Targeted cancer therapies such as antibody-drug conjugates (ADCs) — antibodies that carry a cell-killing payload to tumor cells bearing a specific surface antigen — fail in an unexpected way: patients whose tumors carry little or none of the target antigen sometimes respond well anyway. The reason is the bystander effect: if the released payload is small and membrane-permeable, it diffuses out of the antigen-positive cell and kills neighboring antigen-negative cells, so response depends not on the average amount of antigen but on the spatial arrangement of the few antigen-positive cells relative to everything else. The phase 2 DAISY trial of trastuzumab deruxtecan (T-DXd) in metastatic breast cancer showed exactly this: patients classified HER2-zero by immunohistochemistry still achieved a confirmed objective response rate of 29.7% (95% CI 15.9–47) (Trastuzumab deruxtecan in metastatic breast cancer with variable HER2 expression: the phase 2 DAISY trial), and, strikingly, within that cohort patients with ERBB2 mRNA below the median responded slightly more often (35.7%) than those above it (30%) — mean antigen level, the quantity every current clinical rule thresholds on, carries no signal in this regime.
This proposal builds an interpretable decision-tree predictor whose split variables are bystander-coverage features: auditable spatial quantities, computed from digitized immunohistochemistry slides, that measure what fraction of tumor cells lies within a payload's plausible diffusion radius of an antigen-positive "hub" cell. The core claim is two-part. First, that a small set of physically motivated coverage features — parameterized by the diffusion radius implied by the payload's permeability class — predicts response to bystander-capable ADCs better than bulk antigen measures, precisely in the low-antigen cohorts where current biomarkers fail. Second, that a decision tree trained on these features, using a teacher-weighting scheme in the spirit of ProfWeight to survive small clinical cohorts, can match black-box spatial models on accuracy while producing rules a pathologist can read and audit. No existing work does this: the DAISY translational analysis used unsupervised black-box clustering on slides and found that a low-staining spatial phenotype ("Cluster 6") tracks resistance (Mechanism of action and resistance to Trastuzumab Deruxtecan in patients with metastatic breast cancer: the DAISY trial), but it extracted no inspectable rule, no explicit diffusion model, and no individual-level predictor; the closest white-box precedent, a CART tree predicting imatinib response in chronic myeloid leukaemia (Modelling Predictors of Molecular Response to Frontline Imatinib for Patients with Chronic Myeloid Leukaemia), operates on non-spatial blood biomarkers for a drug with no bystander mechanism at all.
Core Claim
Response to a bystander-capable ADC is governed by an interpretable geometric quantity — bystander coverage — rather than by mean antigen density, and a decision tree over coverage features plus a small number of resistance markers can formalize this into auditable if-then rules of the form "if the fraction of tumor nuclei within radius $r$ of a HER2-positive hub exceeds $c$, predict response." The pharmacological literature motivates the radius parameter: bystander killing depends on linker cleavability and on the free payload being uncharged and lipophilic enough to cross membranes, as with MMAE released from VC-linked constructs (1183 HX111, a FIC OX40-mAb-VC-MMAE ADC for the treatment of lymphoma), the membrane-permeable topoisomerase-I inhibitor in JSKN003 (A biparatopic HER2-targeting ADC constructed via site-specific glycan conjugation exhibits superior stability, safety, and efficacy), and the A2 payload activating cGAS–STING in bladder cancer (A Novel FGFR3‐Targeting Antibody‐Drug Conjugate Induces Tumor Cell Apoptosis through the cGAS–STING Pathway in Bladder Cancer). The design-rules synthesis in Antibody–drug conjugates in cancer therapy: current landscape, challenges, and future directions confirms these permeability classes are the operative pharmacology; our trees consume them as a calibration of the split geometry rather than as a discarded background fact.
