A dataset of small protein conformational ensembles from all-atom molecular dynamics simulations

Abstract

Small proteins, with more pronounced conformational flexibility, are crucial in biological processes and their function is linked to their high flexibility. While several databases are widely used to characterize protein dynamics, a significant data gap remains for small proteins comprising 5–100 amino acids. To address this, we present DynoDB, a dataset of small protein conformational ensembles from all-atom molecular dynamics simulations. In DynoDB, a total of 8,385 small proteins were subjected to all-atom molecular dynamics simulations of 100 ns each. We generated about 9.1 TB of trajectory data, covering diverse functional types of biomolecules, along with structural and functional annotations, as well as dynamic analysis results. This dataset systematically captures the conformational dynamics of the small proteins, serving as a dedicated data resource for the field of peptide design, vaccine development, antimicrobial screening, and AI modeling.

Background & Summary

Proteins are inherently dynamic under physiological conditions¹. A variety of important biological processes, such as ligand recognition, enzymatic catalysis, and membrane translocation, rely on such dynamic behaviors². Advances in molecular dynamics (MD) simulations have enabled researchers to investigate the crucial link between dynamics and function of biological macromolecules in realistic environments³.

Small proteins with a length of 5–100 amino acids (aa), which often exhibit high-frequency conformational transitions, are involved in diverse functional systems, such as antimicrobial and immune defense peptides, membrane-associated regulators, and functional peptides^4,5. Understanding the dynamics of these molecules is thus essential for illuminating the fundamental structure-dynamics-function relationship and providing data-driven support for key practical applications: peptide design, vaccine development, antimicrobial discovery, and AI modeling of biomolecular systems⁶.

To further enhance the research of protein dynamics, efforts have been made to construct general MD datasets. Notable general databases include ATLAS⁷, a database of standardized 100-ns all-atom MD simulations on 1,522 protein chains with 3 replicates, and mdCATH⁸, a database of all-atom MD simulations on 5,398 representative protein domains. In addition, there are several MD databases on specific protein classes, including G-protein-coupled receptors (GPCRs)⁹, SCOV2-MD¹⁰, MemProtMD¹¹, and DynamicPDB¹². However, these existing databases have limited coverage of small proteins (Table 1).

Table 1.

Overview of public molecular dynamics (MD) simulation.

Database	Research Scope / Target	Number of Proteins / Systems	Simulation Time & Replicates
ATLAS	General protein dynamics	1,390 protein chains (1,068 proteins), 1,149 non-redundant ECOD X-class domains	100 ns × 3
mdCATH	Protein domains	5,398 domains, 134,950 trajectories	Five temperatures × 5 replicas; average 464 ns
GPCRmd	G protein-coupled receptors	588 GPCR systems	500 ns × 3
SCoV2-MD	SARS-CoV-2 proteome	>250 systems (covering all SARS-CoV-2 proteins)	Variable
MemProtMD	Membrane proteins	>8,000 membrane proteins	1 μs (coarse-grained)
Dynamic PDB	General protein dynamics	~12,600 proteins, each with 1 μs all-atom MD	1 μs

To address the gap in small protein dynamics data, we introduce a comprehensive dataset of small protein conformational ensembles, available at Hugging Face. This dataset, hereafter referred to as DynoDB, consists of 8,385 conformational ensembles for small proteins. These protein structures were collected from the Protein Data Bank (PDB)¹³ and underwent quality filtering, structure repair, and 100-ns MD simulations. Functional annotations were retrieved from the UniProt database¹⁴, with an automated keyword-based classification system identifying 17 major functional categories. To validate the quality of the repaired structure, we calculated the structural evaluation metrics. Furthermore, we performed MD simulations on 130 proteins from the ATLAS database⁷ to verify the reproducibility of the simulation process. This resource provides a valuable dataset for further investigations into protein dynamics, binder design, and peptide structure prediction.

