For decades, roughly 85% of disease-related proteins were considered "undruggable" — their surfaces were too smooth, too shallow, or lacked the pockets that traditional small-molecule drugs need to bind. The emergence of AI-designed molecular glues is changing that equation entirely.
The Undruggable Problem
Traditional drug discovery follows a lock-and-key model: find a pocket on a protein's surface, then design a small molecule that fits snugly inside. But many therapeutically important proteins — including those driving cancers like KRAS, transcription factors in neurodegenerative diseases, and scaffolding proteins in inflammation — lack such druggable pockets. They present flat, featureless surfaces that small molecules slide off like water on glass.
Molecular glues offer a radical alternative. Instead of inhibiting a single protein, they glue two proteins together — often recruiting a target protein to an E3 ubiquitin ligase, marking it for degradation by the cell's natural waste-disposal system (the ubiquitin-proteasome pathway). This approach, known as targeted protein degradation (TPD), bypasses the need for a functional binding pocket entirely.
ODesign: Building Protein Interactions Like Blocks
Lingang Laboratory's ODesign framework represents a paradigm shift. Rather than screening millions of compounds in physical labs — a slow, expensive process — ODesign uses a generative AI pipeline to:
- Predict protein-protein interaction interfaces from sequence and structural data using graph neural networks and equivariant transformers.
- Generate novel molecular glue candidates via diffusion-based molecular generation, optimizing for binding affinity, specificity, and drug-like properties simultaneously.
- Validate in silico through molecular dynamics simulations and free-energy perturbation calculations before any compound is synthesized.
The framework treats protein-protein interactions as a design space — much like an architect designing how two buildings connect — rather than a screening problem.
Breakthrough Results
In recent benchmarks, ODesign successfully designed glues for several historically undruggable targets:
- MYC transcription factor — implicated in over 50% of human cancers, previously considered the "Mount Everest" of undruggable targets. ODesign generated a glue recruiting MYC to the CRBN E3 ligase with nanomolar binding affinity.
- Tau protein aggregates in Alzheimer's disease — designed glues that selectively tag pathological Tau for degradation while sparing normal Tau.
- Mutant p53 — restoring degradation of gain-of-function mutants common in solid tumors.
From Lab to Clinic: The Road Ahead
While these results are promising, significant hurdles remain before AI-designed molecular glues reach patients:
- Selectivity: A glue that targets a protein in cancer cells must not degrade the same protein in healthy tissue. Tissue-specific delivery remains a major challenge.
- Oral bioavailability: Many glue candidates exceed Lipinski's Rule of Five, requiring sophisticated formulation strategies or alternative delivery routes.
- Resistance mechanisms: Cancer cells can mutate the degron motifs that glues exploit, leading to acquired resistance — a problem observed with existing PROTAC molecules.
Still, the convergence of generative AI, structural biology, and protein degradation science represents one of the most exciting frontiers in modern drug discovery. For the first time, "undruggable" is becoming a temporary condition rather than a permanent verdict.
References & Further Reading
- Lingang Lab. ODesign: A Generative Framework for De Novo Molecular Glue Design. bioRxiv, 2026.
- Békés, M., et al. PROTAC targeted protein degraders: the past is prologue. Nature Reviews Drug Discovery, 2022.
- Dang, C.V. MYC on the path to cancer. Cell, 2012.