Teaching a Force-Trace Decoder Which Ruptures Are Real, Using Only Blocking Controls

When an atomic force microscope (AFM) tip coated with an antibody is pulled away from a cell, the recorded force curve contains a few sudden drops called ruptures. Some are the antibody letting go of its target receptor. Others are the tip simply sticking to the membrane or the glass. Biophysicists separate the two with a competitive displacement assay. They flood the sample with free antibody so the receptors are occupied, then check which ruptures disappear. Because receptors are finite and sticky background surfaces are effectively unlimited, whatever vanishes under blocking is the specific signal. Today this logic is applied by hand and only at the level of whole histograms. In Detecting CD20-Rituximab interaction forces using AFM single-molecule force spectroscopy, specificity was confirmed by noting that rupture peaks "disappeared" under excess free Rituximab. Individual events were never labelled as specific or non-specific.

We propose to turn the blocking control into a training signal. A sequence model reads a raw, unsegmented force trace and emits a sparse string of events, each tagged specific or non-specific. It learns these tags without per-event labels, using only which experimental condition each trace came from. The mechanism is connectionist temporal classification (CTC), a training loss that lets a network map a long signal to a short sequence of symbols by summing over every possible placement of those symbols in time. We extend it with two condition-dependent constraints. Under blocking, the decoder may emit no specific events. Without blocking, the number of non-specific events it emits must match the blocked population, so any excess has to be explained as specific. The claim is that the logic of the displacement assay, specific = total − blocked, can be written as a trace-level loss that yields per-event specificity calls. With those calls, single-cell and patient-level rupture-force distributions can be measured automatically instead of only inferred from population histograms.

What exists and what is missing

The file contains no work that connects displacement assays to CTC. The closest machine-learning work is A Physics-Augmented Deep Learning Framework for Classifying Single Molecule Force Spectroscopy Data. Its model, PemNN, classifies whole curves into no molecule, single molecule and multiple molecules using a raw-trace network fused with a worm-like-chain (WLC) polymer fit. It requires expert labels and never consults blocked controls. Our model instead decodes a variable-length event sequence and replaces expert labels for specificity with the experimental design. PemNN also supplies tools we reuse: its Monte Carlo curve simulator, its four-protein experimental data, and its few-shot setting.

Causalcall: Nanopore Basecalling Using a Temporal Convolutional Network supplies the architecture. It shows that a dilated convolutional network trained with CTC can turn raw biophysical current traces into symbol strings without a separate event-detection stage. Nanopore reads, however, have no notion of specific versus background binding.

Two CTC variants constrain the design. W-CTC: a Connectionist Temporal Classification Loss with Wild Cards adds a wildcard at the sequence edges so that unlabelled margins do not break training. On TIMIT with 50% masking it holds word error near 30%, while vanilla CTC rises to 80%. We need this because the tip-contact and retraction ends of a force curve carry adhesion artefacts that should not be forced into the label string. The same paper also warns that the wildcard slightly hurts performance on fully clean data, so its use is an ablation rather than an assumption. CR-CTC: Consistency regularization on CTC for improved speech recognition argues that CTC's sharp single-frame spikes overfit. We think rupture events are physically near-instantaneous force drops, so spiky emission may suit them better than it suits phonemes. We test this rather than assume it.

The ensemble statistics in Directly measuring single-molecule heterogeneity using force spectroscopy motivate the per-event goal. That work collapses survival curves across loading rates and finds heterogeneity in 5 of 10 published datasets, with disorder values from 1.5 to 13.3. It assumes every rupture is a target event. A contaminating pool of non-specific events could masquerade as heterogeneity. Filtering events by our decoder before applying that test is one downstream use.

Core Claim

Two conditions make competitive displacement identifiable at the event level. First, blocked traces are a clean sample of the non-specific process. Second, the non-specific process is unchanged by the competitor. Together these give enough supervision to train a segmentation-free event decoder whose specificity calls match those of a decoder trained on full per-event labels.

This holds only if specific and non-specific ruptures differ in some trace-visible feature. AFM gives one: a specific event pulled through a polyethylene glycol (PEG) linker shows nonlinear linker stretching at an extension set by the linker length, while tip–membrane adhesion typically ruptures near contact. When no such feature exists, the method should fail, and we treat that as a predicted negative result.

Research Program

The decoder is a dilated causal convolutional encoder in the Causalcall style. It emits, per frame, probabilities over blank, specific rupture $S$, and non-specific rupture $N$. A front wildcard and a flexible end, following W-CTC, absorb the contact and free-retraction regions. Training combines three terms. The first is ordinary CTC on a small labelled set, consisting of synthetic curves plus about 30 expert-labelled traces to match PemNN's few-shot regime. The second and third are the displacement terms. Let $p_t(S)$ and $p_t(N)$ be the decoder's frame-$t$ probabilities, and let $\mathcal{B}$ and $\mathcal{U}$ be the blocked and unblocked trace sets. The displacement loss is

$$ \mathcal{L}_{\text{disp}} = \frac{1}{|\mathcal{B}|}\sum_{X\in\mathcal{B}} -\sum_t \log\big(1-p_t(S)\big) \;+\; \beta\Big(\mathbb{E}_{\mathcal{U}}\big[\textstyle\sum_t p_t(N)\big]-\mathbb{E}_{\mathcal{B}}\big[\textstyle\sum_t p_t(N)\big]\Big)^2 . $$

The first term is exactly the CTC probability that a blocked trace's path contains no $S$ symbol, since summing over all paths using only blank and $N$ factorises over frames. The second term encodes the saturability argument: the expected number of non-specific emissions must be the same with or without competitor. Peaky CTC emissions make $\sum_t p_t(\cdot)$ a good estimate of event counts, which is why the term is written this way. Blocking is rarely complete, so we add a blocking-efficiency parameter $\eta$. A fraction $1-\eta$ of specific events is allowed in blocked traces, and we sweep $\eta$ to measure how robust the method is to incomplete blocking.

