Measuring How Strongly Tissue Boundaries Block Signal, Calibrated on Speech

Tumors that resist immunotherapy often do so by keeping immune cells on the wrong side of a border. Killer T cells gather in the surrounding connective tissue, called stroma, but never reach the cancer cells. Current spatial analyses describe this border crudely. They either draw a fixed 100 µm band around each tumor nest, or they smooth gene expression evenly across neighboring measurement spots. Neither approach asks the question that matters: how much does a particular boundary actually cut off molecular and cellular exchange between its two sides? We propose to estimate a per-boundary "resistance" score. It is the drop in expression similarity between two points that sit on opposite sides of an interface, relative to two points the same distance apart on the same side. We then test whether high-resistance tumor–stroma interfaces are the ones that exclude T cells.

The difficulty is validation. No tissue dataset tells us how strong a boundary "really" is. Speech does. In speech science, "prosody" means the rhythm and melody of spoken language, and phonetics has long measured boundary strength directly. Major phrase boundaries block the blending of neighboring sounds (coarticulation) more than minor ones, as the Italian ultrasound study shows. The Boston University Radio News Corpus (BURNC), about 158 minutes of read news from six speakers, is annotated with graded break indices under ToBI, a standard hand-labeling scheme for pitch and phrasing. These indices are human judgments of boundary strength on a scale from no break to a full intonational phrase break. BURNC therefore gives us a real, noisy, naturally produced signal with gold hierarchical boundary labels and realistic alignment errors. We use it as a calibration bench for the resistance estimator before applying that estimator to tumors. In this proposal, speech supplies the validation regime and the measurement concept. The tumor microenvironment supplies the scientific question.

Core Claim

Our claim is that a single estimator of boundary resistance can serve two purposes. On BURNC, its output should rank ToBI break indices correctly. In spatial transcriptomics of tumors, the same estimator should identify the tumor–stroma interfaces associated with immune exclusion, better than fixed-distance bands. It should also improve fine-grained domain segmentation when it replaces isotropic neighbor smoothing, meaning smoothing that treats every direction and every neighbor the same.

Define $s(x,y)$ as the similarity between two locations: correlation of acoustic feature vectors in speech, or of denoised expression profiles in tissue. Define $\bar s_{\text{within}}(d)$ and $\bar s_{\text{across}}(d)$ as the mean similarity of pairs at distance $d$ on the same side of boundary $b$ and on opposite sides of it. The resistance of $b$ is then

$$R(b) = 1 - \frac{\int w(d)\,\bar s_{\text{across}}(d)\,dd}{\int w(d)\,\bar s_{\text{within}}(d)\,dd},$$

where $w(d)$ is a distance weighting whose width is learned rather than fixed. $R$ near 0 means the boundary blocks nothing, and $R$ near 1 means the two sides do not communicate. Learning the width of $w$ follows directly from Detecting Pitch Accents at the Word, Syllable and Vowel Level. That paper found that acoustic cues spill beyond the syllable. Word-level models reached 84.2% accuracy against 68.5% at the vowel level, and adding context lifted vowel-level accuracy to 77.4%. The right context span is therefore an empirical quantity, and a fixed band like the one used in A Novel Framework for Characterization of Tumor-Immune Spatial Relationships in Tumor Microenvironment assumes it away.

Research Program

The work has three stages. In stage one, we build the estimator and calibrate it on BURNC. In speech, the locations are word-aligned frames, and the features are the pitch, energy and duration measures that Rosenberg and Hirschberg found informative. Boundaries are ToBI break sites. Speech has its own version of a known tissue confound. Pitch tracking and forced alignment smear information across boundaries, which inflates across-boundary similarity, just as spot bleeding does in spatial transcriptomics. We therefore build a correction for measurement smear into the estimator. It is modeled on the directional bleeding correction in BayesTME: A unified statistical framework for spatial transcriptomics, which fits a hierarchical Bayesian model to undo expression leaking between neighboring spots. We then test whether $R$ recovers the ordering of break indices. Coarticulation across prosodic domains in Italian: an ultrasound investigation is a caution here. Only two of its three speakers showed the expected major-versus-minor difference, so the estimator must report per-speaker spread rather than a single pooled score.

In stage two, we transfer the frozen estimator design to tissue. Locations become spots or cells. Features are denoised expression from DUSTED: Dual-Attention Enhanced Spatial Transcriptomics Denoiser, a graph attention autoencoder that recovers dropout-corrupted counts. Boundaries are candidate interfaces from existing domain calls. We use resistance in two ways. First, it serves as a descriptive per-interface measurement. Second, it replaces the isotropic neighbor-reassignment step in DisConST: Deciphering Spatial Domains Using Distribution-aware Contrastive Learning for Spatial Transcriptomics. That clustering method models counts with zero-inflated negative binomial distributions and uses graph contrastive learning, and its authors note that its smoothing step fails on fine-grained structures. In our version, neighbors across a high-resistance edge are not smoothed together.

