Model training method and device

The invention provides a model training method and device. Comprising the steps of obtaining a model training sample which comprises a protein pocket; inputting the model training sample into a molecular generation model in a to-be-trained protein pocket, wherein the molecular generation model in the to-be-trained protein pocket comprises N network layers which are circularly connected; calling N circularly connected network layers to perform atom-by-atom prediction processing on the protein pockets to obtain predicted atom types, predicted central atoms, predicted atom positions and predicted atom bonding relationships corresponding to the protein pockets; on the basis of the predicted atom type, the predicted central atom, the predicted atom position and the predicted atom bonding relationship, calculating to obtain a loss value of the molecular generation model in the to-be-trained protein pocket; and under the condition that the loss value is within a preset range, taking the trained to-be-trained molecul