Research Program
The first work package is the feature layer. From digitized whole-slide IHC, we segment nuclei, classify each as antigen-positive or negative, and compute, for a grid of radii, the coverage function
$$C_i(r) = \frac{\text{fraction of tumor nuclei within distance } r \text{ of at least one antigen-positive nucleus in patch } i}{\text{ }}$$
at the patch and patient level. Every downstream split variable is then a plain number a pathologist can verify: patient-level $C(r)$, hub density (antigen-positive nuclei per mm²), the size of the largest antigen-negative gap normalized by $r$, and stroma fraction. We deliberately avoid black-box embeddings at this stage; the features must remain auditable or the interpretability claim is hollow. We complement them with the two non-spatial resistance markers the DAISY translational work identified: SLX4 mutation status (acquired in 20% of resistant samples, 4/20, versus 2% at baseline, 2/88 — evidence that intracellular DNA-repair resistance can negate bystander delivery (Mechanism of action and resistance to Trastuzumab Deruxtecan in patients with metastatic breast cancer: the DAISY trial)) and bulk ERBB2 mRNA.
The second work package is tree induction under small-sample constraints. DAISY's full analysis set is n = 177, far too small for unconstrained greedy search over a radius grid plus biomarkers. We therefore adopt ProfWeight (Improving Simple Models with Confidence Profiles): a frozen black-box teacher (a CNN trained end-to-end on the slides for response) supplies per-sample confidence profiles via linear probes on its hidden layers, and the student CART is trained on samples weighted by teachability. This is the exact mechanism that lifted a raw CART from 74.3% to 87.3% accuracy on high-dimensional semiconductor data and improved small ResNets by 3–4% on CIFAR-10, while standard logit-matching distillation sometimes made students worse than baseline — a critical property when every patient counts. The radius $r$ enters induction either as a hyperparameter swept per payload class (MMAE-like vs DXd-like permeability) or, in a second iteration, via DEFT-style dynamic node features (Interpretable DNA Sequence Classification via Dynamic Feature Generation in Decision Trees), where a language model proposes executable compound features ("fraction of nuclei within 2$r$ of a hub cluster of size ≥ 5") at each node; DEFT reached 0.869 accuracy and 0.912 AUPRC on genomic sequence tasks while remaining interpretable, evidence that generated-but-readable features can close the gap to foundation models. We treat DEFT-style generation as a stretch extension, not the core deliverable, because its per-node cost is high and our primary claim must stand on the fixed, physically motivated feature set.
The third work package addresses the honest bottleneck: spatial IHC plus linked response labels exist for only one published cohort of this size. We will therefore build a calibrated generative simulator of spatial tumor mosaics — antigen-positive and negative nuclei drawn on a point process with controllable clustering — whose forward model maps a coverage map to a probabilistic response using the payload's permeability class and an SLX4-resistance subpopulation. The simulator is calibrated so that its marginal response rates reproduce the three DAISY cohort ORRs (70.6%, 37.5%, 29.7%) and its resistance fraction matches the reported SLX4 dynamics. It serves two purposes: ground-truth recovery of known coverage rules (does the tree recover the radius that generated the data?), and power analysis for the clinical evaluation.
