Compound Property Learning Based on Molecular Fingerprints and Complex Network Metrics

Molecular fingerprints map molecular structures into vectorized sequences and can be applied to QSAR/QSPR model learning to predict molecular properties. In this paper, we obtain fingerprint embedding matrix based on molecular structure, while deriving complex network metrics based on molecular graph structure. Inserting complex network metrics into molecular fingerprints, we construct a new molecular embedding fingerprint and construct a better CNN-based QSAR/QSPR model for compounds with the same/similar graph structures. Experiments on relative retention time prediction of PCB dataset demonstrate the effectiveness of the proposed method.