (from Mycobacterium tuberculosis) and BGLF4 (from Epstein-Barr virus).

Performance Metrics

In retrospective "leave-one-out" tests on PKIS data, IBR methods significantly outperformed chemical-similarity baselines. As shown in the performance summary below, non-baseline methods like Adaptive Selection (AS) achieved substantially higher median ROCAUC scores compared to traditional chemical structure-based approaches.

Summary statistics for retrospective performance on the PKIS1 dataset, showing that non-baseline IBR methods (RS, CS, AS) consistently achieve higher mean and median ROCAUC scores compared to baselines (BCw, BFw).

Scaffold Hopping and Diversity

Because IBR focuses on biological behavior rather than molecular shape, it successfully identifies active compounds from novel chemical classes—a process known as "scaffold hopping." Prospective lab tests confirmed these results: for both PknB and BGLF4, the RS and CS methods were the only ones to successfully recover multiple hits and diverse chemical classes within the top 10% of their rankings.

Conclusion

Informer-Based-Ranking demonstrates that bioactivity data is a more powerful predictor of target activity than chemical structure alone. Testing just 16 compounds (roughly 4% of the library) was sufficient to achieve high predictive accuracy, offering substantial savings in experimental costs. While the current study focused on kinases, future work will explore the scalability of IBR to ultra-large virtual libraries and its applicability to other protein families.

Summary updated

Abstract

Abstract Prediction of compounds that are active against a desired biological target is a common step in drug discovery efforts. Virtual screening methods seek some active-enriched fraction of a library for experimental testing. Where data are too scarce to train supervised learning models for compound prioritization, initial screening must provide the necessary data. Commonly, such an initial library is selected on the basis of chemical diversity by some pseudo-random process (for example, the first few plates of a larger library) or by selecting an entire smaller library. These approaches may not produce a sufficient number or diversity of actives. An alternative approach is to select an informer set of screening compounds on the basis of chemogenomic information from previous testing of compounds against a large number of targets. We compare different ways of using chemogenomic data to choose a small informer set of compounds based on previously measured bioactivity data. We develop this Informer-Based-Ranking (IBR) approach using the Published Kinase Inhibitor Sets (PKIS) as the chemogenomic data to select the informer sets. We test the informer compounds on a target that is not part of the chemogenomic data, then predict the activity of the remaining compounds based on the experimental informer data and the chemogenomic data. Through new chemical screening experiments, we demonstrate the utility of IBR strategies in a prospective test on two kinase targets not included in the PKIS. Using limited training data in both retrospective and prospective tests, bioactivity fingerprints based on chemogenomic data outperform chemical fingerprints in predicting active compounds in both standard virtual screening metrics and accurate identification of hits from novel chemical classes. Author Summary In the early stages of drug discovery efforts, computational models are used to predict activity and prioritize compounds for experimental testing. New targets commonly lack the data necessary to build effective models, and the screening needed to generate that experimental data can be costly. We seek to improve the efficiency of the initial screening phase, and of the process of prioritizing compounds for subsequent screening. We choose a small informer set of compounds based on publicly available prior screening data on distinct (though related) targets. We then use experimental data on these informer compounds to predict the activity of other compounds in the set against the target of interest. Computational and statistical tools are needed to identify informer compounds and to prioritize other compounds for subsequent phases of screening. Using limited training data, we find that selection of informer compounds on the basis of bioactivity data from previous screening efforts is superior to the traditional approach of selection of a chemically diverse subset of compounds. We demonstrate the success of this approach in retrospective tests on the Published Kinase Inhibitor Sets (PKIS) chemogenomic data and in prospective experimental screens against two additional non-human kinase targets.
BibTeX
@article{zhang2019predicting,
  author = {Huikun Zhang and Spencer S. Ericksen and Ching-pei Lee and Gene E. Ananiev and Nathan Wlodarchak and Julie C. Mitchell and Anthony Gitter and Stephen J. Wright and F. Michael Hoffmann and Scott A. Wildman and Michael A. Newton},
  title = {{Predicting kinase inhibitors using bioactivity matrix derived informer sets}},
  journal = {PLOS Computational Biology},
  year = {2019},
  url = {https://openreview.net/forum?id=uCBgNRkATYz},
}
Figure