Next-generation AI-assisted drug design against cancer: large language models meet conventional in silico methods

Cancer remains a leading cause of death, with limited effective therapies. The AXL-GAS6 pathway promotes tumor growth, invasion, metastasis, and resistance to apoptosis. Large Language Models (LLMs) can predict drug-target interactions, generate novel molecular scaffolds, and optimize lead compounds. This study aims to design novel small molecules through a computational pipeline integrating commercial LLMs, molecular docking, molecular dynamics (MD), and ADMET evaluation. We combined DeepSeek LLM with conventional computational methods to design AXL inhibitors via three strategies: natural product-based, microbiome-derived, and FDA-approved drug-inspired scaffolds. Structured prompt engineering generated novel candidates, filtered for drug-likeness, synthetic feasibility, and docking score (Glide, Schrödinger). Top hits underwent 100 ns MD simulations and ADMET evaluation (SwissADME, ADMETLab3). AIC1 showed the highest binding affinity (- 10.079 kcal/mol), surpassing clinical-stage bemcentinib (- 8.234 kcal/mol). MD confirmed stable complexes (RMSD < 3 Å), with AIC1 and AIC4 forming extensive hydrogen bonds. ADMET profiling indicated favorable pharmacokinetics for all, with AIC2 exhibiting the lowest toxicity (hERG inhibition: 34.2%, hematotoxicity: 36.8%) and optimal drug-like properties. This work pioneers LLM-driven in silico design of AXL inhibitors, offering a scalable blueprint for accelerated anticancer drug development.

The online version contains supplementary material available at 10.1007/s40203-026-00582-y.