SmilesGEN AI model generates drug-like molecules using multi-objective reinforcement learning

A new generative model called SmilesGEN uses a variational autoencoder architecture to create drug-like molecules with potential therapeutic effects. The model addresses limitations in previous methods that rely on expression profiles but overlook how molecules perturb cellular contexts. SmilesGEN integrates a pre-trained drug VAE named SmilesNet with an expression profile VAE called ProfileNet. These components jointly model the interplay between drug perturbations and transcriptional responses within a common latent space. ProfileNet is designed to reconstruct pre-treatment expression profiles by eliminating drug-induced perturbations in the latent space. Meanwhile, SmilesNet uses desired expression profiles to generate drug-like molecules. The approach aims to bridge the gap between phenotype and target for molecular generation. This research focuses on de novo generation of molecules capable of inducing desirable phenotypic changes. The work is documented in the arXiv preprint 2509.21010v2, which was announced as a replace-cross type. The model's development reflects increasing attention toward generating molecules with specific phenotypic impacts.