Baselines
The "do nothing" reference point is the current clinical rule itself: binary IHC stratification with median ERBB2 mRNA splitting. Its performance is documented — under it, every IHC 0 patient would have been excluded from a therapy that in fact produced a 29.7% confirmed ORR and 4.2 months median PFS (Trastuzumab deruxtecan in metastatic breast cancer with variable HER2 expression: the phase 2 DAISY trial), and within that cohort the mRNA split is anti-informative (35.7% below-median vs 30% above-median ORR). Any proposed rule must enrich responders within IHC 0/low patients, not merely reproduce cohort-level separation. The strongest existing white-box competitor is a standard CART on non-spatial features — exactly the design that succeeded in CML, where CART Model D reached 72% testing accuracy and 35% specificity on the Saudi external cohort against 6% for Hasford and 13% for Sokal (Modelling Predictors of Molecular Response to Frontline Imatinib for Patients with Chronic Myeloid Leukaemia) — run here on ERBB2 mRNA, SLX4 status, and clinical variables with the same depth, pruning, and tuning budget as our tree. Beating it is the decisive test of the spatial-coverage hypothesis: same learner, same tuning, only the features differ. The strongest overall competitor is the end-to-end teacher CNN, evaluated on its own predictive merit; our claim there is parity-with-auditability, not superiority. Two ablations isolate the ingredients: raw CART (no teacher weighting) on the coverage features — expected to fall short by an amount comparable to the 13-point ProfWeight gap reported on manufacturing data (Improving Simple Models with Confidence Profiles) — and teacher-weighted CART on non-spatial features, which isolates the contribution of the weighting scheme alone. All models are trained and compared on identical folds of the DAISY-linked spatial data, with tuning budgets matched by fixed search grids, evaluated by cross-validated AUC, by the ORR enrichment of the tree-designated responder group against the non-responder group, and by rule stability under bootstrap resampling.
Why not just…?
Why not just keep using IHC cutoffs and median mRNA splits? Because the file's own evidence shows these rules are not merely weak but inverted in the bystander regime: IHC 0 patients responded at 29.7%, and below-median ERBB2 mRNA out-responded above-median (35.7% vs 30%) (Trastuzumab deruxtecan in metastatic breast cancer with variable HER2 expression: the phase 2 DAISY trial). A threshold on the wrong variable cannot be repaired by moving the threshold. Our decisive check is whether $C(r)$ at the patient level separates responders from non-responders within the IHC 0 and IHC low strata where every existing cutoff is uninformative.
Why not just train a black-box CNN end-to-end on the slides? This is the real threat, and we say so plainly: on pure AUC, a CNN may win, and DAISY's own clustering analysis shows slide-level spatial signal exists (Mechanism of action and resistance to Trastuzumab Deruxtecan in patients with metastatic breast cancer: the DAISY trial). Three things temper it. The cohort is small (n = 177 analysis set), the regime where deep nets overfit and where ProfWeight-style guidance was designed to help; black-box pathology models face documented auditability barriers in high-stakes oncology settings (Clinical applications of artificial intelligence in the histopathology of lymphoma: diagnosis, treatment and prognosis); and interpretable trees have already matched foundation models on at least one hard biological benchmark (DEFT, 0.869 accuracy vs deep sequence models). Our experiments settle it directly: the CNN runs as a baseline and as the teacher, so a CNN win tells us how much accuracy auditability costs, and a coverage-tree win tells us the geometry was the signal all along.
Why not just fit off-the-shelf scikit-learn CART to the spatial features? This is included as an ablation rather than argued away, because it is the honest null model for our own method. The evidence says it will underperform: unguided CART sat 13 points below teacher-guided CART (74.3% vs 87.3%) on high-dimensional tabular data (Improving Simple Models with Confidence Profiles), greedy orthogonal splits failed on raw spatially-structured biological inputs where depth-aware feature construction succeeded (Interpretable DNA Sequence Classification via Dynamic Feature Generation in Decision Trees), and a clinical CART left 32% of deteriorating COPD patients undetected (Fast Prediction of Deterioration and Death Risk in Patients With Acute Exacerbation of Chronic Obstructive Pulmonary Disease Using Vital Signs and Admission History: Retrospective Cohort Study). If raw CART on coverage features matches teacher-weighted CART on the DAISY folds, the weighting ingredient is unnecessary and we will report that; the spatial-feature claim survives either way, which is the more important half of the proposal.
