Transformer-Decoder GPT Models for Generating Virtual\nScreening Libraries of HMG-Coenzyme A Reductase Inhibitors: Effects\nof Temperature, Prompt Length, and Transfer-Learning Strategies

Attention-based decoder\nmodels were used to generate libraries\nof novel inhibitors for the HMG-Coenzyme A reductase (HMGCR) enzyme.\nThese deep neural network models were pretrained on previously synthesized\ndrug-like molecules from the ZINC15 database to learn the syntax of\nSMILES strings and then fine-tuned with a set of ∼1000 molecules\nthat inhibit HMGCR. The number of layers used for pretraining and\nfine-tuning was varied to find the optimal balance for robust library\ngeneration. Virtual screening libraries were also generated with different\ntemperatures and numbers of input tokens (prompt length) to find the\nmost desirable molecular properties. The resulting libraries were\nscreened against several criteria, including IC50 values predicted\nby a dense neural network (DNN) trained on experimental HMGCR IC50\nvalues, docking scores from AutoDock Vina (via Dockstring), a calculated\nquantitative estimate of druglikeness, and Tanimoto similarity to\nknown HMGCR inhibitors. It was found that 50/50 or 25/75% pretrained/fine-tuned\nmodels with a nonzero temperature and shorter prompt lengths produced\nthe most robust libraries, and the DNN-predicted IC50 values had good\ncorrelation with docking scores and statin similarity. 42% of generated\nmolecules were classified as statin-like by k-means clustering, with\nthe rosuvastatin-like group having the lowest IC50 values and lowest\ndocking scores.