ElixirSeeker: A Machine Learning Framework Utilizing Attention-Driven Fusion of Molecular Fingerprints for the Discovery of Anti-Aging Compounds

Abstract

Despite the growing interest in anti-aging drug development, high cost and low success rate pose a significant challenge. We present ElixirSeeker, a new machine-learning framework designed to help speed up the discovery of potential anti-aging compounds by utilizing the attention-driven fusion of molecular fingerprints. Our approach integrates molecular fingerprints generated by different algorithms and utilizes XGBoost to select optimal fingerprint lengths. Subsequently, we assign weights to the molecular fingerprints and employ Kernel Principal Component Analysis (KPCA) to reduce dimensionality, integrating different attention-driven methods. We trained the algorithm using DrugAge database. Our comprehensive analyses demonstrate that 64-bit Attention-ElixirFP maintains high predictive accuracy and F1 score while minimizing computational cost. Using ElixirSeeker to screen external compound databases, we identified a number of promising candidate anti-aging drugs. We tested top 6 hits and found that 4 of these compounds extend the lifespan of Caenorhabditis elegans, including Polyphyllin Ⅵ, Medrysone, Thymoquinone and Medrysone. This study illustrates that attention-driven fusion of fingerprints maximizes the learning of molecular activity features, providing a novel approach for high-throughput machine learning discovery of anti-aging molecules.