Methods

Data collection and annotation

Data collection and filtering

A total of 9,247 potential protein structures (5–100 AAs) were identified in the Protein Data Bank (PDB)¹³ as of May 13, 2025. Following the exclusion of 27 entries due to retrieval failures, 9,220 structures were processed through our quality control pipeline (Fig. 1a). First, we removed protein structures with low resolution (resolution > 3 Å)¹⁵ or lacking resolution information (Table 2), except for Nuclear Magnetic Resonance (NMR)-resolved files. For NMR-resolved files (6,766 of 9,220 retrieved structures), although the structures lack resolutions, they generally exhibit good agreement with experimentally observed conformations¹⁶. A total of 9,159 structures were retained. Next, we excluded incomplete proteins missing more than 10 consecutive residues (268 structures). For the remaining 8,891 protein structures, we performed structural repair.

Fig. 1.

Flowchart for constructing the small protein dynamics dataset. (a) Data filtering: PDB entries are filtered by date, length (5–100 aa), resolution (≤3 Å for non-NMR), and completeness (removal of >10 missing residues). (b) Data cleaning and repair: water/ligands are removed, missing atoms are filled using PDBFixer v1.11, missing residues are filled using PDBFixer v1.11 or Boltz-1, and modified residues are repaired using Rosetta. (c) Molecular dynamics simulation: the systems are prepared using GROMACS (CHARMM36m, TIP3P, ions), followed by energy minimization, NVT/NPT equilibration, and MD.

Table 2.

Summary of structure filtering by experimental method.

Experimental method	Total structures	Removed (low resolution >3 Å or missing)	Retained
Nuclear Magnetic Resonance (NMR)	6,761 (73.32%)	0	6,761
X-ray crystallography	2,330 (25.27%)	30	2,300
Electron Microscopy (EM)	100 (1.08%)	21	79
Neutron diffraction	12 (0.13%)	0	12
Multiple methods	13 (0.14%)	6	7
Other	4 (0.04%)	4	0
Total	9,220	61	9,159

Structure repair

To address the structural gaps in the original PDB files, we used the internal records of the PDB files as the reference for reconstruction. Specifically, REMARK 465 records provided information on missing residues, REMARK 470 records indicated missing atoms, and SEQRES entries defined the theoretical full-length amino acid sequence. These records served as the ground truth to guide the subsequent repair procedures. First, water and ligand molecules were removed. Then, PDBFixer v1.11 was applied to identify and add missing atoms. The findMissingAtoms function was used to detect incomplete residues by comparing the ATOM records against the standard atom definitions in the CHARMM36m force field topology. The identified heavy atoms and hydrogens were then automatically rebuilt using the addMissingAtoms function. Next, missing residues were addressed based on the sequence defined in the SEQRES records. If the number of consecutively missing residues was less than or equal to five, they were reconstructed using PDBFixer’s findMissingResidues and addMissingResidues functions. For segments exceeding five residues, the entire protein structure was re-predicted using the Boltz-1 method based on the full-length SEQRES sequence. Finally, modified residues were processed according to their positions to ensure compatibility with downstream MD simulations. Modifications at the N- or C-terminus were removed directly. For internal modified residues, we utilized the Rosetta software suite (v3.14) to replace them with standard amino acids, followed by a FastRelax protocol using the Cartesian-space ref2015_cart score function to alleviate potential steric clashes. As a result, 8,385 repaired and processed structures were obtained as initial conformations for molecular dynamics simulations. For NMR-resolved data, we used the first model as the initial conformation for subsequent MD simulation.

Data annotation

For structural annotation, we calculated and summarized the information including secondary structure composition, numbers of missing residues and atoms, and the number of non-standard amino acids. As for functional annotation, we retrieved protein types and functional information by extracting key descriptions from the Uniprot resource, specifically the function section, UniProt keywords section, and Gene Ontology (GO) section¹⁴. As a result, a total of seventeen functional categories were determined by the custom script for this dataset.

