Abstract
Two-dimensional (2D) chemical fingerprints are essential numerical features for quantifying structural similarity in chemical compounds, a cornerstone of similarity-based virtual screening (VS). This study employs an eigenvalue-based entropy approach to identify "related" fingerprints—those with negligible contributions to the feature matrix's eigenvalue distribution. We examine how these redundant features influence molecular similarity scores calculated via the Tanimoto coefficient. Our findings reveal significant redundancy in standard fingerprint schemes, which tends to artificially suppress similarity scores. Removing these related fingerprints improves the identification of potential hits for compounds with target biological activity.
Introduction
In drug discovery, virtual screening (VS) is guided by the Structure-Activity Relationship (SAR) principle: molecules with similar chemical structures are likely to exhibit similar biological activities. To quantify this similarity, researchers typically use 2D fingerprints—binary vectors where each bit represents the presence or absence of specific structural fragments.
The industry standard for comparing these vectors is the Tanimoto coefficient. However, common fingerprint dictionaries, such as MACCS (166 bits) or Pubchem (881 bits), are predefined and often contain features that are highly correlated or provide no additional information for a specific dataset. These "related fingerprints" act as noise, diluting the similarity signal and potentially causing researchers to overlook valid "hits" during the screening process.
Methodology
We propose an unsupervised method to prune fingerprint dictionaries using eigenvalue-based entropy, identifying features that do not contribute meaningfully to the dataset's complexity.
1. Feature Matrix Analysis
A dataset of $m$ compounds and $n$ fingerprint bits is represented as an $m \times n$ matrix $A$. We focus on the eigenvalue distribution of the matrix $A^T A$.
2. Eigenvalue Entropy ($H$)
The entropy of the eigenvalue distribution (ranging from 0 to 1) indicates the dataset's complexity. A low entropy suggests high redundancy, meaning the data can be described by fewer principal components than the number of bits present.
3. Contribution Measure ($h_i$)
To identify specific "related" fingerprints, we employ a leave-one-out strategy. We systematically remove one fingerprint bit at a time and measure the resulting change in total entropy. A fingerprint is labeled "related" if its removal has a negligible effect on the eigenvalue distribution, indicating its information is already captured by other features (a quasi-linear relationship).
4. Similarity Comparison
We compared Tanimoto scores calculated using the full fingerprint set against "reduced" sets where related fingerprints were removed based on varying threshold levels ($z$).
Experimental Setup
The study utilized the following datasets and schemes:
- HMDB (Human Metabolome Database): A core set of 1,023 unique metabolites found in human blood.
- Fingerprint Schemes: 166-bit MACCS keys and 881-bit Pubchem fingerprints.
- Expert-Validated Set: A subset of 33 drug-like molecular pairs from DrugBank 3.0, where at least 80% of 143 human experts agreed the molecules were structurally similar.
Key Insights and Results
Prevalent Redundancy in Standard Schemes
Public fingerprint schemes were found to be significantly rank-deficient. For the metabolite dataset, the MACCS matrix (166 bits) had a rank of only 144. More strikingly, the Pubchem matrix (881 bits) had a rank of only 377, meaning over half of its features were mathematically redundant in this context. Using our entropy threshold, up to 63% of fingerprints were identified as "related."
The "Dilution" Effect on Similarity
The presence of related fingerprints consistently decreased similarity scores. In the Tanimoto formula, redundant bits cause the denominator to grow faster than the numerator, artificially lowering the perceived similarity. Removing these bits consistently increased average Tanimoto scores across all metabolite classes (drugs, microbial, plant, and endogenous).
Improved Hit Detection
In the expert-validated dataset of 33 similar pairs, removing related bits aligned the numerical scores more closely with human judgment:
- MACCS: The number of pairs meeting a "high similarity" threshold (sim ≥ 0.8) increased from 24 to 25, and one "false negative" (previously scored < 0.5) was eliminated.
- Pubchem: The number of high-similarity pairs increased from 29 to 31.
These results suggest that removing redundant bits can qualitatively change the outcome of a virtual screen by pushing potential hits above the selection threshold.
Conclusion
This eigenvalue-based entropy approach offers a powerful, unsupervised method for feature selection in chemical informatics. Unlike supervised methods, it requires no activity labels, making it ideal for early-stage discovery. By identifying and removing redundant fingerprints, researchers can mitigate the dilution of similarity signals, leading to more accurate Tanimoto scores and more effective identification of biologically active compounds. Future work should explore the impact of this pruning on non-linear dependencies and other similarity metrics like Dice or Cosine coefficients.
Summary updated