Fig. 1: Overall architecture of Uni-Mol+.

a In contrast to prior methods that directly predict QC properties from 1D/2D data, Uni-Mol+ uses a different approach. It first generates raw 3D conformation from 1D/2D data using cheap tools like RDKit, and then iteratively updates it towards the DFT equilibrium conformation. Finally, it predicts QC properties using the learned conformation. The abbreviation HOMO-LUMO gap represents the Highest Occupied Molecular Orbital— Lowest Unoccupied Molecular Orbital gap. b The Uni-Mol+ backbone consists of L blocks, each of which maintains two tracks of representations—atom and pair, initialized by atom features and 2D graph/3D conformation, respectively. These representations communicate with each other at every block. Based on this backbone model, Uni-Mol+ iteratively updates the raw conformation (i.e., 3D coordinates of atoms) towards the DFT equilibrium conformation for R iterations. The abbreviation FFN represents the Feed-Forward Neural network and QC property represents Quantum Chemical property. c A linear noisy interpolation between raw conformation and DFT conformation is used to generate a pseudo trajectory, effectively augmenting the input conformations. Uni-Mol+ uses a mixture of Bernoulli distribution and Uniform distribution to sample the noise interpolation weight q during training. The symbol q represents the interpolation weight between raw conformation and DFT conformation.