Research Projects

The central aim of this project is to develop robust pre-clinical drug candidates for cancer therapy for EphA2 (EPH receptor A2) and other receptor tyrosine kinases (RTK), leveraging a synergistic framework that combines state-of-the-art artificial intelligence/machine learning (AI/ML) protocols with physics-based simulations and biochemical experiments. Building on our proven expertise in large-scale non-equilibrium molecular dynamics (MD) simulations on 5000 protein-ligand conformations (PLCs), and normal MD simulations across 2000 kinases, the latter being a subset of ~16000 PLCs (with computationally predicted binding affinity), the aim is to develop predictive and generative AI/ML models, to generate new candidates and finally validate with physics-based simulations and biochemical assays validation. The EphA2 RTK is widely overexpressed in numerous solid tumors—evidently including non‑small cell lung cancer, HER2+ and triple‑negative breast cancer, glioblastoma, prostate, ovarian, and colorectal cancers—where its ligand‑independent (“non‑canonical”) signaling drives proliferation, invasion, metastasis, therapy resistance, and maintenance of cancer stemness. In healthy tissues, EphA2 binds its membrane‑bound ephrin‑A1 ligand to trigger receptor autophosphorylation, internalization, and degradation, in restraining cell migration. However, many tumors downregulate ephrin‑A1 or disrupt cell–cell contacts, resulting in surface accumulation of unliganded EphA2. Non‑canonical phosphorylation at sites such as S897 by Akt, RSK/ERK, or PKA then activates PI3K/Akt and Ras/MAPK pathways, promoting malignant hallmarks and correlating with poor patient prognosis. Despite this clear clinical importance, no selective, approved EphA2‑targeted therapy exists today. To bridge this gap, the current project will leverage our group’s unique physics‑based datasets and cutting‑edge AI/ML methods. We have already generated protein–ligand conformations (PLCs) with binding free energies from large‑scale equilibrium molecular dynamics (MD) and have simulated more than 2,000 kinase–ligand systems to predict binding affinities. In addition, non-equilibrium simulations on EphA2 and related RTKs will provide information on non-native protein ligand complexes ad well as drug residence‑time determinants often missed by docking alone. These high‑resolution datasets form the backbone for our predictive and generative AI/ML models. Our first aim is to compile and curate these proprietary conventional and enhanced MD data—augmented with publicly available binding‑free‑energy (BFE) datasets—into a standardized, high‑quality training corpus, involving both true examples and decoys. Next, we will train supervised ML algorithms to predict binding affinities and key drug‑like properties, using the curated BFE data to improve model robustness. Leveraging these predictors, we will develop generative models (reinforcement learning, normalizing flows, diffusion models) to propose novel small‑molecule scaffolds tailored to block EphA2’s ATP‑binding and allosteric sites, with multi‑objective optimization for isoform selectivity and ADMET properties. Generated candidates will undergo rigorous physics‑based validation—to refine and rank hits. A transfer‑learning loop will feed validation outcomes back into the generative pipeline, continuously improving molecular designs. Finally, top candidates will be synthesized for biochemical (kinase activity, ADP‑Glo) and cell‑based assays (viability, invasion, stemness), assessing inhibitory potency, target selectivity, and safety. By closing the AI/ML–physics–experimental loop, we aim to deliver 2–3 pre‑clinical EphA2 inhibitors with superior efficacy and resistance‑proof profiles. Beyond EphA2, this integrated framework is readily extensible to other resistance‑prone RTKs, offering a generalizable platform for precision oncology discovery.