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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Ambient Proteins - Training Diffusion Models on Noisy Structures
Authors: Giannis Daras, Jeffrey Ouyang-Zhang, Krithika Ravishankar, Constantinos Daskalakis, Adam Klivans, Daniel Diaz
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, Ambient Protein Diffusion yields major improvements: on proteins with 700 residues, diversity increases from 45% to 86% from the previous state-of-the-art, and designability improves from 68% to 86%. |
| Researcher Affiliation | Academia | Giannis Daras CSAIL, MIT EMAIL Jeffrey Ouyang-Zhang Computer Science, UT Austin EMAIL Krithika Ravishankar Computer Science, UT Austin EMAIL Costis Daskalakis CSAIL, MIT EMAIL Adam Klivans Computer Science, UT Austin EMAIL Daniel J. Diaz Computer Science, UT Austin EMAIL |
| Pseudocode | Yes | Algorithm 1 Ambient Protein Diffusion: Training Algorithm. |
| Open Source Code | Yes | All of our code, models and datasets are available under the following repository: https://github.com/jozhang97/ambient-proteins. |
| Open Datasets | Yes | The Alpha Fold Database (AFDB), which contains over 214M predicted structures from Uni Prot KB sequences [13, 46]. In parallel, high-throughput tools for sequence and structure comparison, such as MMSeqs2 and Fold Seek, have facilitated the curation of large, diverse training datasets from AFDB [5]. Among them, the 2.3M AFDB cluster dataset, has already been shown to improve the capabilities of generative models for protein structure design [34, 23]. |
| Dataset Splits | No | The paper describes training on subsets of the AFDB cluster dataset based on protein length and pLDDT scores across three stages, for example, training on 196,000 proteins in the first stage. However, it does not specify a standard train/validation/test split for the original dataset in the traditional sense, but rather how training data was composed for different training stages and evaluation was done on generated samples. |
| Hardware Specification | Yes | Training is performed on 48 GH200 GPUs and runs in 18, 48, and 48 hours for each stage respectively. |
| Software Dependencies | No | The paper mentions several tools like Protein MPNN [19], ESMFold [35], MMSeqs2 [26], Fold Seek [45], and biotite, but it does not provide specific version numbers for any of the software, libraries, or frameworks used in the implementation (e.g., Python, PyTorch, etc.). |
| Experiment Setup | Yes | Table 16: Hyperparameters of the diffusion protein model. Dashes (-) indicate that the value is the same as the previous column. The Ambient walls correspond to the assigned diffusion times based on the protein s p LDDT (times are from 1 to 1000). Proteins with p LDDT > 90 are used everywhere. Proteins with p LDDT > 80 are used for times in [600, 1000] and proteins with p LDDT > 70 are used for times in [900, 1000]. |