MD simulation pipeline

We constructed an all-atom MD simulation pipeline, referencing the protocol described in the ATLAS database work (Fig. 1c). First, the simulation systems were constructed using the CHARMM36m force field (July 2020 release)¹⁷. Then, each protein was placed in a periodic cubic box and solvated with TIP3P^18,19 water molecules. Systems were neutralized with Na⁺/Cl⁻ ions at a concentration of 150 mM to mimic physiological salt conditions. Next, energy minimization was performed using the steepest descent algorithm for 5,000 steps to optimize system geometry. Equilibration was conducted in the canonical (NVT) ensemble for 1 ns with a 2-fs time step, followed by equilibration in the isothermal–isobaric (NPT) ensemble for 200 ps with a 1-fs time step, using the leap-frog integrator. The temperature was maintained at 300 K using the velocity-rescaling (V-rescale) thermostat^20,21 with a coupling constant of 1 ps in both NVT and NPT ensembles. During NPT equilibration, the pressure was maintained at 1 bar using the isotropic Parrinello–Rahman barostat²² with a relaxation time (τp) of 5 ps. During the minimization and equilibration phases, heavy atom positions were restrained with a harmonic potential of 1,000 kJ · mol^−¹ · nm^−². Each system was then subjected to a 100-ns production MD simulation with a 2-fs time step, and coordinates were saved every 10 ps for subsequent analysis. Covalent bonds involving hydrogen atoms were constrained using the LINCS algorithm²³, and long-range electrostatic interactions were handled using the particle-mesh Ewald (PME) method²⁴.

The all-atom MD simulations were conducted using GROMACS v2024.1²⁵, leveraging a computational node equipped with 128-core AMD EPYC CPUs and eight NVIDIA RTX 4090 GPUs for acceleration.

MD trajectory analysis

To assess the behaviors of small proteins, we developed a custom analysis script that integrates the built-in GROMACS tools for comprehensive trajectory analysis. Users can perform an integrated analysis workflow for this dataset using the Jupyter notebook (https://github.com/RPDGroup/DynoDB), including: (1) analyzing conformational properties including root mean square deviation (RMSD), root mean square fluctuation (RMSF), and radius of gyration (Rg), Dictionary of Secondary Structure of Proteins (DSSP), and dihedral angles (φ/ψ); (2) identifying representative states through principal component analysis (PCA) and conformational clustering; and (3) calculating the solvent-accessible surface area (SASA) to describe physicochemical properties.

Data Overview

DynoDB serves as a data resource to facilitate research in small protein structure prediction, design, and dynamics. When compared with six widely used protein trajectory databases, this dataset fills the gap by providing small protein trajectory data. Notably, 7,139 of 8,385 trajectories in this resource are unique (Fig. 2a). The number of overlapping trajectories between this dataset and six existing resources is summarized as follows: DynamicPDB (729, 8.69%), mdCATH (378, 4.51%), ATLAS (130, 1.55%), MemProtMD (118, 1.41%), Chameleon (2, 0.02%), DPFs (0), and GPCRmd (0). We also summarized the protein structure determination methods (Fig. 2b). Among them, 6,241 structures (74.43%) were determined by nuclear magnetic resonance (NMR), and 2,050 structures (24.45%) were determined by X-ray crystallography. Specifically, when the protein length exceeds 55 amino acids, the number of X-ray crystallography-determined structures increases markedly.

Fig. 2.

Composition and feature analysis of the DynoDB dataset. (a) Overlap analysis with other protein dynamics databases. (b) Distribution of protein length and experimental determination method. (c) Distribution of functional categories. (d) Distribution of species. (e) Molecular weight distribution. (f) Atom count distribution. (g) Ternary plot of secondary structure composition.

We summarized the functional categories and species origins of the proteins within the dataset, including 17 functional categories and 1,392 species (Fig. 2c–d). Membrane-associated proteins, functional domain proteins, and metabolic enzyme fragments are the major functional types (4,983, 59.43%). In terms of species distribution, Homo sapiens accounts for the largest proportion, with 2,426 entries (31.12%), approximately seven times that of Mus musculus (327, 4.19%). Mus musculus ranks second, followed by synthetic construct (293, 3.76%) and Escherichia coli (216, 2.77%).

