Research map / Probabilistic and scientific modeling / Molecules and proteins
A Physics-Augmented Deep Learning Framework for Classifying Single Molecule Force Spectroscopy Data
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 PemNN, a dual-branch 1D CNN-LSTM that fuses a worm-like-chain physics branch with a raw force-trace branch to classify SMFS curves as no, single, or multiple molecules, and releases simulated and experimental datasets.
Problem
Classifying single-molecule force spectroscopy (SMFS) force curves as no-molecule, single-molecule, or multiple-molecule events currently requires slow manual visual inspection by experts, and no public non-specific-pulling multi-domain SMFS time-series datasets existed.
Approach
A dual-branch deep network (PemNN) where one branch filters and encodes force-extension data using worm-like-chain contour-length estimates and the other learns directly from force traces, fused via early or late fusion, plus a physics-based Monte Carlo engine to generate labeled simulated pretraining data.
Key results
- PemNN achieves 98.9 ± 1.9% accuracy on four simulated datasets, outperforming baselines by 6-15%.
- On experimental data with ~30 training samples (5%), PemNN reaches 79.6 ± 5.2% average accuracy, 11.4% above the best baseline LSTMFCN.
- Pretraining only on simulated data yields 72.0 ± 5.9% average accuracy on experimental data; PemNN's Titin I27O most probable unfolding force is 206.68 pN vs expected 204 ± 26 pN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the polymer-elastic-model physics-based branch to the force-trace model (ablation) | PemNN · Four simulated protein datasets | accuracy: 98.9± 1.9 %; outperformed baselines by 6% | Helped |
| Use the force-trace branch alongside the physics-based branch (ablation) | PemNN · Simulated datasets with contour-length threshold 1,024,000 nm | accuracy: at least 7% | Helped |
| Use the force-trace branch alongside the physics-based branch (ablation) | PemNN · Simulated datasets with no contour-length threshold | accuracy: 25% | Helped |
| Use the force-trace branch alongside the physics-based branch (ablation) | PemNN · Simulated datasets with persistence length Lp = 0.3 pm | accuracy: at least 3% | Helped |
| Use the force-trace branch alongside the physics-based branch (ablation) | PemNN · Simulated datasets with persistence length Lp = 3 pm | accuracy: 10% | Helped |
| Train with 5% rather than more experimental training data (ablation) | PemNN · Experimental datasets | accuracy: 79.6% and 85.2% | Mixed |
| Pre-train on corresponding simulated datasets without experimental training data (ablation) | PemNN · Experimental datasets | accuracy: 72.0±5.9% | Helped |
Surprising findings
- Pretraining solely on simulated data with a single homogeneous domain model still classifies heterogeneous experimental proteins at 72.0% accuracy.
Limitations
- Experimental labels come from visual inspection by experts, which may carry human bias.
- Physics model assumes all protein domains behave identically, which is violated for heterogeneous proteins like utrophin and dystrophin.
- Performance on experimental data (79.6%) is well below near-perfect simulated accuracy.
Experimental setup
Datasets: Titin I27O experimental bact UtrN-R3 experimental insect UtrN-R3 experimental DysN-R3 experimental DDRs (Waite et al., 2023) Monte Carlo simulated SMFS datasets
Benchmarks: simulated SMFS datasets experimental SMFS datasets (Titin I27O, bact UtrN-R3, insect UtrN-R3, DysN-R3) DDRs dataset
Metrics: accuracy average rank IQR of unfolding force most probable unfolding force
Baselines: Triplet network FCN ResNet InceptionTime LSTMFCN
Builds on: worm-like chain model InceptionTime LSTM-FCN Triplet network
Closest papers by meaning
- Towards Fast, Specialized Machine Learning Force Fields: Distilling Foundation Models via Energy Hessians (ICLR 2025 · similarity 0.87)
- Predicting mutational effects on protein binding from folding energy (ICML 2025 · similarity 0.86)
- UniSim: A Unified Simulator for Time-Coarsened Dynamics of Biomolecules (ICML 2025 · similarity 0.85)
- Sliced Denoising: A Physics-Informed Molecular Pre-Training Method (ICLR 2024 · similarity 0.85)
- Protein Conformation Generation via Force-Guided SE(3) Diffusion Models (ICML 2024 · similarity 0.84)
- A Systematic Evaluation of Co-folding Model Representations for Small-Molecule Learning (NeurIPS 2026 · similarity 0.84)
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.