Traditional drug discovery is costly, slow, and has low FDA approval rates. Deep learning models such as GAN, VAE, and reinforcement learning (RL) have emerged as a solution, using SMILES or graphs to create drugs tailored to target molecules/ligands, or employing heuristics like QED and Lipinski’s rule. Additionally, molecule vector representations enable various downstreamed transfer learning (TL). However, existing works face the following limitations: generate small molecules, no implicit way of using information from scientific literature, use on-policy RL. This paper extends the Soft Mixture of Experts architecture introduced for images by DeepMind for network graphs, namely Soft Graph Mixture of Experts (SGMoE). Using multi-headed spectral graph transformers, which are stacked in an adversarial VAE for generating latent vectors. It has been benchmarked on GuacaMole, DOCKSTRING and others for de novo drug design, and MoleculeNet and DOCSTIRNG tasks for molecular property prediction. Finally the vector representation of the molecule combined with embeddings from Galactica AI and ChemLLM are passed to a Deep Q Learning model for generating molecules. GraphChem can drastically reduce costs and time for drug development, enhancing accessibility to life-saving treatments worldwide and advancing global health equity and outcomes.