PemNN: Polymer Physics for Single-Molecule Force-Spectroscopy Classification

0. Link verification table

All URLs were accessed during report production. PMLR directly provides the final PDF. OpenReview exists but automated access encountered verification, so only the PMLR-provided link is confirmed; OpenReview metadata is not quantitative evidence. Exact-title searches found no independently verifiable arXiv, Hugging Face Daily Papers, Papers with Code, or Semantic Scholar entry. ACL Anthology is unrelated to this topic.

1. Topic-selection audit: pool, deduplication, and random draw

The day's category is AI4Science and Interdisciplinary Research. Index deduplication excluded recently covered Foundation Molecular Grammar (molecular generation), TIMING (scientific/medical time-series explanation), and coordinate-robust regression. This round primarily used official ICML 2025 proceedings , cross-checked through OpenReview links, exact-title arXiv searches, and aggregators. Keeping venue status consistent prevents the random pool from hiding a main-conference versus preprint quality bias.

The pool contains 29 ICML 2025 main-conference papers, all within the 24-month limit:

As required, the system shell executed shuf -i 1-29 -n 1returned 8; the eighth paper was PemNN. The final paper, appendices, code, and data were sufficient for methodological and quantitative auditing, sothere were zero redraws. Exact-title searches on Hugging Face, Papers with Code, Semantic Scholar, and ACL Anthology returned no verifiable core-paper entry. This means those searches could not verify it, not that the paper does not exist.

2. Motivation: reliable labels, not prediction, are the expensive part

Single-molecule force spectroscopy (SMFS) stretches proteins with a probe and records force against extension. The scientific objective is often a domain's unfolding force, but a trial may capture no molecule, one molecule, or several, with overlapping curve shapes. The paper formulates three-way classification:no molecule / single molecule / multiple molecules. Experts traditionally inspect curves individually. The paper claims analysis drops from a day to under an hour (abstract p.1, conclusion p.8), but provides no timing table, sample size, or hardware range. This is a workflow-scale statement, not a rigorously measured speedup.

There are three difficulties. First, labels are scarce. Per-class counts across four experimental proteins are Titin [181,164,191], bacterial Utrophin [181,181,178], insect Utrophin [175,166,200]、Dystrophin [191,185,177]; each protein spans five to eleven experimental days and three to four pulling speeds (Table 1, p.6). The default low-data setting uses only about thirty training curves, with mean accuracy across four proteins of 79.6±5.2%(pp.1,7, Figure 6). Second, experimental curve lengths range from 736 to 9974 points, while networks require uniform input (Table 1, p.6). Third, purely data-driven models may exploit instrument noise, speed, or day-specific drift, while handwritten physics cannot fully represent protein heterogeneity. PemNN uses physical transforms as inputs to a learnable branch rather than demanding physics alone classify the curves.

A quantitative reference is Waite et al.'s DDRs data: a Triplet network achieves overall accuracy 66.70%, class accuracies 73.30/63.30/63.30%, and overall F1 66.70%(core reproduction Table 3, p.19; original work in Patterns full text). PemNN is not applied to DDRs because extension is absent, preventing WLC contour-length calculation. The method is therefore not a plug-and-play force-only classifier.

3. Argument, assumptions, and trade-offs

The narrative is clear: build a physical Monte Carlo engine generating 600 three-class curves for each of four proteins; let the network observe raw force and WLC-inverted contour length; select architecture on simulation, then validate few-shot and simulation-only pretraining on real experiments (pp.4–8; Appendices A–E, pp.13–24).

Four assumptions underlie this design.

