Research map / Image generation / Autoregressive and latent generation
Domain-Adapted Diffusion Model for PROTAC Linker Design Through the Lens of Density Ratio in Chemical Space
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DAD-PROTAC, a diffusion model for PROTAC linker design pre-trained on ZINC fragments and adapted to PROTAC-DB via a density-ratio-based score correction term learned by a lightweight network.
Problem
Diffusion models for PROTAC linker design are pre-trained on small molecules, creating a chemical-space distribution mismatch with PROTACs, and direct fine-tuning on the small PROTAC-DB overfits.
Approach
Decomposes the target PROTAC score estimator into a pre-trained small-molecule score plus a score correction term obtained via density ratio estimation (binary classifier), learned as a lightweight network instead of full model retraining.
Key results
- DAD-PROTAC reaches 94.8% validity and 45.7% recovery on PROTAC-DB, versus 82.9% validity and 32.8% recovery for LinkerNet-Fine-tuning.
- Ablation: full fine-tuning gives 82.5% validity and direct score correction gives 81.9%, both below the full DAD-PROTAC model.
- Fine-tuning cost of baselines ranges from 1.35x to 2.8x that of DAD-PROTAC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use density-ratio score correction instead of pre-training LinkerNet on ZINC and fine-tuning it on PROTAC-DB | PROTAC linker generation · PROTAC-DB test | Valid (%): 82.9 ± 0.3 -> 94.8 ± 0.4 | Helped |
| Replace density-ratio score correction with standard full fine-tuning of all EGNN parameters (ablation) | DAD-PROTAC · PROTAC-DB test | Valid (%): 82.5 ± 0.4 -> 94.8 ± 0.4 | Helped |
| Replace density-ratio score correction with standard partial fine-tuning of the EGNN decoder (ablation) | DAD-PROTAC · PROTAC-DB test | Valid (%): 83.1 ± 0.6 -> 94.8 ± 0.4 | Helped |
| Use density-ratio estimation rather than direct score-correction approximation (ablation) | DAD-PROTAC · PROTAC-DB test | Valid (%): 81.9 ± 0.8 -> 94.8 ± 0.4 | Helped |
| Estimate the density ratio from noise-perturbed samples instead of fixed clean samples (ablation) | DAD-PROTAC · PROTAC-DB test | Valid (%): 90.0 ± 4.7 -> 94.8 ± 0.4 | Helped |
Surprising findings
- Direct score correction without density ratio estimation performs worst among ablations, with higher Emin (203.8) than full fine-tuning.
Limitations
- Assumes the number of linker atoms is pre-specified.
- Pre-training does not consider fragment rotation or protein context.
- Direct fine-tuning on the small PROTAC-DB dataset overfits.
Experimental setup
Datasets: ZINC PROTAC-DB
Benchmarks: PROTAC-DB
Metrics: validity uniqueness novelty recovery rate QED SA Emin RMSD
Baselines: 3DLinker DiffLinker LinkerNet full fine-tuning partial fine-tuning
Builds on: DDPM multinomial diffusion denoising concrete score matching EGNN DiffLinker
Closest papers by meaning
- Chain-of-Generation: Progressive Latent Diffusion for Text-Guided Molecular Design (NeurIPS 2026 · similarity 0.85)
- Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · similarity 0.85)
- DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization (AAAI 2024 · similarity 0.85)
- DMol: A Highly Efficient and Chemical Motif-Preserving Molecule Generation Platform (NeurIPS 2025 · similarity 0.85)
- Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings (NeurIPS 2025 · similarity 0.84)
- Budgeted Quotient-Residual Guidance for Frozen Pocket-Conditioned Molecular Diffusion (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.