Proposed Experiments and Decisive Checks
The clinical evaluation uses the DAISY cohorts as the anchor: within each IHC stratum, we test whether the induced tree's responder-designated group shows a materially higher confirmed ORR than the designated non-responder group, with the pre-registered headline comparison being enrichment within cohorts 2 and 3 (pooled low/zero, n = 114), where existing rules fail. Secondary endpoints are cross-validated AUC and the fraction of IHC 0 patients the rule would have retained as candidates versus zero under current practice. A translational validation uses the RC48 colorectal cancer data (Synergistic antitumor activity between HER2 antibody-drug conjugate and chemotherapy for treating advanced colorectal cancer), where efficacy did not track surface HER2 density across COLO205, P53R, RKO, and HCT116 models (monotherapy IC50 1.5–3.8 μg/mL; PDX tumor growth inhibition up to 92.15%; 74% HER2 positivity among 538 human CRC specimens): we test whether the induced coverage rules, applied to mixed-antigen co-culture spatial measurements, predict which seeding ratios are sterilized by bystander diffusion — the quantitative threshold extraction that study explicitly left qualitative. The untreated HER2-positive spatial transcriptomics cohort (Spatial deconvolution of HER2-positive breast cancer delineates tumor-associated cell type interactions) cannot validate prediction (no outcomes), but serves to check that coverage features generalize across staining platforms and that the feature layer captures the heterogeneity it was built for — 36 sections from 8 patients showing that morphologically uniform tissue hides extensive molecular heterogeneity. The simulator experiments close the loop: trees induced on simulated mosaics must recover the generating radius within the grid resolution and the SLX4-resistance fraction within its calibration interval; failure there falsifies the identification strategy before any clinical claim is made.
Two further checks guard against self-deception. Because T-DXd staining correlated with HER2 IHC only borderline in a tiny biopsy set (Pearson r = 0.75, P = 0.053, n = 7) (Trastuzumab deruxtecan in metastatic breast cancer with variable HER2 expression: the phase 2 DAISY trial), we pre-commit to reporting rule stability under nucleus-classification label noise, since the coverage features inherit any segmentation error. And because CAR-T therapies lacking diffusible payloads face antigen-escape barriers that bystander mechanisms would otherwise solve (Hurdles of CAR-T cell-based cancer immunotherapy directed against solid tumors), we include a negative-control condition: coverage trees fit with the radius shrunk toward zero, where the bystander mechanism is disabled and the predictor should collapse to bulk-antigen behavior. If it does not, the features are picking up something other than diffusion geometry, and the core claim fails. The deliverable is a compact, published rule set — a handful of thresholds on quantities a pathologist can compute from a slide — that turns the bystander effect from a retrospective explanation into a prospective, auditable patient-selection instrument.
References
- Trastuzumab deruxtecan in metastatic breast cancer with variable HER2 expression: the phase 2 DAISY trial
- Mechanism of action and resistance to Trastuzumab Deruxtecan in patients with metastatic breast cancer: the DAISY trial
- Modelling Predictors of Molecular Response to Frontline Imatinib for Patients with Chronic Myeloid Leukaemia
- 1183 HX111, a FIC OX40-mAb-VC-MMAE ADC for the treatment of lymphoma
- A biparatopic HER2-targeting ADC constructed via site-specific glycan conjugation exhibits superior stability, safety, and efficacy
- A Novel FGFR3‐Targeting Antibody‐Drug Conjugate Induces Tumor Cell Apoptosis through the cGAS–STING Pathway in Bladder Cancer
- Antibody–drug conjugates in cancer therapy: current landscape, challenges, and future directions
- Improving Simple Models with Confidence Profiles
- Interpretable DNA Sequence Classification via Dynamic Feature Generation in Decision Trees
- Clinical applications of artificial intelligence in the histopathology of lymphoma: diagnosis, treatment and prognosis
- Fast Prediction of Deterioration and Death Risk in Patients With Acute Exacerbation of Chronic Obstructive Pulmonary Disease Using Vital Signs and Admission History: Retrospective Cohort Study
- Synergistic antitumor activity between HER2 antibody-drug conjugate and chemotherapy for treating advanced colorectal cancer
- Spatial deconvolution of HER2-positive breast cancer delineates tumor-associated cell type interactions
- Hurdles of CAR-T cell-based cancer immunotherapy directed against solid tumors