The structural analysis information such as molecular weight, atom count, and secondary structure composition is shown in Fig. 2e–g. A comprehensive description of the metadata fields for all 8,385 protein entries is provided in Supplementary Table 1 (detailed in the metadata file on Hugging Face). As illustrated in Fig. 2e and Fig. 2f, almost all proteins in this dataset have molecular weights ranging from 2 kDa to 10 kDa, and their atom counts are below 800. Finally, we analyzed the secondary structure composition of the static conformations of small proteins, and the results indicate that coils constitute the major component.

Data Records

Organization

The dataset is available at Hugging Face under a CC BY 4.0 license²⁶. Considering that the large size of the trajectory data (~9.1 TB) makes direct transmission inconvenient, we performed frame extraction on all trajectory files. Specifically, each trajectory was sampled every 100 ps to generate a structure file containing 1,000 conformations in PDB format. Users can download individual ensemble through the website interface, and batch downloads can also be performed via the API. Due to the large data volume and the presence of multiple file types, including structural, simulation, and annotation files, we adopted a hierarchical directory structure for data storage and management. At the top level, data are categorized by file type, and each category contains corresponding files for all 8,385 PDB entries. The detailed organization is as follows:

• static_structures/: Contains the initial PDB conformations for all entries (8,385 in total). These structures are consistent with those in the RCSB PDB. Files are named following the convention {PDB_ID}_static.pdb, e.g., 1A11_static.pdb.

• processed_structures/: Contains the standardized and preprocessed PDB files for all 8,385 entries. The preprocessing steps include completion of missing atoms (using PDBFixer v1.11), modeling of missing residues (using Boltz-1 or PDBFixer v1.11), and correction of modified residues (using Rosetta mutation and relaxation). Files are named following the convention {PDB_ID}_processed.pdb, e.g., 1A11_processed.pdb. These files can be directly used for molecular dynamics simulations.

• trajectory_frames/: Contains the frame-extracted results from molecular dynamics simulations. Each entry includes a 100-ns trajectory, sampled every 100 ps to generate 1,000 frames in PDB format. Files are named following the convention {PDB_ID}_frame.pdb, e.g., 1A11_frame.pdb. These files are publicly available and can be directly used for downstream analyses. Full 100-ns trajectory files are available upon request.

• metadata/: Contains all metadata and analytical results associated with each entry, including basic information (PDB ID, experimental method, resolution, sequence, etc.), structural and functional annotations (e.g., secondary structure composition, missing residue information, UniProt mapping, and functional labels).

Size

The dataset comprises a total of 8,385 independent molecular dynamics trajectories, with each trajectory spanning 100 ns. The cumulative simulation time amounts to 838.5 µs, with a total storage volume exceeding 9.1 TB. The key simulation metrics, including total atom counts, residue statistics, and storage requirements, are summarized in Table 3. However, considering that the simulated systems are short proteins with relatively low conformational complexity, the predefined simulation duration and number of replicates are relatively limited. To facilitate extension of the simulations, we provide the XTC files as breakpoints so that users can resume running if needed (full 100-ns XTC trajectories are available upon request). The conformational ensembles hosted at Hugging Face occupy approximately 611 GB.

Table 3.

Overall statistics of the DynoDB.

