Fig. 2. Schematic depiction of the cG-SchNet architecture with inputs and outputs.

“ ⊕ " represents concatenation and “ ⊙ " represents the Hadamard product. Left: Atom-wise feature vectors representing an unfinished molecule are extracted with SchNet⁶⁷ and conditions are individually embedded and then concatenated to extract the conditional features vector. The exact embedding depends on the type of the condition (e.g., scalar or vector-valued). Middle: The distribution for the type of the next atom is predicted from the extracted feature vectors. Right: Based on the extracted feature vectors and the sampled type of the next atom, distributions for the pairwise distances between the next atom and every atom/token in the unfinished molecule are predicted. See Methods for details on the building blocks.