Molecular Dynamics Eval

molecular-dynamics-eval

Molecular Representations in Implicit Functional Space via Hyper-Networks — Wang et al. (2026) (arXiv:2601.22327, 2026)

What this evaluates

Evaluates a model's ability to reconstruct and predict time-varying 3D molecular surfaces in a continuous, resolution-independent manner. It measures volumetric overlap, point-to-point geometric distance, and surface normal alignment across diverse protein trajectories.

Datasets

• Sun et al. (2023) Protein Trajectories — total ?; splits: test (-1)

Metrics

• volumetric IoU (primary) — range: [0, 1]
  • Intersection over Union of predicted and ground-truth volumetric masks. Higher is better.
• Chamfer Distance (CD) — range: other
  • Average squared Euclidean distance between points in the predicted surface and the ground-truth surface, and vice versa. Lower is better.
• Normal Consistency (NC) — range: [0, 1]
  • Mean cosine similarity between predicted and ground-truth surface normals. Higher is better.

Input / output format

Input: 3D molecular trajectory data (time-varying atomic coordinates or surface points) representing protein conformations.

Output: Reconstructed continuous implicit neural representation or discrete surface mesh/predicted future frames.

Scoring recipe

iou = np.sum(pred_mask & gt_mask) / np.sum(pred_mask | gt_mask)
cd = 0.5 * (np.mean(np.min(pdist(pred_pts, gt_pts), axis=1)) + np.mean(np.min(pdist(gt_pts, pred_pts), axis=1)))
nc = np.mean(np.sum(pred_normals * gt_normals, axis=1))

Common pitfalls

• IoU and CD are sensitive to grid resolution and sampling density, requiring consistent voxelization across methods.
• NC measures surface normal alignment, which can be unstable for highly fluctuating or low-signal trajectories.

Evidence (verbatim from paper)

We compare against NDF, DSR, and CanFields, using volumetric IoU, Chamfer Distance (CD), and Normal Consistency (NC) as evaluation metrics.

Citation

@misc{wang2026molfield,
  title={Molecular Representations in Implicit Functional Space via Hyper-Networks},
  author={Wang et al. (2026)},
  year={2026},
  note={arXiv:2601.22327}
}

• arXiv: 2601.22327