Dataset Metric	Value
Number of trajectories	8,385
Length per trajectory	100 ns
Cumulative simulation time	838.5 µs
Total number of atoms	3,990,516
Average system size (atoms)	~476
Total number of residues	491,395
Average system size (residues)	~59
Storage size	~9.1 TB

Technical Validation

Repaired structure validation

We first assessed structural completeness at both the atomic and residue levels (Fig. 3a–b). The results demonstrated a pronounced reduction in the number of missing atoms and contiguous missing residues. This confirms that the pipeline effectively reconstructed critical structural fragments and markedly enhanced overall model integrity. Next, we quantified the conformational fidelity between the repaired models and their reference structures using the TM-scores (Fig. 3c). The majority of repaired models achieved TM-scores between 0.8 and 1.0, indicating that the process successfully restored structural completeness while preserving the original conformational framework. Finally, Ramachandran plot analysis showed that the backbone dihedral angle distributions of the repaired structures were highly consistent with high-quality experimental models (Fig. 3d), with the majority of residues located within favored or allowed regions. Collectively, these results confirm the efficacy of the repair procedure in maintaining both geometric validity and conformational fidelity.

Fig. 3.

Evaluation of structural integrity and geometric quality before and after repair. (a) Violin plots of missing atoms (log-scaled) before and after repair. (b) Violin plots of maximum consecutive missing residues before and after repair. (c) TM-score distribution of repaired models compared to their original structures. (d) Ramachandran plot analysis of original (red) and repaired (blue) structures.

MD simulation pipeline validation

To ensure the reproducibility of our MD simulation pipeline, we conducted 100-ns MD simulations on 130 systems, comprising 30 Dual Personality Fragments (DPFs) and 100 chameleon sequences. The detailed list of these identifiers is provided in the metadata file hosted at Hugging Face. We then used Jensen-Shannon distance (JS) to characterize the similarity between our simulation samples (ours) and samples of ATLAS replicates (R1, R2, and R3). The JS distance for each pair of systems is calculated through the estimated distribution of α-carbon coordinates (Fig. 4a-f). The histogram plots of JS distances show that trajectories generated by our pipeline are highly consistent with the trajectories from the ATLAS (ranging from 0 to 0.05). A representative example is shown in Fig. 4b, where the kernel density contours of the protein (PDB ID: 2jg7_A) from our simulation (red) largely overlap with those from the three ATLAS replicates (R1–R3). Next, we assessed local flexibility by computing the Pearson correlation coefficients of residue-wise root-mean-square fluctuations (RMSF) between trajectories (Fig. 4h). Most correlation coefficients exceeded 0.8, suggesting strong consistency in residue-level dynamic behavior. In addition, we examined the global conformational compactness by analyzing the distributions of the radius of gyration (Rg) and evaluating their pairwise JS distances (Fig. 4i). The JS distances between Rg distributions were mostly below 0.1, reflecting minimal variation in overall compactness among replicates. Collectively, these results demonstrate that the established MD simulation pipeline achieves high stability and reproducibility consistent with the benchmark data from the ATLAS database.

Fig. 4.

Reproducibility and consistency analysis of MD simulations. (a) Distribution of JS distances for conformational sampling between all replicate pairs (ours, R1, R2, R3). (b) An example of PCA projection of protein 2jg7_A with KDE profiles, showing the overlap of conformational ensembles between repeated experiments (JS distances range from 0.003 to 0.0091). (c) Violin plots of Pearson correlation coefficients for residue-wise RMSF between replicates. (d) Violin plots of JS distances for Rg distributions between replicates.

Analysis of the randomness and sufficiency of the simulation pipeline

The current version of the dataset primarily encompasses 100-ns all-atom trajectories of small proteins (<100 amino acids) in the PDB. To assess how effectively these 100-ns simulations explore the conformational space, we analyzed three key factors: simulation length, the selection of the initial NMR conformation, and the exclusion of ions.

