Research map / Probabilistic and scientific modeling / Molecules and proteins
Representing local protein environments with machine learning force fields
Meital Bojan, Sanketh Vedula, Sai Advaith Maddipatla, Nadav Bojan, Anar Rzayev, Federico Napoli, Paul Schanda, Alexander Bronstein
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MLFF embeddings (AIMNet) instead of ESM-based or learned features | GCN pKa predictor · pKa, Glutamic acid | MAE: 0.518 (learned) -> 0.265 | Improved |
| use MACE embeddings instead of UCBShift for CA | chemical shift predictor · backbone CA | MAE (ppm): 0.653 -> 0.584 | Improved |
| use MACE embeddings instead of UCBShift for HA | chemical shift predictor · backbone HA | MAE (ppm): 0.165 -> 0.180 | Hurt |
| use 5 A radius instead of 4 A or 6 A for the environment (ablation) | MACE + GCN · backbone chemical shifts | MAE (ppm): CA 0.620 / 0.627 -> 0.584 | Improved |
| concatenate first and second MACE message-passing layers (L1+2) (ablation) | MACE + GCN · backbone N shift | MAE (ppm): 1.771 -> 1.642 | Improved |
| use deeper Orb layer (L15 instead of L5/L10) (ablation) | Orb + GCN · backbone N shift | MAE (ppm): 1.889 -> 2.015 | Hurt |
| use six backbone atoms instead of CA only for secondary structure (ablation) | Egret + GCN · secondary structure | weighted F1 (%): 94.817 -> 98.633 | Improved |
| use six backbone atoms instead of CA only for amino acid identity (ablation) | Egret + GCN · amino acid identity | weighted F1 (%): 95.629 -> 99.135 | Improved |
| use MLFF embeddings instead of hand-crafted LocoHD descriptors | GCN · secondary structure | weighted F1 (%): 76.323 -> 95.693 | Improved |
| use kernel-density likelihood in MLFF embedding space as confidence | MACE shift predictor · chemical shift CA and N | prediction error | Improved |
| use MACE instead of pKa-ANI for histidine pKa | GCN · PKAD-R histidine | MAE: 0.899 -> 0.925 | Hurt |
| use AIMNet embeddings for pKa vs MACE embeddings | GCN · pKa prediction | MAE: Lysine 0.320 (MACE) vs 0.270 (AIMNet) | Mixed |
Closest papers by meaning
- Towards Fast, Specialized Machine Learning Force Fields: Distilling Foundation Models via Energy Hessians (ICLR 2025 · similarity 0.89)
- A Physics-Augmented Deep Learning Framework for Classifying Single Molecule Force Spectroscopy Data (ICML 2025 · similarity 0.87)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.86)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.86)
- AtomSurf: Surface Representation for Learning on Protein Structures (ICLR 2025 · similarity 0.86)
- Generalized Implicit Neural Representations for Dynamic Molecular Surface Modeling (AAAI 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.
