Inductive Biases in Feature Reduction for QSAR: SHAP vs. Autoencoders

Machine learning models in drug discovery often depend on high-dimensional molecular descriptors, many of which may be redundant or irrelevant. Reducing these descriptors is essential for improving model performance, interpretability, and computational efficiency. This study compares two widely used reduction strategies: SHAP-based feature selection and autoencoder-based compression, within the context of Quantitative Structure-Activity Relationship (QSAR) classification. LightGBM is used as a consistent modeling framework to evaluate models trained on all descriptors, the top 50 and 100 SHAP-ranked descriptors, and a 64-dimensional autoencoder embedding. The results show that SHAP-based selection produces interpretable and stable models with minimal performance loss, particularly when using the top 100 descriptors. In contrast, the autoencoder achieves the highest test performance by capturing nonlinear patterns in a compact, low-dimensional representation, although this comes at the cost of interpretability and consistency across data splits. These findings reflect the differing inductive biases of each method. SHAP prioritizes sparsity and attribution, while autoencoders focus on reconstruction and continuity. The analysis emphasizes that descriptor reduction strategies are not interchangeable. SHAP-based selection is suitable for applications where interpretability and reliability are essential, such as in hypothesis-driven or regulatory settings. Autoencoders are more appropriate for performance-driven tasks, including virtual screening. The choice of reduction strategy should be guided not only by performance metrics but also by the specific modeling requirements and assumptions relevant to cheminformatics workflows.