1. Low-dimensional physical coordinates separate classes better. If single-molecule unfolding events are more stable in contour-length space than raw-force space, explicitly supplying Lc can reduce sample complexity.
2. Physics can be approximate but directionally useful. WLC, FJC, and FRC differ by no more than about three accuracy points on simulated/experimental data (Appendix E.1, pp.19–20, Figure 10), motivating the claim that exact model choice is not decisive. Only curves are given, without full matrices or CIs, so statistical equivalence cannot be assessed.
3. Simulation provides useful representation initialization. Pretraining only on corresponding simulations, with no experimental labels, yields mean four-protein accuracy 72.0±5.9%and mean ROC-AUC 0.88(p.7; Appendix E.4, pp.23–24). This shows nonzero transfer, not completed sim-to-real adaptation.
4. Early fusion learns complementarity more readily than late decision fusion. Early-Conv performs best in the critical-difference plot; late fusion of the two branches' final softmax outputs is weaker (Appendix E.1, pp.19–20, Figure 9).

The tradeoffs are equally clear. A physics branch introduces dependencies on kBT, fixed persistence length, contour-length thresholds, and extension quality. Simulation needs protein energy-barrier and unfolding parameters. In return, low-data behavior is more stable and screened unfolding-force distributions can be checked against physical expectations.

4. Formalization and procedure

For trajectory i with extension X^(i)=(x_1,…,x_T)、force F^(i)=(f_1,…,f_T) and label y^(i)∈{0,1,2}, the classifier outputs p_θ(y|X,F)and minimizes standard categorical cross-entropy:

L_CE(θ) = -(1/N) Σ_i Σ_{c=0}^2 1[y^(i)=c] log p_θ(c | X^(i),F^(i)).

This is Eq. (1), p.3. Physical enhancement occurs in the input transform. WLC maps each positive (x,f) pair to contour length Lc：

f = (k_B T / L_p) · [ 1/{4(1-x/L_c)^2} - 1/4 + x/L_c ].

This is Eq. (2), p.4. The authors fix L_p=300 pm, numerically solve pointwise for L_c, retain physically valid points satisfying L_c≤1000 nm , and separately resample force and contour-length sequences to 400 points (pp.4–5; Appendix D, p.18). Note:Lc is not determined solely by F ; it is jointly derived from (X,F) . If generic baselines see only F, PemNN receives both an extra sensor variable and a physical coordinate transform.

The default network can be written as:

h_F  = Branch_F(Resample400(Normalize(F)))
L_c  = SolveWLC(X,F; L_p, L_c,max)
h_L  = Branch_L(Resample400(Normalize(L_c)))
h     = Fuse_early(h_F,h_L)
p     = Softmax(Head(h))

Each branch starts with 1D convolution, batch normalization, and ReLU. Early-Conv concatenates features before two further convolution blocks and global average pooling. Another implementation uses an LSTM with hidden dimension eight and dropout 0.8 in each branch before concatenation (p.5; Appendix D, p.18). The complete training/deployment flow is:

                  ┌────────── Monte Carlo simulator ──────────┐
Protein physical parameters ────>│ WLC/FJC/FRC + DHS off-rate + stochastic unfolding │──> Simulated three-class curves
                  └────────────────────────────────────────────┘          │
Real AFM: (X,F) ───────────────────────────────────────────────────────────┤
                                                                          v
                ┌──────────── force branch ─────────────┐       Supervised cross-entropy
(X,F) ─────────>│ F → normalization/400 points → Conv/LSTM │──┐       │
                └────────────────────────────────────────┘  ├─fusion─>Three-way classification
                ┌──────────── physics branch ───────────┐  │
                │ (X,F)→WLC inversion Lc→filter/400 points→Conv/LSTM│──┘
                └───────────────────────────────────────┘
                                     │
                                     └─ single-class curves → unfolding-force distribution

The simulator does more than draw plausible curves: it links chain extension, cantilever equilibrium, Dudko–Hummer–Szabo (DHS) force-dependent off-rates, and stochastic unfolding. At each step, Algorithm 1 increases base extension, solves force balance, computes DHS off-rates/event probabilities, then samples domain unfolding or molecular detachment (Appendix A, pp.13–14). Table 2 gives Titin, bacterial Utrophin, Dystrophin, and insect Utrophin values for ln k0 as -9.21/-6.50/-4.50/-2.50，Δx‡ and for 0.40/0.85/0.60/0.41 nm，ΔG‡ and for 20/14.5/11.5/9.8 kBT(Table 2, p.15). These define the simulation domain and are also potential sources of transfer error.

