Antigen Sinks: Scoring HER2 Slides by Their Worst-Dosed Regions

Pathologists summarize a HER2-stained tumor slide with one number or one grade. The H-score is a weighted percentage of weak, moderate, and strong cells. The HER2 category (0, 1+, 2+, 3+) is set by how many cells show complete membrane staining. Both treat strong staining as good news for anti-HER2 antibodies such as trastuzumab. Wireless networks offer a cautionary parallel. When phones attach to whichever tower has the strongest signal, the powerful tower is swamped and nearby small cells sit idle, so a good signal does not translate into good service. The tumor analogue works like this. An antibody leaves a blood vessel and is captured by the first dense cluster of receptors it meets. The most strongly stained regions can therefore act as sinks that soak up drug, while moderately stained or poorly vascularized regions further away receive little. Two slides with the same H-score could then have very different worst-dosed regions. Population-genetics work suggests those under-dosed pockets are exactly where resistant cells survive and regrow: Spatial structure facilitates evolutionary rescue by drug resistance shows with island models that isolated subpopulations shelter rare resistant mutants from eradication.

We propose to turn a stained slide into a map of predicted drug exposure and to score the tumor by a tunable average that interpolates between the slide mean and its worst region. The method has three steps. A cell-level intensity map from the IHC slide and a vessel map from the paired H&E slide feed a simple steady-state diffusion model in which receptors consume drug. Each region gets a predicted target engagement. These are pooled with the "alpha-fair" family of averages used in wireless scheduling. The claim is testable and could fail. In HER2-positive patients treated with anti-HER2 therapy, the tail of predicted engagement should predict response better than H-score, HER2 grade, or a tail statistic computed from staining alone. The improvement should come specifically from the drug-consumption term.

Core Claim

Spatial arrangement of HER2 expression, and not only its composition, carries response information. It does so through a sink effect that standard scores cannot represent. The sink model predicts something counterintuitive that a composition-only view cannot. A 2+ region lying downstream of a dense 3+ cluster should be engaged worse than an identical 2+ region in a uniformly 2+ tumor. Response should therefore be worse for tumors whose strongest staining is concentrated near vessels with weaker staining at the periphery, even after matching on H-score and on the distribution of staining levels. If the data instead show that the slide's composition explains everything, the claim is wrong.

Relation to Prior Work

Automated HER2 scoring is mature but aims only to reproduce the pathologist's grade. In Her2 Challenge Contest, 18 pipelines were benchmarked on 172 slides. Algorithms matched experts on 3+ cases and overcalled 2+ cases far less often than pathologists. PTree-Net scores slides from multi-magnification patch trees under weak labels. HER2Net predicts gastric HER2 status from H&E by thresholding the high-expression area at 10%. HEROHE explicitly excluded intra-tumoral HER2 heterogeneity from evaluation. These papers supply our tools for extracting cell-level maps. They motivate the work by showing that the grade itself is the target and that spatial layout is discarded, but none links layout to treatment outcome.

Response-prediction models motivate the gap from the other side. DeepCDR and MMCL-CDR work on cell lines, which have no tissue architecture. THERAPI models patients as mixtures of cell lines from bulk expression, so it captures heterogeneity but not space. MuMo predicts anti-HER2 response in 429 gastric cancer patients by fusing pathology, CT, and clinical data into global embeddings, without any regional notion of dose. Mechanistic models constrain our design. Raimúndez et al. show that receptor internalization and turnover shape targeted-antibody response, which justifies a consumption term. They also show that kinetic parameters are poorly identifiable, which forces us to collapse the transport model to a single dimensionless parameter. Steering the Evolutionary Game controls clonal dynamics in well-mixed compartments and names spatial reaction-diffusion as its open problem.

