Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Chemistry-Inspired Diffusion with Non-Differentiable Guidance

Authors: Yuchen Shen, Chenhao Zhang, Sijie Fu, Chenghui Zhou, Newell Washburn, Barnabás Póczos

ICLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments demonstrate that our method: (1) significantly reduces atomic forces, enhancing the validity of generated molecules when used for stability optimization; (2) is compatible with both explicit and implicit guidance in diffusion models, enabling joint optimization of molecular properties and stability; and (3) generalizes effectively to molecular optimization tasks beyond stability optimization. (Abstract)
Researcher Affiliation Academia Yuchen Shen , Chenhao Zhang , Sijie Fu , Chenghui Zhou, Newell Washburn , Barnab as P oczos Carnegie Mellon University Pittsburgh, PA 15213, USA EMAIL EMAIL
Pseudocode Yes Algorithm 1 Bilevel guided diffusion sampling with noisy neural guidance... Algorithm 2 CHEMGUIDE diffusion sampling with evolutionary algorithm... Algorithm 3 Bilevel guided diffusion sampling with clean neural guidance
Open Source Code Yes https://github.com/A-Chicharito-S/Chem Guide (Abstract) ... Our implementation is available at https://github.com/A-Chicharito-S/Chem Guide. (Section B)
Open Datasets Yes The models in our experiment are trained on the QM9 dataset (Ramakrishnan et al.) and the GEOM dataset (Axelrod & G omez-Bombarelli).
Dataset Splits No The paper mentions training on QM9 and GEOM datasets and sampling molecules for evaluation (e.g., "We sample 500 molecules from QM9"), but it does not specify explicit training/validation/test splits, percentages, or methodology used for partitioning the datasets.
Hardware Specification Yes Hardware & Time We use a 48 Gi B A6000 GPU with AMD EPYC 7513 32-Core Processors for our experiments. (Section B)
Software Dependencies No The paper mentions using specific models like EDM and Geo LDM, and a method like GFN2-x TB, and an external tool Gaussian16, but it does not provide specific version numbers for programming languages or libraries (e.g., Python 3.x, PyTorch 1.x) that were part of their implementation.
Experiment Setup Yes We choose s (Eq. 6) from {1, 10 1, 10 2, 10 3, 10 4} for all experiments, and additionally {2, 5, 10, 20, 25, 30, 40, 50} for the 6 properties. ... We add guidance to the last 400 of the 1000 diffusion steps (Han et al., 2024) (Section 4.1)