In stage three, we ask the biological question. Does interface resistance predict immune exclusion? The motivating biology comes from several papers. Human Hyaluronidase PH20 Potentiates the Antitumor Activities of Mesothelin-Specific CAR-T Cells Against Gastric Cancer shows that a hyaluronic-acid-rich matrix physically limits T-cell entry, and that degrading it restores infiltration. The MYBL2–CCL2 axis promotes tumor progression and resistance to anti-PD-1 therapy in ovarian cancer by inducing immunosuppressive macrophages shows a diffusible chemokine recruiting suppressive macrophages. Single-nucleus and spatial transcriptomics of archival pancreatic cancer reveals multi-compartment reprogramming after neoadjuvant treatment links classical-like tumor states to macrophage-rich, immune-excluded microniches. These papers generate a falsifiable prediction. Resistance of the expression profile across an interface should rise with stromal matrix gene expression, and it should correlate negatively with CD8 T-cell density on the tumor side. Chemokine signals such as CCL2 may cross boundaries even when T cells do not, so a gene-program-specific resistance is also worth reporting.

Baselines

The naive reference in tissue is the fixed 100 µm isotropic band from the pancreatic multiplex-imaging study. That study reported that CD8 and CD163 distributions deviate from randomness (median absolute deviation 1.36 and 0.99). Our baseline predicts CD8 exclusion from the density of immune cells and the stromal signal inside that band, using the same interfaces, the same cell calls, and the same cross-validated regression as our model. Resistance beats it only if it adds explained variance in tumor-side CD8 density beyond band density under held-out-slide evaluation. For interface-level effects, the confidence interval must exclude zero.

The strongest existing method for the segmentation use is DisConST. On the 12-slice DLPFC 10X Visium benchmark, it reports an average adjusted Rand index of 0.62 and a peak of 0.86 (DisConST). DUSTED reports a median of 0.515. We rerun DisConST from its released design with identical preprocessing, seeds, and hyperparameter budget. The only change is swapping its neighbor refinement for resistance-weighted refinement. Beating it means a higher mean ARI across the 12 slices, with gains concentrated in the thin cortical layers. Gains confined to thick layers would suggest generic regularization rather than boundary modeling.

In speech, the naive reference is a raw pause-duration threshold for break strength. Ivan Chow identifies silent pauses as the primary juncture cue. A stronger reference is a supervised logistic regression on word-level pitch, energy and duration features in the Rosenberg–Hirschberg style, trained to predict break index. Lacuna has no reported result for break-index prediction on BURNC, so we report our own run of that baseline. We do not use the 84.2% figure, which measures pitch accents. Resistance is unsupervised, so success means matching the supervised baseline's rank correlation with break index within its cross-speaker variance.

Two ablations isolate the key ingredients. The first removes the smear correction, which tests whether bleeding alone creates apparent low resistance. The second fixes the width of $w(d)$ instead of learning it, which tests the context-span claim.

Why not just…?

Why not just calibrate on simulated tissue instead of speech? This is the real threat, and we treat it as a competing arm, not a strawman. Simulations with planted boundary strengths are cheap and match tissue geometry. Their weakness is that the noise follows the simulator's own assumptions. BURNC offers independently produced human-graded boundaries with real alignment and tracking errors. We will tune the estimator on BURNC only, on simulation only, and on both, then compare downstream DLPFC ARI and immune-exclusion prediction. If simulation-only tuning is as good, BURNC's role reduces to a robustness check. We will report that plainly.

Why not just compare expression between DisConST or BayesTME domains? Mean differences between domains measure how different two regions are, not how much their boundary blocks exchange. Two domains can differ sharply yet share a gradient across their interface, which BayesTME's spatial programs in fact show within cell types. The distance-matched within-versus-across comparison separates these two cases. The domain-difference score will be run as a covariate to show whether resistance carries separate information.

Why not just use bulk deconvolution like CIBERSORT? The glioblastoma study MTAP loss correlates with an immunosuppressive profile in GBM and its substrate MTA stimulates alternative macrophage polarization found M2 macrophage enrichment with bulk data (p = 0.0064). However, bulk profiles cannot tell present-but-excluded T cells from absent ones, which is precisely the failure mode the hyaluronidase study targets. We will include slide-level pseudo-bulk deconvolution as a predictor so we can check whether it already explains exclusion.

Proposed Experiments

The first decisive check runs on BURNC. Across the six speakers, the smear-corrected resistance should rise monotonically with break index, measured as Spearman correlation per speaker. If it fails to separate minor from major breaks for most speakers, we stop and revise before touching tissue.

The second check runs on DLPFC. Resistance-weighted refinement must raise DisConST's average ARI above 0.62, with the gain concentrated in the thin layers.

The third check uses the ovarian cancer array analyzed by DUSTED, the melanoma arrays from BayesTME, and the 23-slide pancreatic imaging cohort. Tumor–stroma interface resistance should correlate positively with matrix and hyaluronan-synthesis programs and negatively with tumor-side CD8 density, beyond the fixed-band baseline, under leave-one-slide-out evaluation.

Finally, as a transfer test on the archival pancreatic data, interfaces in classical-like regions should show higher resistance than those in basal-like regions. That comparison is observational and cannot separate plasticity from clonal selection, as the source paper itself notes.