The aggregation step reuses wireless results. Ye et al. show that once you count each user's share of a congested tower's resources, signal strength alone mispredicts delivered rate, and that a logarithmic (proportional-fair) utility is a natural compromise. Zhu et al. use smooth approximations of the worst-user rate to avoid the collapse that strict max-min fairness causes. We borrow both the "share of a congested source" logic and the family of utilities between average and minimum. We do not borrow any wireless optimizer. The idea that antibody binding near vessels limits penetration is a known concern in antibody pharmacology, but no paper in our material connects it to semi-quantitative IHC scoring or tests it against clinical response.

Research Program

The pipeline has three parts.

The first part produces maps. We train a cell-level HER2 membrane-intensity classifier (0 to 3+) and validate it against the Her2 Challenge annotations. Tumor regions and vessels are segmented on the co-registered H&E. The output is a receptor-density field $\rho(x)$ and a set of vessel source locations on a tissue graph of roughly 50 µm tiles.

The second part predicts exposure. We solve a steady-state diffusion-with-consumption equation on that graph:

$$D\nabla^2 c(x) - k\,\rho(x)\,c(x) = 0,\quad c = c_0 \text{ at vessels},$$

Here $c$ is free antibody concentration, $D$ is a diffusivity, $k$ is a binding/internalization rate, and $c_0$ is the plasma level. Only the penetration length $\lambda = \sqrt{D/(k\bar\rho)}$ affects the relative exposure pattern, where $\bar\rho$ is the mean receptor density. The scale $c_0$ sets absolute occupancy, and we fix it to a typical trough concentration. Regional engagement is $e_i = \rho_i \cdot c_i/(c_i + K_d)$, where $K_d$ is the antibody's binding constant. This is the density of receptors actually bound in tumor region $i$.

The third part pools regions. Tumor-level engagement is the power mean

$$U_\alpha = \Big(\tfrac{1}{n}\textstyle\sum_{i=1}^n e_i^{\,1-\alpha}\Big)^{1/(1-\alpha)},$$

taken over $n$ tumor regions. With $\alpha=0$ this is the plain average, closely related to the H-score. With $\alpha=1$ it is the geometric mean, which corresponds to the proportional-fair choice. As $\alpha\to\infty$ it approaches the worst region. Two numbers are fitted on training folds only: $\alpha$ and $\lambda$, the latter scanned over a physiologically plausible range. Each answers a scientific question. A fitted $\alpha$ near zero says response follows the average. A large $\alpha$ says it follows the tail. A finite $\lambda$ that beats $\lambda=\infty$ says the drug-consumption effect matters.

The essential dataset is a retrospective cohort of HER2-positive gastric or breast cancer patients. Each patient needs a pre-treatment HER2 IHC slide, a paired H&E slide, and a recorded response to trastuzumab-based therapy. MuMo's 429-patient gastric cohort shows that such cohorts exist. We will seek access to it or assemble a comparable multi-center cohort of a few hundred patients. Without response labels the project reduces to simulation, and we state this as the main feasibility risk.

Baselines

The naive reference is the clinical HER2 grade (2+ versus 3+), alone and combined with standard clinical covariates. Next comes the continuous digital H-score and the high-expression area fraction, the quantities HER2Net thresholds. We fit both by logistic regression on identical folds. No reported result was found on Lacuna for H-score-based prediction of anti-HER2 response.

The ablation that isolates our key ingredient is the identical power-mean pipeline with the consumption term switched off ($\lambda=\infty$). Exposure is then uniform and $U_\alpha$ reduces to a tail statistic of staining alone. It gets the same $\alpha$ search, so the only difference from the full model is the sink effect.

The strongest learned competitor is attention-based multiple-instance learning as in Ilse et al.. This model learns a weighted average over patches and is trained end to end on the IHC and H&E patches to predict response. We give it the same folds and patch extraction and a hyperparameter budget at least as large as our two-parameter scan. A PTree-Net-style multi-magnification scorer is a secondary learned baseline.

