DemoDiff

DemoDiff is a molecular design foundation model that adapts to new tasks the way large language models do in-context: from a handful of demonstrations. In molecular design, a task is usually defined by an optimization target—binding to a protein, a desired material property, a synthetic constraint—and conventional generators must be trained or fine-tuned separately for each. DemoDiff instead conditions a diffusion model on a small set of example molecule-score pairs that define the task, then generates new molecules that fit that context, avoiding per-task retraining.

The core method is a demonstration-conditioned diffusion model built on a denoising graph transformer. To make pretraining over millions of tasks tractable, the authors introduce a molecular tokenizer based on Node Pair Encoding—a graph analogue of byte-pair encoding that merges recurring substructures into motif-level tokens—reducing representation length by roughly 5.5x relative to atom-level graphs. This compression is what lets a compact 0.7-billion-parameter model act as a general in-context designer.

DemoDiff was developed by researchers at the University of Notre Dame, IBM Research, and MIT, and released as a preprint in October 2025 with open code and weights.

#Key Features

• In-context molecular design: Tasks are specified by example molecule-property pairs rather than text prompts or fine-tuning, so the model adapts to a new objective from a few demonstrations at inference time.
• Graph diffusion transformer: A denoising diffusion process over molecular graphs, driven by a transformer backbone, generates chemically valid structures conditioned on the demonstration context.
• Node Pair Encoding tokenizer: A motif-level graph tokenizer (3,000 base tokens plus 300 ring tokens) shrinks molecular representations by about 5.5x, enabling efficient pretraining across millions of context tasks.
• Cross-domain coverage: Pretraining spans both drug-like molecules and materials, giving one model competence over chemically distinct design spaces.
• Open release: Code and the 0.7B checkpoint are publicly available on GitHub and Hugging Face.

#Technical Details

DemoDiff is a discrete graph diffusion transformer with roughly 700 million parameters: 24 transformer layers, 16 attention heads, a hidden size of 1,280, and a 500-step diffusion schedule with cosine noise. Molecules are encoded with a Node Pair Encoding vocabulary of 3,000 tokens plus 300 ring tokens, handling up to 150 tokens per molecule. The model is pretrained on millions of curated context tasks drawn from drug and materials datasets. Evaluated on 33 design tasks spanning six categories, DemoDiff achieves an average rank of 3.63, compared with 5.25 to 10.20 for domain-specific baselines, and matches or exceeds molecular language models that are 100 to 1,000 times larger. The code is released under the MIT license.

#Applications

DemoDiff is aimed at researchers in drug discovery and materials science who need to generate candidate molecules for tasks that lack large labeled datasets. Because a new objective is specified with only a few example molecule-score pairs, a chemist can steer generation toward a target—docking to a receptor, tuning a polymer property, satisfying a structural constraint—without assembling a training set or training a bespoke model. This lowers the barrier to applying generative chemistry to the many niche objectives that arise in real projects.

#Impact

DemoDiff transfers the in-context learning paradigm that reshaped language modeling into molecular generation, showing that a compact 0.7B model with demonstration conditioning can rival or beat both specialized generators and far larger molecular language models across a broad task suite. Its motif-level graph tokenizer is a reusable contribution for scaling graph diffusion. As a preprint, its benchmark results are computational; its practical influence will grow as designs are validated and the open checkpoint is adopted.