Fig. 1: The use of combined machine learning strategies to identify novel mitophagy inducers.
a, The workflow for model pre-training: (i) Molecules within the pre-training dataset were transferred into SMILES sequences, molecular interaction features and 3D conformers fingerprint in the data preparation stage; (ii) Three encoders (for 1D, 2D and 3D representations) were then designed to embed the input data, and these representational embeddings were aggregated into the encoder model of the multi-representation; (iii) The multi-representational embeddings were then passed to the representation decoder to pre-train the multi-representation molecule model. ‘F’ and ‘G’ stand for ‘Functional encoder’ and ‘Generator’ respectively. b, The workflow for the virtual screening process: (i) The virtual screening library contained 3,274 molecules from a traditional Chinese medicine dataset, named Macau Library; (ii) The 1D, 2D and 3D molecular representations for each compound were generated on the basis of the pre-trained molecule representation models; (iii) The representations were then aggregated and clustered, and a hyper-space filter was applied to the representations to filter out outliers; (iv) The similarity scores for each compound were calculated to generate the top N candidate compounds.