5. Key mechanisms and engineering details

How far does the public repository support reproduction? Data directories contain four experimental/simulated dataset types and training scripts;CallScripts/ML_models/Titin contains two approximately 8.39 MB Keras checkpoints, but no complete four-protein weight set was found. The latest repository commit read for this report was 7875181(2025-06-05). Code/data are public, but pretrained weights do not directly verify every reported result.

6. Datasets, splits, and metrics

All counts come from Table 1, p.6. Four additional simulated datasets each contain 600 class-balanced curves for architecture comparison and pretraining (p.6). Default experiments use stratified random splits repeated five times. Low-data tests fix 80% for testing and progressively enlarge training subsets (pp.6–7; Appendix E.2, p.21). No leave-one-day, leave-one-speed, or leave-one-protein-out results are reported.

Accuracy is primary, with class-level ROC-AUC. Scientifically, predicted single-molecule curves are selected and their unfolding-force mode/IQR compared. Accuracy is interpretable for Table 1's roughly balanced classes, but single molecules may be rare at deployment, requiring precision, recall, PR-AUC, and misses per hundred curves. A threshold rule reportedly reaches average precision 0.93without corresponding recall (Appendix E.4, p.24, text after Eq.19).

7. Main results: complete matrices, recomputation, and boundaries

7.1 Experimental label budgets: six methods × four settings

This is the only main matrix giving per-method means and standard deviations for six methods. Parentheses are five-run standard deviations; higher accuracy is better, bold marks column bests. The header 5/80 means 80% fixed testing and only about 5% training, whereas later 40/60 settings use ordinary complementary splits. They are not one continuously defined ratio axis (Table 4, p.21).

Source: Table 4, p.21. This report recomputes gains over each column's strongest non-PemNN method:

Recomputation formula: (PemNN−competitor)/competitor; source values are from Table 4. The simple four-column average is PemNN 85.87%、LSTMFCN 81.64%, a difference of 4.23 pp. These columns share data and use different training sizes, so their average is not an effect size over independent samples.

The key result is not winning every column, but gains shrinking rapidly with labels: about thirty training curves yield 11.35 pp, while ample labels leave only 0.88 pp. This supports reduced low-data sample complexity through physical coordinates, not a large performance ceiling advantage at every budget.

7.2 DDRs: baseline matrix where PemNN is inapplicable

Source: Table 3, p.19. No PemNN result can be added because DDRs lacks extension. The matrix shows generic InceptionTime is strong with equal available information and reminds deployment teams to audit acquisition formats first.

7.3 Scientific downstream task: do screened unfolding forces look more physical?

First three columns: Table 5, p.21; last two recomputed using (baseline IQR−PemNN IQR)/baseline IQR . All four proteins show substantially narrower IQRs, stronger evidence than similar modes. But narrower does not automatically mean truer: a classifier might discard genuine heterogeneous conformations.

The text says PemNN and heuristic modes differ by no more than 10 pN, but Table 5 gives Titin |206.68−217.41|=10.73 pN; the other three differences are 2.44/5.92/1.84 pN. Thus the table and summary slightly disagree. Titin's expected literature range is 204±26 pN, containing Raw, Heuristic, and PemNN modes alike (p.7). The main improvement is IQR, not moving an incorrect mode to a correct location.

7.4 Sim-to-real: class AUC after simulation-only versus experimental training

Source: Table 7, p.24. The simple mean of twelve class AUCs increases from simulation training 0.819 to experimental training 0.932, a recomputed difference of 0.113. The difficult, scientifically relevant Single class still trails by 0.20–0.27on three heterogeneous proteins; therefore, 72.0±5.9% simulation-only accuracy is a useful cold start, not a replacement for expert labels.

