GLI-NT: Cascaded Transfer Learning for Virtual Screening of Novel GLI1 Zinc Finger Inhibitors in Hedgehog-Driven Cancers

Overview

Medulloblastoma, the most common malignant pediatric brain tumor, is driven in ~30% of cases by aberrant Hedgehog signaling through GLI transcription factors. FDA-approved upstream inhibitors fail due to rapid resistance, and no direct GLI1 inhibitor exists clinically. Targeting GLI1's zinc finger domain would circumvent resistance at the pathway's terminal node, but this surface is shallow and solvent-exposed, rendering conventional screening ineffective. GLI-NT solves this through cascaded transfer learning. The model captures target biology through frozen ESM-2 protein embeddings (650M parameters), drug chemistry through ChemBERTa embeddings, and structural specificity through 2,048-bit Morgan fingerprints, all integrated by a trainable prediction head. Training cascades in three stages: pretraining on 150,935 BindingDB pairs, zinc finger adaptation on 1,573 pairs, and GLI-specific fine-tuning via leave-one-compound-out cross-validation across five seeds. Despite training on only 28 known inhibitors, the model correctly identified up to 22, establishing that cascaded transfer learning enables drug discovery in extreme data scarcity. A 2.0% false positive rate protects downstream researchers from years spent pursuing dead-end compounds. These predictions held under independent validation: screening over one million compounds uncovered 500 candidates, 84.6% representing scaffolds absent from existing literature. Docking energies reached −8.70 kcal/mol, and 110 ns ternary molecular dynamics confirmed sustained target engagement at the GLI1 zinc finger site. Because aberrant GLI signaling also drives basal cell carcinoma and pancreatic cancer, and the architecture generalizes to any transcription factor, GLI-NT establishes a new paradigm for AI-guided cancer drug discovery.