TL;DR
This paper introduces structured datasets of compounds with both multi-target (MT) and single-target (ST) activity derived from biological assays. These datasets are crucial for understanding compound promiscuity, a phenomenon vital in drug discovery, and are made publicly available to facilitate further research.
What Problem Does This Paper Address?
In drug discovery, many therapeutic compounds interact with multiple biological targets, a phenomenon known as promiscuity or polypharmacology. Understanding this promiscuity is key to designing more effective and safer drugs, especially for complex diseases. However, analyzing promiscuity requires comprehensive data that links compounds to their activity against various targets. Existing datasets often lack this crucial information, particularly regarding confirmed inactive compounds and experimental test frequencies. This paper addresses this gap by curating and releasing structured datasets that include both MT and ST compound activities, along with information on confirmed inactivity, enabling more robust promiscuity analysis and machine learning studies.
Methods
The authors employed a rigorous data curation process to build their datasets. They extracted compound and target annotations from PubChem BioAssay, focusing on assays for human targets and excluding those with known issues like antitarget activity or high hit rates. A key methodological step involved filtering out potentially problematic compounds using established criteria for pan-assay interference compounds (PAINS), colloidal aggregators, and luciferase inhibitors to ensure data quality. They then defined "promiscuity degree" (PD) as the number of targets a compound was active against. MT-CPDs were identified with a PD threshold (e.g., PD ≥5), and corresponding ST-CPDs (PD=1) were selected based on confirmed activity against a specific target and inactivity against others within a defined group. This ensured that each compound in a group was tested against all targets in that group, preventing underestimation of promiscuity.
Datasets
The paper introduces two primary curated datasets:
- Mixed Assay Set: Comprises compounds from both biochemical and cell-based assays, with an MT-CPD threshold of PD ≥5. This set includes 817 assays and 398 diverse biological targets.
- Biochemical Assay Set: A subset of the mixed set, focusing solely on biochemical assays. This set was further divided into two subsets based on different MT-CPD thresholds:
- PD ≥5: 259 assays and 143 targets.
- PD ≥3: An additional set of biochemical assays was used to increase the dataset size.
Additionally, reduced datasets were created by randomly removing 50% of compound groups or systematically removing nearest-neighbor (NN) groups to assess the impact of data reduction on machine learning predictions. These datasets are made available on the Zenodo open access platform.
Key Insights and Ideas
- Data Quality is Paramount: The meticulous filtering of assays and compounds for potential interference and inconsistencies is highlighted as crucial for reliable promiscuity analysis.
- ST-CPDs are More Frequently Tested: ST-CPDs generally exhibit higher test frequencies compared to MT-CPDs, which is an important consideration when building comparable datasets.
- Nearest Neighbor Relationships Matter: The study demonstrates that nearest neighbor relationships between compounds significantly influence the accuracy of machine learning models for predicting MT vs. ST activity. Removing NN groups led to a larger drop in prediction accuracy than random removal.
- Structured Data Enables Deeper Analysis: The availability of structured datasets with confirmed inactivity and test frequencies allows for more nuanced "local" promiscuity analysis, focusing on specific target combinations, rather than just global assessments.
Key Results
- The curated datasets include a substantial number of compounds: for the mixed assay set with PD ≥5, there are 2858 MT-CPDs and 15,839 ST-CPDs. For the biochemical assay set with PD ≥5, there are 1242 MT-CPDs and 6629 ST-CPDs.
- Machine learning models (SVM, k-NN, RF) achieved high prediction accuracy for distinguishing MT-CPDs from ST-CPDs on the original datasets.
- Random removal of 50% of compound groups resulted in only a small reduction in predictive performance.
- Nearest neighbor removal of 50% of compound groups led to a more substantial decrease in prediction accuracy, underscoring the importance of these relationships.
Limitations and Future Work
The authors acknowledge that the primary limitation of the current datasets stems from the inherent variability and lower confidence of activity measurements in screening assays compared to data derived from equilibrium constant determinations. However, they argue that for promiscuity analysis and exploring structure-promiscuity relationships, qualitative target annotations are sufficient.
Future work will involve using these datasets for "local" promiscuity analysis, examining differences between MT- and ST-CPDs for specific target combinations. They also hope these datasets will be valuable for others exploring various aspects of compound promiscuity from computational and medicinal chemistry perspectives.
Summary updated