TL;DR
Molecular fingerprints, a common tool for virtual drug screening, are largely ineffective at identifying diverse active drug molecules from massive libraries. Despite their widespread use, fingerprint similarity offers poor discrimination between active and inactive compounds, and when it does find potential hits, they often share the same core structure as the starting molecule, limiting the discovery of truly novel scaffolds.
What Problem Does This Paper Address?
The quest for new drugs often involves sifting through billions of potential small molecules to find those that can interact with a specific biological target (like a protein). A common strategy is "virtual screening," where computational methods predict which molecules are most likely to be active. A key assumption in many of these methods is the "similar property principle" (SPP): molecules with similar structures are likely to have similar biological activities. Molecular fingerprints are widely used to quantify this structural similarity, generating compact digital representations of molecules that can be quickly compared. However, the authors question whether these fingerprints are truly effective for large-scale screening, especially for identifying diverse active molecules, and whether they can reliably predict potency or even distinguish actives from decoys in a meaningful way.
Methods
The researchers evaluated the performance of 32 different molecular fingerprints across four established virtual screening benchmark datasets. They primarily used two metrics:
- ROC AUC (Area Under the Curve): A standard measure of a classifier's ability to distinguish between classes (actives vs. decoys) across all possible thresholds.
- DRF (Decoy Retention Factor): A novel metric introduced by the authors to specifically assess early enrichment. It quantifies how effectively decoys are filtered out when looking at the top fraction of molecules ranked by similarity. A DRF of 1 means decoys are retained at the same rate as actives, while a DRF of 0 indicates perfect separation.
They also analyzed the distribution of Tanimoto similarity scores (a common measure of fingerprint similarity) between active-active and active-decoy pairs. Additionally, they investigated the correlation between fingerprint similarity and compound potency (how strongly a molecule binds to its target) using AC50 values.
Datasets
The study utilized four widely recognized virtual screening benchmark datasets:
- DUD-E (Directory of Useful Decoys, Enhanced): A large dataset with ~224 actives and ~14,000 computational decoys per target.
- DEKOIS (Demanding Evaluation Kits for Objective In silico Screening): A dataset derived from BindingDB, with a focus on demanding evaluation for in silico screening.
- MUV (Maximum Unbiased Validation): A dataset based on PubChem bioactivity data, designed to provide an unbiased spread of actives and decoys.
- LIT-PCBA: A curated subset of PubChem BioAssay data, featuring large numbers of actives and inactives for certain targets.
Key Insights and Ideas
- Fingerprints struggle with true novelty: When fingerprints identify molecules similar to a known active, these "new" actives often share the same core chemical scaffold. This means they could likely be found through simpler, structure-based enumeration rather than sophisticated similarity searching, limiting the discovery of entirely new chemical classes.
- Similarity doesn't guarantee potency: Even among molecules identified as active, fingerprint similarity scores did not correlate well with their actual potency (how strongly they bind to the target). This suggests fingerprints are poor predictors of drug efficacy.
- The "activity cliff" is a major hurdle: The paper implicitly highlights the challenge of "activity cliffs," where minor structural changes can lead to significant differences in biological activity. Fingerprints, being primarily structural descriptors, may not capture these subtle but critical differences.
- The need for context: The authors argue that molecular similarity for biological activity is context-dependent, relying on the interaction with the specific target protein. Fingerprints, which often focus on intrinsic molecular features, may not adequately capture this crucial context.
Key Results
- Poor enrichment: Across all 32 fingerprints tested and four benchmark datasets, performance in enriching for active molecules was generally poor. AUC values were often below 0.6, and DRF0.1 values were close to 1.0, indicating minimal enrichment of actives over decoys.
- Indistinguishable distributions: For most targets, the distributions of Tanimoto similarity scores between active-active pairs and active-decoy pairs were largely overlapping, making it difficult to set a similarity threshold that effectively separates them.
- Limited correlation with potency: Kendall rank correlation coefficients between fingerprint similarity and compound potency (AC50) were generally low, ranging from -0.53 to 0.54, with most values close to zero. This indicates fingerprints are not reliable for ranking active compounds by potency.
- Scaffold hopping is limited: For the few cases where early enrichment was observed, the identified "novel" actives were typically bioisosteres or minor variations of the query molecule, not truly novel scaffolds.
Limitations and Future Work
The authors acknowledge that their study focuses on ligand-based virtual screening and that other methods, such as structure-based approaches or deep learning, might offer different capabilities. They suggest that future successful strategies will need to move beyond simple fingerprint representations. This will likely involve representing molecules in a way that incorporates the context of the binding target, potentially by focusing on surface properties and representing both molecules and targets as collections of interacting patches. This would allow for a more context-dependent emphasis on features relevant to specific interactions.
Summary updated