We build three test beds. First, a synthetic benchmark with ground truth, made by extending the PemNN simulator. It mixes Bell–Evans-style specific ruptures behind a PEG linker with non-specific adhesion events, and controls class separability, noise, blocking efficiency and the fraction of specific events. Second, a semi-real benchmark from PemNN's four-protein AFM data (Titin I27O, bacterial and insect UtrN-R3, DysN-R3). Expert-labelled no molecule traces play the blocked pool and the full set plays the unblocked pool, so the method trains without event labels and is scored against expert labels. Third, a new wet-lab dataset that repeats the Rituximab–CD20 protocol with and without free Rituximab on purified CD20, Raji cells and primary B cells from more than one donor. This last test bed is the actual application. It costs real lab time, and its value depends on the first two succeeding.

Baselines

The naive reference is histogram subtraction, the displacement assay as practised in the CD20–Rituximab paper. Rupture-force histograms from unblocked and blocked conditions are subtracted, and the residual is read off. It gives no per-event calls, so we compare on the estimated specific fraction and on the shape of the specific force distribution, measured as Wasserstein distance to the synthetic ground truth. Any learned method that cannot recover a closer distribution than plain subtraction has no reason to exist.

The strongest existing learned system is PemNN. It reports 79.6% average three-class accuracy with about 30 labelled samples, 72.0% zero-shot from simulation, and an 11.4% margin over InceptionTime and LSTM-FCN (PemNN). Our decoder's count of specific events maps directly to PemNN's three classes: zero, one, or more than one. On PemNN's own data we run it under the same 30-label budget and the same simulator pretraining. Beating it means exceeding 79.6% while using no expert labels for the blocked/unblocked split. For per-event specificity there is no reported result found on Lacuna, so we report against our own baselines.

The most dangerous baseline involves no sequence model. It runs prominence-based peak detection with a WLC fit around each peak, then trains a per-event classifier in positive–unlabelled fashion, with events from blocked traces as known negatives and events from unblocked traces as an unlabelled mixture. It receives identical traces, the same labelled set and the same hyperparameter search budget. We also run vanilla CTC trained on full per-event labels as an oracle upper bound. Ablations remove the displacement terms (simulator plus few-shot only), the wildcard edges, and the rate-matching term (keeping only the blocked "no $S$" term). The rate-matching ablation isolates the saturability assumption: without it, the decoder can trivially label every event $N$, so a large drop is expected and confirms that this term carries the identification.

Why not just…?

Why not just detect peaks and train a positive–unlabelled classifier on them? This is the real threat. On clean curves with well-separated ruptures it may match our method, and if so the CTC machinery is unnecessary there. The case for sequence decoding rests on traces where peak detection fails. On cells, closely spaced multi-bond ruptures and adhesion events inside the linker-stretch region break per-peak segmentation, and PemNN notes that heuristic thresholds cannot separate adhesion spikes from true events without context. The synthetic benchmark sweeps event spacing and noise precisely to find the crossover. If no crossover appears even at Raji-cell noise levels, we will report that the contribution is the displacement-derived supervision and that a peak-level classifier is the right way to use it.

Why not just train a trace-level classifier to tell blocked from unblocked curves? An InceptionTime-style classifier trained this way does learn something correlated with specificity. It cannot say which event in a trace is specific or how many there are, and these are what single-bond force estimates need. PemNN's 11.4% margin over such classifiers in the low-label regime also suggests they are weak here. We include it as a trace-level baseline on the semi-real benchmark.

Why not just use histogram subtraction or the loading-rate collapse test? Both are ensemble methods. They need hundreds of curves per condition, and they cannot clean individual traces or produce per-cell estimates. The heterogeneity analysis assumes no non-specific contamination. We test directly whether filtering events with our decoder changes the disorder parameter on synthetic data with injected adhesion.

Why not just use vanilla CTC with the specific/non-specific labels? That requires the per-event labels this project aims to avoid, so we use it only as the oracle. Without wildcard edges, W-CTC's TIMIT result (80% versus about 30% error under unlabelled margins) predicts failure on contact artefacts. The wildcard ablation checks whether that transfers to force traces.

Decisive Checks

The idea succeeds if three things hold. On the synthetic benchmark, displacement-trained CTC comes within a few points of per-event F1 of the fully supervised oracle, and it beats the peak-detection positive–unlabelled baseline wherever events overlap. On the four PemNN proteins it exceeds 79.6% three-class accuracy. On new CD20 data it recovers distinct specific-force modes for purified CD20, Raji cells and primary cells, consistent with the reported roughly 121, 89 and 126 pN, with no per-event annotation.

It fails informatively in two ways. If removing the rate-matching term does not hurt, the identification argument is wrong. If performance collapses as the specific and non-specific feature distributions overlap, that confirms the stated boundary of the method rather than refuting it.