The paper also compares screened mode/IQR after simulated versus experimental training: Titin 216.35(50.17) versus 217.75(36.17)，Bact 81.58(64.07) versus 85.38(22.28)，Insect 89.45(70.86) versus 96.65(40.20)，Dys 91.25(65.64) versus 79.75(25.09)(Table 6, p.23). Modes are broadly similar across four groups, but simulation-trained IQR is wider in every group, indicating incomplete classification-boundary transfer.

8. Layered ablations

Ablations cover three levels, with only some exact numbers.

Branch level. The paper compares dual-branch PemNN with force-only, physics-only, and baselines receiving physical inputs. Figure 5a reports mean accuracy across four simulated datasets of 98.9±1.9%; versus the strongest baseline, the largest gap is about fifteen points on Titin and the smallest about six on Insect UtrN (p.6, Figure 5a text). Without a per-dataset×method table, a complete five-benchmark×six-method simulation matrix cannot be reconstructed.

Fusion level. Early-Conv, Early-LSTM, and late fusion are compared. Early-Conv ranks best; the dual-LSTM variant has mean rank 2.4167(Appendix E.1, pp.19–20, Figure 9). No per-fusion accuracy means, standard deviations, or pairwise p-values are given, so rankings support early fusion without quantifying individual component effects.

Physical-model level. WLC, FJC, and FRC differ by no more than about three points on simulated/experimental curves (Appendix E.1, pp.19–20, Figure 10). An appropriate family of physical coordinates may matter more than one uniquely correct chain model, but equivalence tests are absent.

The crucial missing ablation is aF-only、F+raw X、F+shuffled Lc、F+learned monotonic transform(X,F)、F+WLC Lc five-way head-to-head. Because the physics branch introduces Xthat baselines lack, current ablations cannot distinguish extra observations, nonlinear transforms, and physical correctness.

9. Hyperparameter sensitivity and transitions

Contour-length thresholds range from 1,000 to 1,024,000 nm in fourfold increments, plus no filtering.1,000–64,000 nm is nearly perfect; at 1,024,000 nm PemNN still leads baselines by at least seven points, and without filtering by about twenty-five (p.6, Figure 5b). This is not a proved phase transition, but shows a mechanism: looser thresholds admit nonphysical solutions, generic models degrade faster, while the force branch provides a fallback.

Persistence length Lp Under 0.3–30,000 pm is scanned logarithmically over30–30,000 pm , where both sides approach perfect performance; under extreme mismatches at 0.3 pm and 3 pm , PemNN leads by at least about three and ten points, respectively (p.6, Figure 5c). This supports redundancy-based fault tolerance rather than a need for exact WLC parameters.

No sensitivity is reported for learning rate, batch, dropout, LSTM hidden size, resampling length, epochs, or simulation quantity, nor are there performance–compute curves. Physical inputs appear robust to two physical parameters; overall training robustness and tuning-free transfer to other instruments/proteins are not established.

10. Statistical significance and uncertainty

The paper uses Friedman testing followed by Wilcoxon signed-rank tests and Holm correction. Its critical-difference plot gives PemNN mean rank 1.4167and claims significance over five baselines (p.6, Figure 4). This is a reasonable multi-dataset framework, but Friedman statistics, degrees of freedom, raw/corrected p-values, effect sizes, and CIs are absent. Rankings over just eight tasks—four simulated and four experimental—may be affected by task dependence.

Table 4 reports five-run standard deviations without clarifying fixed versus resampled splits or providing paired-seed results. Under5/80 , PemNN's standard deviation 5.24 is smaller than LSTMFCN's 12.12(Table 4, p.21), suggesting higher low-data means and lower variability. Without individual scores, paired tests or credible intervals cannot be computed. Table 5's modes/IQRs describe distributions, not classification-performance significance.

