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..
TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles
Authors: Yaoyao Xu, Di Wang, Zihan Zhou, Tianshu Yu, Mingchen Chen
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our method achieves significant improvements over existing approaches. Our work makes several key contributions: ... Performance and Metrics. Our method achieves state-of-the-art performance in both structural accuracy and computational efficiency in various metrics, outperforming existing methods in matching MD ground truth while requiring fewer computational resources. Extensive Analysis. We demonstrate TEMPO s ability to capture biologically meaningful protein motions through comprehensive case studies and analyses. |
| Researcher Affiliation | Academia | School of Data Science, The Chinese University of Hong Kong, Shenzhen Changping Laboratory, Beijing EMAIL, EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1: Autoregressive Training of Protein Dynamics Model with Multiple Timesteps |
| Open Source Code | Yes | The code is available in the supplemental material. |
| Open Datasets | Yes | We conduct experiments on two comprehensive molecular dynamics datasets: md CATH [29] and ATLAS [42]. |
| Dataset Splits | Yes | For md CATH, we randomly sampled 1,000 proteins... We randomly selected 50 proteins for validation and 64 proteins for testing, ensuring no overlap with the training set. For ATLAS, we follow the data split and processing protocol established by MDGen [21]. |
| Hardware Specification | Yes | In contrast, our multi-scale decomposition enables training on a single NVIDIA A100 GPU, with both slow-scale and fast-scale models operating within memory constraints. |
| Software Dependencies | No | The paper does not explicitly list specific version numbers for key software components or libraries, beyond mentioning a general 'environment needed to run' for code reproducibility. |
| Experiment Setup | Yes | Our multi-scale modeling approach captures protein dynamics at two temporal resolutions. The low-resolution model generates trajectories at 20ns intervals, characterizing major structural transitions, while the high-resolution model operates at 1ns resolution to capture local fluctuations... During training, both scale-models simulate the forward process of protein dynamics SDE through autoregressive sampling with noise scales uniformly sampled from [0.01, 0.05]. At inference time, while the low-resolution model maintains similar noise levels, we increase the noise scale to 5.0 for the high-resolution model. |