Abstract
In the early stages of drug discovery, computational models are used to predict activity and prioritize compounds for experimental testing. However, new targets commonly lack the data necessary to build effective models, and the screening needed to generate that experimental data can be costly. This study introduces Informer-Based-Ranking (IBR), a strategy to solve the cold-start problem in drug discovery. By screening a small, strategically selected informer set of compounds, researchers can use prior bioactivity data to accurately predict which of the remaining thousands of compounds will hit a brand-new biological target. 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.
Introduction
In the early stages of drug discovery, researchers often face a cold-start target: a protein or enzyme they want to inhibit, but for which they have zero experimental data. To find a hit (an active compound), they usually have to screen massive libraries. Because screening is expensive and time-consuming, the standard practice is to pick a small initial batch based on chemical diversity—choosing molecules that look structurally different from one another. However, chemical similarity does not always equal biological similarity. This paper addresses the inefficiency of this blind initial screening by asking: Can we use the historical behavior of compounds across other targets to pick a smarter starting set?
Methods
The authors propose three novel IBR strategies that treat the drug discovery process like a recommendation engine (similar to how Netflix predicts what you’ll like based on other users' habits):
- Regression Selection (RS): This method clusters known targets based on their bioactivity profiles. It then uses multinomial regression to identify a handful of informer compounds whose results are most predictive of which cluster a new target belongs to.
- Coding Selection (CS): Drawing from information theory, this method treats a compound’s binary activity (active/inactive) across targets as a code word. It seeks an informer set that creates unique, high-resolution fingerprints for different target behaviors.
- Adaptive Selection (AS): This starts with a base informer set and grows it by iteratively adding compounds that minimize the distance between the informers and the rest of the library in bioactivity space.
These are compared against Chemometric baselines, which rely purely on chemical structures (fingerprints) to pick diverse sets.
Datasets and Validation
The study utilizes the Published Kinase Inhibitor Sets (PKIS1 and PKIS2), which are high-quality chemogenomic matrices containing the activity of hundreds of compounds against hundreds of human kinases. To prove the model works in the real world, the authors performed prospective screening on three novel microbial kinases that were not in the original datasets: PknB (from Mycobacterium tuberculosis), BGLF4 (from Epstein-Barr virus), and ROP18 (from Toxoplasma gondii).
Key Insights and Results
- Bioactivity over Chemistry: A compound's historical track record across various proteins is a much more powerful predictor of its future performance than its 2D chemical structure. In retrospective tests, the RS method achieved a median ROCAUC of 0.92, dwarfing the standard chemical-diversity baseline of 0.67.
- Target Clustering: Biological targets fall into functional families. If an informer set can quickly identify which family a new target belongs to, the model can instantly prioritize thousands of other compounds known to hit that family.
- Scaffold Hopping: IBR strategies excelled at identifying diverse active structures that traditional chemical-similarity models missed, facilitating the discovery of molecules that work but look structurally distinct from known drugs.
- High Enrichment: For the target BGLF4, the AS method identified 10 out of 11 total hits in the library by only looking at the top 10% of the ranked list.
Conclusion
Informer-Based-Ranking provides a robust solution to the cold-start problem in drug discovery. Even though the training data was based on human kinases, the models successfully predicted inhibitors for microbial kinases that were phylogenetically distant, proving the method's generalizability. While the requirement for a pre-existing bioactivity matrix limits its use to well-studied families, the approach significantly improves the efficiency of prioritizing compounds for screening.
Summary updated