Fig. 1: Overview of Lingo3DMol model development.
From: Generation of 3D molecules in pockets via a language model
a, Lingo3DMol architecture. Three separate models are included: the pretraining model, the fine-tuning model and the NCI/anchor prediction model. These models share the same architecture with slightly different inputs and outputs. b, Illustration of FSMILES construction. The same colour corresponds to the same fragment. c, Illustration of pretraining perturbation strategy. Step 1, original molecular state; step 2, removal of edge information during pretraining; step 3, perturbation of the molecular structure by randomly deleting 25% of the atoms; step 4, perturbation of the coordinates using a uniform distribution within the range (−0.5 Å, 0.5 Å) and step 5, perturbation of 25% of the carbon element types. These perturbations are applied in no particular order, and the pretraining task aims to restore the molecular structure from step 5 to step 1.
