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..
Fractional Diffusion Bridge Models
Authors: Gabriel Nobis, Maximilian Springenberg, Arina Belova, Rembert Daems, Christoph Knochenhauer, Manfred Opper, Tolga Birdal, Wojciech Samek
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate FDBM on both tasks: predicting future protein conformations from aligned data, and unpaired image translation. In both settings, FDBM achieves superior performance compared to the Brownian baselines, yielding lower root mean squared deviation (RMSD) of Cα atomic positions in protein structure prediction and lower Fréchet Inception Distance (FID) in unpaired image translation. 5 Experiments We evaluate the performance of FDBM on both paired and unpaired data translation tasks; see Section I for a detailed description of the evaluation metrics. In the paired setting, we first show in a proof-of-concept on synthetic data that the alignment of training data is preserved and then predict conformational changes in proteins. In the unpaired setting, we consider image-to-image translation across visually distinct domains. Detailed architectural specifications, compute resources, training protocols, and dataset descriptions are provided in Sections F and G, additional experiments are reported in Section K, and an additional use case on cell differentiation is presented in Section J. |
| Researcher Affiliation | Collaboration | Gabriel Nobis Fraunhofer HHI Maximilian Springenberg Fraunhofer HHI Arina Belova Fraunhofer HHI Rembert Daems Ghent University imec Flanders Make MIRO Christoph Knochenhauer Technical University of Munich Manfred Opper Technical University of Berlin University of Potsdam University of Birmingham Tolga Birdal Imperial College London Wojciech Samek Fraunhofer HHI Technical University of Berlin |
| Pseudocode | No | The paper describes methods and mathematical frameworks but does not include any explicitly labeled "Pseudocode" or "Algorithm" blocks or figures. |
| Open Source Code | Yes | Our contributions are: We accompany our work with several publicly available implementations to facilitate the adoption of our framework in both paired and unpaired settings, as well as a stand-alone reimplementation of the method proposed by Bortoli et al. [27]. 1https://github.com/GabrielNobis/FDBM_paired 2https://github.com/mspringe/FDBM_unpaired 3https://github.com/mspringe/Schroedinger-Bridge-Flow |
| Open Datasets | Yes | We evaluate FDBM on both tasks: predicting future protein conformations from aligned data, and unpaired image translation. In both settings, FDBM achieves superior performance compared to the Brownian baselines, yielding lower root mean squared deviation (RMSD) of Cα atomic positions in protein structure prediction and lower Fréchet Inception Distance (FID) in unpaired image translation. Following the training and evaluation setup of Somnath et al. [31], we use their curated subset of the D3PM dataset [49] to evaluate the ability of FDBM to predict 3D ligand-bound (holo) structures from given 3D ligand-free (apo) unbound protein conformations. Unpaired data translation is evaluated for the cat and wild subsets of the AFHQ dataset [51]. We evaluate FDBM on the Moons and T-Shape datasets introduced by Somnath et al. [31], and depicted in Figure 7 |
| Dataset Splits | Yes | This results in a cleaned dataset of 1, 591 pairs, which is split into training, validation, and test sets of 1, 291/150/150 examples, respectively. |
| Hardware Specification | Yes | Each trial of 300 training epochs for the protein conformational change task was completed within 24 hours on a single NVIDIA A100 GPU (40 GB VRAM). All pretrainings of 100K steps for the AFHQ-32 and AFHQ-256 datasets were completed in 16 hours (A100) for the Di T-B/2 variant and 54 hours (A100) for the Di T-L/2 variant. All computations of this section were performed on an NVIDIA A100 GPU (40 GB VRAM). |
| Software Dependencies | No | Table 3: Hyperparameters for experiments with Diffusion Transformers. Model Optimizer Learning Rate EMA Rate Linear Warmup Cosine Decay Online Finetuning Euler Maruyama Steps Parameters Di T-B/2 lion [103] 0.0001 0.999 10K 90K 4K 200 130M Di T-L/2 lion [103] 0.0001 0.999 10K 90K 4K 200 458M We use the Adam W [98] optimizer with an initial learning rate of 0.001 and a training batch size of 2. The paper mentions specific models and optimizers but does not provide version numbers for general software dependencies like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | Table 3: Hyperparameters for experiments with Diffusion Transformers. Model Optimizer Learning Rate EMA Rate Linear Warmup Cosine Decay Online Finetuning Euler Maruyama Steps Parameters Di T-B/2 lion [103] 0.0001 0.999 10K 90K 4K 200 130M Di T-L/2 lion [103] 0.0001 0.999 10K 90K 4K 200 458M Conformational changes in proteins. The results reported in Table 1 were obtained by averaging over 5 training trials, each run for 300 epochs, and performing one sampling trial per trained model, generating a single path over 100 time steps. The remaining training set-up closely follows Somnath et al. [31]. We use the Adam W [98] optimizer with an initial learning rate of 0.001 and a training batch size of 2. |