The supported boundary is therefore:means and rankings consistently favor PemNN; statistically guaranteed gains in every new experimental domain are not established.

11. Genuine contributions, reusable design, and transfer potential

The contribution is neither WLC nor dual branches alone, but connecting a physics simulator that generates labels, fault-tolerant fusion of physical coordinates and raw observations, and scientific validation through unfolding-force distributions. The most reusable AI4Science design isa fallible physics branch plus a fallback data branch: rather than forcing a possibly mismatched equation, let it generate another representation and let fusion learn how much to trust it.

A second reusable idea separates sim-to-real deployment into two modes: with zero experimental labels, simulation pretraining provides 72.0±5.9% cold-start accuracy; with a few labels, roughly thirty curves raise average accuracy to 79.6±5.2%(pp.1,7). This is more decision-relevant than a generic claim that synthetic data works.

A third reusable idea is checking downstream physical quantities. IQR shrinks relative to raw data by 41.52%–77.56%(recomputed from Table 5), showing that classification changes scientific statistics. Such checks must jointly consider mode, IQR, known ranges, and independent experiments, rather than equating narrower distributions with greater accuracy.

Potential transfers include isotope/fragmentation-rule branches in mass spectrometry, equivalent-circuit parameters in electrochemistry, physiological equations for medical waveforms, and radiative-transfer coordinates in remote sensing. Physical representations must be stably computable from observations, with a raw branch protecting against mismatch. If extra sensors supply the physics branch, baselines must receive equal information.

12. Related-work chain: origins of the problem and design transfer

12.1 Dudko–Hummer–Szabo: force-dependent escape rates in the simulator

Problem solved. DHS theory connects force-driven barrier crossing with rupture/unfolding rates, enabling dynamic force spectroscopy to infer intrinsic rates and barrier locations/heights (PNAS 2008）。 Remaining gap. It explains single-event dynamics, not whether a complex AFM trace contains zero, one, or multiple molecules. Real instruments add cantilever effects, noise, and heterogeneous pathways.
Core response. PemNN's Monte Carlo engine samples unfolding/detachment at each extension step using DHS off-rates, generating three labeled curve classes (Appendix A, pp.13–15).
Design transfer. The transferred object is a generative mechanistic model, not a fixed classification threshold: dynamics theory generates the training distribution.

12.2 Waite et al.: from manual curve inspection to few-shot classification

Problem solved. Waite et al. formulate three-way AFM classification with few-shot Triplet learning and release DDRs data/labels. The core paper reproduces overall accuracy 66.70%(Table 3, p.19;Patterns 2023）。 Remaining gap. The data contain only force sequences from one day and speed, without explicit polymer physics; few-shot performance remains limited.
Core response. PemNN collects four-protein, multi-day, multi-speed data and adds WLC inputs and simulated pretraining.
Design transfer. The three-class task, public data, and Triplet baseline remain; few-shot learning shifts from pure representation learning toward reduced sample complexity through physical coordinates.

12.3 InceptionTime: a strong generic time-series backbone as a stress test

Problem solved. InceptionTime combines multiscale convolutions and ensembles for accurate, scalable time-series classification. Its original paper reports HIVE-COTE taking over eight days on N=1500,T=46 versus about one hour for InceptionTime, which can process eight million sequences in thirteen hours (arXiv:1909.04939v3, abstract). Remaining gap. It does not exploit SMFS extension–force relationships or provide a sim-to-real interface.
Core response. PemNN assigns observed morphology to generic convolutions and task-specific coordinates to physics. On DDRs, InceptionTime still ranks first at 85.48±1.14% , demonstrating a strong baseline rather than a strawman (Table 3, p.19).
Design transfer. Multiscale local convolutions become branch building blocks, no longer solely responsible for learning physical invariants.

