Loading…
PROTAC degrader generation pipeline that screens target-binding fragments, then builds molecules under structure and physicochemical constraints.
No providers recorded yet. Browse all providers
PROTACs degrade a protein rather than inhibit it, by tethering a target-binding warhead to an E3 ligase ligand and letting the ubiquitin-proteasome system do the rest. That mechanism opens targets with no druggable pocket, but it also breaks most computer-aided design tools, which depend on modeling the ternary complex of target, degrader, and ligase — a structure that is rarely resolved and poorly predicted.
LM-PROTAC, from Nantong University and Guangxi Medical University, sidesteps ternary complex modeling entirely by working at the level of molecular fragments and their affinities. Proteins and molecules are segmented into fragments with the VOLT optimal-transport vocabulary algorithm, an affinity model scores which molecular fragments bind which protein fragments of the target, and a generative transformer assembles degraders around the winning fragments under simultaneous structural and physicochemical constraints.
This entry documents the pipeline rather than the generative model alone. The paper names the generative component DCT — a dual-constraint transformer — but DCT is steered by fragments the affinity stage selects and filtered by a downstream property model, and its published results are all obtained inside that arrangement. Two other catalog entries approach targeted degradation from different angles: AiPROTAC predicts degradation activity for existing chimeras rather than generating them, and SynPROTAC generates degraders from reaction templates to guarantee synthesizability.
The generative model is built on the conditional transformer (C-Transformer) formulation and trained on 250,000 drug-like small molecules drawn from ChEMBL and ZINC. Fragment vocabularies came from splitting ZINC SMILES into 4,577,207 candidate fragments, narrowed to 870,834 after Chembridge-style filtering; protein fragments came from BindingDB. Benchmarked on MOSES against CharRNN, AAE, VAE, JT-VAE, Organ, and MolGPT, DCT reported validity 0.998, uniqueness 0.999, novelty 0.997, an FCD of 0.083, and KL divergence 0.921.
For the case study, Wnt3a — an early driver in liver cancer — was segmented and screened, yielding nine high-affinity fragment pairs with interaction scores from 0.861 to 0.972. A PROTAC-DB v2-trained generator produced 10,000 molecules, 100 per pair. RDKit validation kept 9,802; MDAM property screening cut that to 798, then 737 on synthetic accessibility and 372 on QED and Lipinski criteria; retrosynthetic analysis selected 12, of which 3 went to 500 ns molecular dynamics against the Wnt3a-CRBN complex. Two were successfully synthesized and confirmed by NMR. In HepG2 hepatocellular carcinoma cells, western blot showed candidate compound 1 reducing Wnt3a levels dose-dependently from 0.16 µM upward, while candidate compound 2 produced a weaker and less consistent reduction beginning at 5 µM.
The pipeline suits medicinal chemistry teams pursuing targeted degradation of proteins that resist conventional inhibition, particularly secreted and scaffolding proteins like Wnt3a where there is no well-defined pocket for a classical warhead. Because property and synthesizability constraints are applied before candidates leave the computer, the output is a short synthesis list rather than a large virtual library, which matters for groups whose bottleneck is bench capacity rather than compute.
LM-PROTAC's contribution is less any single component than the demonstration that a fragment- language view of proteins and ligands can carry a degrader program from target sequence to a western blot without ever modeling a ternary complex. The end-to-end validation — generation, simulation, synthesis, and cellular degradation — is more than most generative PROTAC papers attempt. The caveats are real: only one of two synthesized compounds showed convincing degradation, the authors note that molecular diversity is limited by reliance on similar scaffolds and that CPI prediction accuracy still lets weak molecules through, and the wet-lab evidence rests on a single target in a single cell line. No code, weights, or hosted version has been released, and the work remains a preprint.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.