First, small proteins typically exhibit pronounced conformational flexibility. To evaluate whether a 100-ns simulation length is sufficient to adequately sample the conformational space, we selected 15 representative systems from the dataset of 8,385 proteins, extended their simulation times to 1 μs, and performed three independent replica simulations (Supplementary Figs. 1–15). These systems primarily comprise proteins characterized by high loop content (loop_fraction > 0.7) or a pronounced disorder propensity (IUPred_score > 0.7). Across these systems, the RMSD–Rg distributions of the initial 100-ns trajectories showed a certain degree of overlap with the RMSD–Rg distributions of the trajectories extended over the subsequent 900 ns; while the 1-μs trajectories explored a broader conformational space, particularly in systems containing high loop content or a high disorder propensity (Supplementary Figs. 1–14). Based on the Rg ranges, the conformational compactness sampled by the 100-ns trajectories substantially overlaps with that of the 1-μs simulations, indicating that the overall topological structure contour has been largely explored. However, projection onto the time-independent component analysis (tICA) space demonstrated that a 100-ns simulation duration is insufficient to access distinct, highly flexible states within the local loop regions. Furthermore, this structural transition is already observable within a 100-ns trajectory for the unfolded states of disorder-prone proteins (Supplementary Fig. 9). Therefore, we have included structural metrics such as the loop_fraction and IUPred disorder scores²⁷ in the metadata, enabling researchers to either selectively extend the simulations or deploy the existing trajectories directly based on their specific requirements.

Second, the dataset comprises a substantial proportion of NMR-resolved structures. Given that initiating simulations from different models within an NMR ensemble may induce variations in the resulting trajectories, we evaluated five randomly selected NMR systems. For each protein, alongside the primary model (Model 1), three alternative conformers displaying pronounced structural divergence were selected to undergo 1-μs simulations. For four of the five systems, the 100-ns trajectories initiated from different starting models exhibited substantial overlap (Supplementary Figs. 16–27). Extending the simulations to 1 μs further increased this conformational overlap while simultaneously exploring broader conformational regions. However, the 100-ns trajectories of 2YSA initiated from different NMR models exhibited significant discrepancies (Supplementary Fig. 28); their sampled conformational spaces gradually converged upon extension to 1 μs (Supplementary Fig. 29). These findings indicate that for NMR systems characterized by a degree of dependence on the initial conformation, a 100-ns simulation initiated from a single model may be insufficient to fully explore the conformational space. Consequently, we have included the structure determination method in the metadata, enabling researchers to filter or extend the trajectories according to their specific needs.

Third, during structural processing and simulation setup, all ligands, metal ions, cofactors, and chemical modifications were removed, which could potentially destabilize the native folds of certain systems. The observed late-stage (50 ns) RMSD increases across multiple systems suggest that these proteins may be transitioning toward alternative metastable states induced by ligand deprivation (Supplementary Fig. 31). To quantify this impact, we calculated the mean (rmsd_avg_last50ns) and standard deviation (rmsd_std_last50ns) of the RMSD relative to the starting conformation over the final 50-ns interval for all 8,385 trajectories. These statistical metrics have been integrated into the metadata, enabling researchers to assess trajectory convergence and equilibration according to their specific analytical criteria.

In summary, despite limitations regarding simulation duration, starting structure selection, and the simplification of the chemical environment, this dataset provides a foundational resource for the conformational analysis of small proteins. The current version is supported by comprehensive metadata to streamline analysis and maximize utility for the research community. These limitations will be addressed in subsequent updates.

Author contributions

Y.H., X.Y. and Z.R. conceived the study and wrote the original manuscript. Y.H. performed data processing, annotation, and built the database. X.Y. conducted molecular dynamics simulations. Z.R. provided project supervision. L.J., Z.F., M.N. and X.B. revised the manuscript and offered advice. Y.H., X.Z. and Z.S. carrying out initial background research for the project. All authors reviewed and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (Grant No. 32501102).

Data availability

The DynoDB dataset generated and analyzed in this study is publicly available on the Hugging Face Hub (https://huggingface.co/datasets/renzilin/DynoDB). A permanent DOI has been assigned to this dataset and can be accessed via the Hugging Face repository (https://doi.org/10.57967/HF/7057).

Code availability

Companion scripts providing automated routines for MD simulation, protein annotation, and trajectory analysis are available at GitHub (https://github.com/RPDGroup/DynoDB). These scripts ensure the reproducibility of the dataset construction. In addition, a subset of these scripts provides the operation content of the Technical Validation section.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41597-026-07759-2.