12.4 Dingeldein et al.: from classification to simulator posteriors

Problem solved. Simulation-based inference encodes mechanisms in simulators and uses neural density estimation for posterior inference over hidden molecular parameters, bypassing intractable likelihoods of non-Markovian trajectories (arXiv:2209.10392v2）。 Remaining gap. The objective is parameter inference, not zero/single/multiple screening; simulator bias still affects the posterior.
Core response. PemNN chooses direct classification and uses a raw branch to reduce dependence on a perfectly correct simulator.
Design transfer. The physics-simulator-plus-neural-network paradigm remains, but amortized posteriors become synthetic pretraining and explicit physical features.

12.5 FUSION: retaining humans in the discovery loop for rare pathways

Problem solved. FUSION encodes trajectories as images and iterates model/expert labeling to identify single-molecule pathways, including pathways absent from initial labels (Nano Letters 2023）。 Remaining gap. Human participation reduces cold-start risk, but multi-protein scaling still requires expert iteration; physical coordinates do not explicitly enter the classifier.
Core response. PemNN generates three-class labels in simulation and reduces few-shot demand through WLC inputs.
Design transfer. They are complementary: PemNN supplies physical cold starts; FUSION supplies human correction after new pathways emerge. Combining them is more promising than choosing only one.

12.6 Nonlinear technical evolution tree

                         ┌─ WLC/FJC/FRC: chain-elasticity coordinates ─┐
SMFS physics ─ DHS dynamics ─┤                                ├─> Physical Monte Carlo labels
                         └─ cantilever / loading protocol ─────┘           │
                                                                            ├─> PemNN
Time-series learning ─ FCN/ResNet ─ InceptionTime ─ raw-force branch ────────┤
    │                                                                       │
    └─ metric few-shot ─ Waite/Triplet ─ three-class task/public benchmark ──┘

Mechanistic generation ─ simulation-based inference ─ parameter posterior
       ╲                                      ╱
        ╲──────────── sim-to-real calibration gap ───╱
                         │
Human-in-the-loop discovery ─ FUSION ──> Unknown pathways / anomaly classes ──> Next-generation open-set PemNN

Three shifts drive this evolution: from manual thresholds to few-shot classification; from raw sequences to explicit physical coordinates; and from requiring simulators to match reality to treating them as fallible but useful auxiliary views. Next, simulator discrepancy, day-specific drift, and unknown classes should be modeled together.

12.7 Cross-paper comparison

13. Author-stated limitations

1. Simulation parameters are hard to obtain. The authors note that Monte Carlo simulation needs energy-landscape parameters unavailable for many proteins. Multiple pulling speeds and constant-force experiments are needed to constrain the simulator more fully (Limitations, p.25).
2. Homogeneous-domain assumption. Training/test curves from the same protein/setup are treated as one domain. Heterogeneous domains and domain adaptation are future work (Limitations, p.25).
3. Insufficient sim-to-real transfer for heterogeneous proteins. The appendix notes greater conformational/experimental heterogeneity in Utrophin and Dystrophin; experimental labels increase Single-class AUC by at least 0.20(Appendix E.4, p.24, Table 7). This quantifies the second limitation rather than adding an independent one.

