AI-Assisted Designing of Novel Anticancer Compounds

Anti-cancer drug design has historically been a challenging and resource-intensive process. Artificial intelligence (AI) and deep learning now enable rapid generation and prioritization of novel drug-like compounds. This study presents a SMILES-to-SELFIES Recurrent Neural Network (RNN) model that was trained on existing anticancer compounds to generate structurally diverse candidate molecules followed by ADMET filtering, molecular docking, synthetic accessibility scoring, and QSAR prediction. A total of 100 new molecules were generated, of which top candidates showed strong predicted potency and favorable pharmacokinetic properties. Particularly, Molecule-43 demonstrated the highest predicted inhibitory potential and drug-likeness score. Several candidates exhibited strong predicted binding against five cancer-associated target proteins (EGFR, P53, AKT1, BCL-2, and IKKβ), with Molecule 43 showing consistent multi-target affinity. QSAR prediction using Random Forest models further supported the inhibitory potential of prioritized compounds. These results demonstrate a promising computational strategy for accelerating anticancer drug discovery. This work is a purely computational study, and all results represent predictive hypotheses requiring experimental validation.