Reversible Deep Learning for 13C NMR in Chemoinformatics

A novel reversible deep learning framework for 13C NMR spectroscopy utilizes a single conditional invertible neural network to facilitate bidirectional mapping between molecular structures and their corresponding spectra. This network incorporates i-RevNet style bijective blocks, allowing for both forward predictions and inverse generation from the same trained model. It generates a 128-bit binned spectrum code based on graph-encoded structures, with latent dimensions that capture residual variability. During inference, the inverted network produces structural candidates from the spectrum code, effectively addressing the one-to-many challenge of spectrum-to-structure inference. On a filtered subset, the model demonstrates numerical invertibility on trained examples, surpassing random spectrum-code prediction and yielding coarse structural signals from validation spectra, proving the efficacy of invertible architectures in chemoinformatics.