14. Independent critique: because X, therefore Y

1. [Causal attribution] Because the physics branch's Lc is computed jointly from (X,F) while generic baselines receive only F, PemNN's +11.35 pp low-data advantage cannot be attributed uniquely to physics; it may also reflect extra extension information. A F+raw X with F+shuffled Lc control is essential.
2. [Leakage / extrapolation] Because data span five to eleven days and three to four speeds but only curve-level stratified random splits are reported (Table 1, p.6), day-specific instrument drift and speed-specific shapes may occur in both train and test. Random splits may overestimate cross-day deployment performance.
3. [Model selection] Because the appendix selects the best model by minimum training loss, checkpoints monitor loss, and the API uses validation_split=0.001(Appendix D, p.18; official code), the 200-epoch result may favor training optimum rather than generalization optimum.
4. [Reproducibility] Because the public API's train_test_split Moderate random_state is commented out while example scripts fix seeds separately, API users and paper scripts follow different split protocols. Identical data/hyperparameters do not guarantee Table 4's five-run means.
5. [Statistics] Because Holm-corrected significance is claimed without p-values, statistics, or effect sizes, and ranks derive from only four simulated and four experimental tasks (Figure 4, p.6), readers cannot audit whether correlated tasks are effectively counted multiple times.
6. [sim-to-real] Because mean AUC across twelve classes rises from simulation training 0.819 to experimental training 0.932, with Single-class gaps on three heterogeneous proteins of 0.20–0.27(recomputed from Table 7), simulation is suitable for cold start but cannot yet replace labels. “No expert input” should be restricted to the operating point with mean accuracy 72.0±5.9% .
7. [Deployment metrics] Because at threshold α=10 the paper reports only mean precision 0.93without recall (after Eq.19, p.24), misses per hundred true single-molecule curves cannot be computed. Accuracy and precision alone do not establish rare-event screening usability.
8. [Scientific validity] Because Table 5's main improvement is narrower IQR while all three Titin modes lie within the expected 204±26 pN range, the model may improve purity or discard real heterogeneous pathways. Without an independent gold standard or blinded review, narrower cannot mean truer.
9. [Baseline fairness] Because baselines largely use sktime recipes and uniform 200-epoch training while PemNN's architecture/inputs are specialized for SMFS, and no equal-budget hyperparameter search is reported, architectural superiority is confounded with task priors and tuning effort.
10. [Internal consistency] Because the text claims mode differences no greater than 10 pNbut Table 5's Titin difference is 10.73 pN, the summary exceeds its rounding tolerance. This does not overturn the conclusion, but scientific quantities should follow the table.

15. Open challenges: directly executable head-to-head experiments

Theory

1. Extra information or physical inductive bias? With identical backbone, parameter count, and training budget, compare F、F+X、F+Lc、F + randomly permuted Lc、F + learned monotone transform(X,F); evaluate leave-one-day-out accuracy, Single PR-AUC, and calibration. If F+X≈F+Lc, extra information explains the gain; only a consistent advantage from true Lc supports the physical-coordinate account.
2. How should simulation-parameter uncertainty propagate? Compare fixed DHS/WLC parameters, prior-sampled domain randomization, and SBI posterior-sampled pretraining. Report sim-only AUC and learning curves with five, fifteen, and thirty experimental labels on four proteins. Test whether posterior coverage narrows Table 7's mean AUC gap of 0.113 .
3. When do dual branches cause negative transfer? Shift Lp, energy barriers, and cantilever stiffness by orders of magnitude from truth. Compare hard physics, gated fusion, and uncertainty-weighted fusion to locate where physics changes from helpful to harmful, beyond Figure 5's two parameters.

Engineering and applications

1. Real deployment splits. For each protein, compare leave-one-day/speed/instrument-out with the paper's random split at equal budgets. Report wall-clock time, human-review minutes, missed singles per hundred curves, and p95 latency. This directly tests the claimed day-to-under-an-hour end-to-end improvement.
2. Fair baselines. Give InceptionTime, LSTMFCN, and Transformers the same (F,X) two channels, hyperparameter-search counts, and parameter/FLOP limits as PemNN. A remaining advantage would rule out input/budget unfairness.
3. Unknown classes and active review. Add adhesion artifacts, partial folding, rupture, and new proteins as unknowns. Compare closed-set softmax, energy-based OOD, deep ensembles, and FUSION-style human review using risk–coverage, unknown AUROC, and expert-review workload.
4. Complete reproduction package. Freeze split manifests, all four protein checkpoints, containers, and seeds. An independent team should reproduce Table 4 on Linux GPUs and Apple Silicon. Success means means within one originally reported standard deviation, not one similar run.