The strongest published number on the closest task is MuMo's AUC of 0.821 for anti-HER2 therapy on its gastric cohort (MuMo). This comparison is only direct if we obtain that cohort. On any other cohort we will reimplement its pathology-plus-clinical branch rather than quote the figure.

Beating the baselines means two things. First, $U_\alpha$ at a finite $\lambda$ must improve AUC over both H-score and the $\lambda=\infty$ ablation, with a bootstrap confidence interval excluding zero. Second, it must add significant information when included alongside clinical covariates and the attention-MIL score. Matching attention-MIL with two interpretable parameters would itself be a useful result. Exceeding it would be strong evidence.

Why not just…?

Why not just let attention-based MIL learn the spatial pattern end to end? With a few hundred patients, an attention model must discover from scratch that a region's value depends on its distance to vessels and on the receptor density along the path. The sink effect is a non-local, path-dependent interaction, which attention pooling over independent patch embeddings does not represent naturally. Attention also tends to concentrate on the most discriminative patches. We cannot rule out that attention-MIL learns enough. The head-to-head comparison on identical folds settles it. So does a residual test: if adding $U_\alpha$ to the attention score adds nothing, our mechanism is redundant.

Why not just use composition-only heterogeneity statistics, such as the fraction of 0/1+ tumor or a low percentile of local HER2? This is the most serious threat, and the $\lambda=\infty$ ablation is exactly this alternative given the same tail-pooling. If it performs as well as the full model, spatial arrangement adds nothing beyond composition, and the sink hypothesis is refuted for this endpoint. We additionally run a matched test. Within strata of equal H-score and equal staining histogram, we compare tumors whose strong staining sits near vessels against tumors whose strong staining is distal. Only the sink model predicts that the near-vessel group responds worse.

Why not just fuse everything into one multimodal model, as MuMo does? MuMo's reported AUC (0.821) is competitive. But its risk scores lean on macroscopic clinical features such as peritoneal metastasis, and its global embeddings cannot say which region is under-treated. Our goal is a measurable, falsifiable mechanism rather than the top AUC. We test whether $U_\alpha$ remains informative when added to a fused model, which is the relevant question.

Why not just score the single worst region? This is the $\alpha\to\infty$ end of our family, so it is tested rather than assumed. The wireless experience (Zhu et al.) and the noise of single-tile estimates both suggest that the pure minimum will be unstable. A fitted intermediate $\alpha$ would confirm that a soft tail is better.

Proposed Experiments and Decisive Checks

We begin in simulation to confirm the mechanism can be detected at realistic cohort sizes. We take real cell-level HER2 maps extracted from public scoring datasets (the Her2 Challenge and HEROHE slides) and simulate exposure and regional killing. We then generate synthetic responses under two hypotheses: response driven by the sink-affected tail, and response driven by composition alone. This determines whether our pipeline recovers the true $\lambda$ and $\alpha$, and how many patients are needed to separate the hypotheses at 80% power. Map quality is checked separately by reproducing slide grades. Agreement on 3+ and on 2+ overcalling should be comparable to the Her2 Challenge pipelines, so that downstream failures cannot be blamed on bad maps.

On the clinical cohort the primary endpoint is objective response. Secondary endpoints are progression-free survival via a Cox model and the log-rank separation of patients split at the median of $U_\alpha$. All model selection uses nested cross-validation, and an external center is held out if available.

Three results would be decisive. The first is whether finite $\lambda$ beats $\lambda=\infty$ at the selected $\alpha$. The second is whether the fitted $\alpha$ is clearly above zero, meaning the tail matters more than the average. The third is the matched near-vessel versus distal-strong comparison at equal composition. Only the full sink model predicts that one. A positive result on all three would show that semi-quantitative IHC discards decision-relevant spatial information in a specific, mechanistic way. A negative result on the first and third would cleanly show that, for anti-HER2 response, how much of each staining level a tumor contains is enough, regardless of layout.