Research map / Probabilistic and scientific modeling / Molecules and proteins
Sliced Denoising: A Physics-Informed Molecular Pre-Training Method
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes sliced denoising (SliDe), a molecular pre-training method that adds BAT (bond/angle/torsion) noise derived from a quadratic classical mechanical energy and uses random slicing to avoid computing the Jacobian when learning the force field.
Problem
Existing denoising-based molecular pre-training uses ad-hoc noise and coarse energy functions, so the learned force fields deviate from true physical force fields, limiting downstream generalization.
Approach
Derive a quadratic intramolecular energy in bond lengths, angles, and torsions, sample conformations with BAT Gaussian noise whose variances come from force-field parameters, and regress the Cartesian force field via random slicing (projections onto Gaussian random vectors) so that the Jacobian of the internal-coordinate transform never has to be computed explicitly.
Key results
- SliDe reaches a force-field correlation coefficient of 0.895 with DFT versus 0.616 (Coord) and 0.631 (Frad), a 42% improvement.
- On QM9, SliDe achieves state-of-the-art on 10 of 12 tasks, reducing MAE by 12.4% relative to the prior best.
- On MD17 force prediction with 1000 training samples, SliDe achieves the best MAE on six of eight molecules (e.g., Aspirin 0.1740 vs 0.2087 for Frad).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SliDe instead of Frad for denoising pre-training | Molecular force-field learning · PCQM4Mv2 force-field evaluation | correlation coefficient: 0.631 (0.046) -> 0.895 (0.071) | Helped |
| Use SliDe instead of Coord for denoising pre-training | Molecular force-field learning · PCQM4Mv2 force-field evaluation | correlation coefficient: 0.616(0.047) -> 0.895 (0.071) | Helped |
| Add pre-training regularization (ablation) | SliDe · QM9 homo | MAE (meV): 15.0 -> 13.6 | Helped |
| Add pre-training regularization (ablation) | SliDe · QM9 lumo | MAE (meV): 14.8 -> 12.3 | Helped |
| Add pre-training regularization (ablation) | SliDe · QM9 gap | MAE (meV): 27.7 -> 26.2 | Helped |
| Use GET instead of ET (ablation) | SliDe pre-training and MD17 fine-tuning · MD17 force prediction: Aspirin | MAE: 0.2045 -> 0.1740 | Helped |
| Use GET instead of ET (ablation) | SliDe pre-training and MD17 fine-tuning · MD17 force prediction: Benzene | MAE: 0.1810 -> 0.1691 | Helped |
| Increase random slicing sample count from N_v=32 to N_v=128 at σ=0.001 | SliDe · PCQM4Mv2 force-field evaluation | Pearson correlation coefficient ρ: 0.536(0.067) -> 0.895 (0.071) | Helped |
| Increase random slicing sample count from N_v=128 to N_v=512 at σ=0.001 | SliDe · PCQM4Mv2 force-field evaluation | Pearson correlation coefficient ρ: 0.895 (0.071) -> 0.896 (0.067) | No clear effect |
Surprising findings
- SliDe is worse than Frad on ZPVE and SphereNet/ET outperform it on a few QM9 tasks, and it is worse than Coord on Malonaldehyde and Uracil on MD17.
Limitations
- Noisy Nodes fine-tuning was hard to optimize with SliDe, requiring an added regularization term during pre-training.
- Ring bonds, angles, and torsions are not perturbed, and atoms bonded to more than two atoms need special handling for independence.
- Pre-training uses only equilibrium conformations from PCQM4Mv2 and the energy function omits electrostatic and van der Waals terms.
Experimental setup
Datasets: PCQM4Mv2 QM9 MD17
Benchmarks: QM9 MD17
Metrics: MAE Pearson correlation coefficient
Baselines: Coord Frad 3D-EMGP SE(3)-DDM Transformer-M TorchMD-NET (ET) SphereNet PaiNN DimeNet++ SchNet
Builds on: coordinate denoising fractional denoising Sliced Wasserstein distance TorchMD-NET Noisy Nodes
Closest papers by meaning
- Learning 3D Anisotropic Noise Distributions Improves Molecular Force Fields (NeurIPS 2025 · similarity 0.90)
- SubgDiff: A Subgraph Diffusion Model to Improve Molecular Representation Learning (NeurIPS 2024 · similarity 0.87)
- 3D Denoisers Are Good 2D Teachers: Molecular Pretraining via Denoising and Cross-Modal Distillation (AAAI 2025 · similarity 0.87)
- Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation (AAAI 2024 · similarity 0.86)
- Structure-based drug design by denoising voxel grids (ICML 2024 · similarity 0.85)
- The dark side of the forces: assessing non-conservative force models for atomistic machine learning (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.