Emerging directions

1. Physics-foundation-model pretraining. Compare separate per-protein simulator pretraining with cross-protein masked modeling plus physics tokens, then five-shot-adapt to unseen proteins. Test whether shared chain physics yields transferable representations.
2. Joint classification and parameter posteriors. Jointly train three-class and SBI posterior heads; compare classification-only, inference-only, and multitask models. Evaluate posterior coverage and unfolding-force credible intervals alongside accuracy rather than only point predictions.
3. Closed-loop experimental design. Use uncertainty to choose the next curve for labeling or pulling speed. Compare random labeling, entropy sampling, and FUSION-only review. Plot expert minutes against leave-day-out Single PR-AUC to measure real data efficiency.

16. Recent-work radar

1. Synthetic data-driven deep learning for label-free autonomous atomic force microscopy(Nature Communications 2026): synthetic AFM data supports label-free autonomous imaging, moving from offline screening toward choosing the next measurement. Examine whether cross-sample generalization is stricter than PemNN's within-protein random splits.
2. Automating the Expert Eye: A System-Agnostic Deep Learning Framework for Rare Event Discovery in Imbalanced Force Spectroscopy(arXiv:2606.09541, 2026-06-08): with target events at `13/970=1.34%`, reports accuracy `0.9196`, recall 0.9231, and removal of 880 nontarget trajectories (abstract/results). It directly addresses rare-event deployment absent from PemNN's balanced classification.
3. Atomic Force Microscopy Based Single Molecule Recognition: Deciphering Biological Interactions at the Nanoscale(Journal of Molecular Recognition 2026): traces errors from probe functionalization through force spectroscopy to recognition, providing experimental context for whether better classification improves biological conclusions.
4. Automated atomic force microscopy analysis using convolutional neural networks(2025): CNN-based AFM automation shows generic vision/signal backbones remain strong. Compare equal sensor inputs, not just model names, against PemNN.
5. Machine learning approaches for improving atomic force microscopy instrumentation and data analytics(Frontiers in Physics 2024): covers instrument control, imaging, and force-curve analysis, emphasizing acquisition–analysis loops. PemNN currently supplies only offline filtering within that loop.
6. Deep Learning Regressors of Surface Properties from Atomic Force Microscopy(Applied Sciences 2024): extends AFM from classification to continuous surface-property regression, suggesting joint unfolding-parameter and class prediction instead of hard labels alone.
7. Machine Learning for Analyzing AFM Images Generated from Polymer Blends(arXiv:2409.11438v2, 2024-09-20): compares frequency-domain features and ResNet on 144 images, reminding us that handcrafted transforms can be more data-efficient than large backbones in AI4Science—a methodological parallel to PemNN's physics branch (abstract/version history).

17. Conclusion and evidence boundaries

PemNN's strongest finding is low-label sample efficiency: with roughly thirty training curves, Table 4 gives 79.61±5.24%, exceeding the strongest LSTMFCN baseline by 11.35 pp; at 80% training, the gap shrinks to 0.88 pp. It is a case where physical priors matter most with few examples, not a new general-purpose time-series SOTA.

The key unresolved issues are attribution and domain shift. The physics branch also adds extension, random curve splits do not isolate days/speeds, and sim-only Single AUC trails experimental training on three heterogeneous proteins by 0.20–0.27. Evidence supports physical dual views as a useful engineering design, but not WLC itself as the cause or expert-free deployment.

Report footnote. The candidate pool, random command, URLs, final PDF, and public code were checked in this run. Exact numbers come only from read tables/text or explicitly marked arithmetic recomputations.ml-report-index.md This report was appended. HTML uses the Warm Editorial template with light/dark themes and mobile rules and passed its static validator with zero warnings. Playwright is installed, but /usr/bin/chromium、/usr/bin/chromium-browser with /opt/pw-browsers contain no executable Chromium, soreal-browser light, dark, and mobile screenshot checks could not be completed; static validation is not presented as a substitute for that review.