Why Deep Models Often cannot Beat Non-deep Counterparts on Molecular Property Prediction?

Specifically, the Multi-Layer Perceptron (MLP) could be applied to Molecular property prediction (MPP) is a crucial computed or handcrafted molecular fingerprints; Sequencebased task in the drug discovery pipeline, which has recently neural architectures including Recurrent Neural Networks gained considerable attention thanks to advances (RNNs) (Medsker & Jain, 1999), 1D Convolutional in deep neural networks. However, recent Neural Networks (1D CNNs) (Gu et al., 2018), and Transformers research has revealed that deep models struggle (Honda et al., 2019; Rong et al., 2020) are exploited to beat traditional non-deep ones on MPP. In this to encode molecules represented in Simplified Molecular-study, we benchmark 12 representative models Input Line-Entry System (SMILES) strings (Weininger (3 non-deep models and 9 deep models) on 14 et al., 1989). Later, it is argued that molecules can be naturally molecule datasets. Through the most comprehensive represented in graph structures with atoms as nodes and study to date, we make the following key observations: bonds as edges.