Human Epidermal Growth Factor Receptor 2 Quantification Using Computational Pathology to Identify Novel Biomarkers for Trastuzumab Deruxtecan–Treated Human Epidermal Growth Factor Receptor 2–Positive Gastric Cancer Authors: Ansh Kapil, Henrik Failmezger, Yu Fu, Andreas Spitzmüller, Jessica Chan, Susanne Haneder, Anatoliy Shumilov, Alice Barkell, Toshihiro Oguma, Koichiro Inaki, Fumitaka Suto, John Allard, Jeeyun Lee, Arthur W Lambert, Markus Schick, J Carl Barrett, Günter Schmidt, Hadassah Sade, Mark Gustavson, Danielle Carroll, Fabiola Cecchi Published: 2026 ABSTRACT AbstractPURPOSETrastuzumab deruxtecan (T-DXd) is an antibody-drug conjugate approved for human epidermal growth factor receptor 2 (HER2)–positive gastric cancer (GC) and gastroesophageal junction adenocarcinoma (GEJA). Eligibility for T-DXd therapy currently relies on conventional HER2 immunohistochemistry (IHC) scoring per ASCO/CAP guidelines. Computational pathology tools such as quantitative continuous scoring (QCS) offer precise quantification of HER2 expression, enabling improved scores reflective of T-DXd's mechanism of action.METHODSWe used HER2-QCS scoring to identify biomarker signatures in the DESTINY-Gastric01 cohort, with external validation in DESTINY-Gastric02.RESULTSQCS-derived continuous spatial proximity score (cSPS) and density of HER2-positive cells in tumor epithelium (DPC) displayed significant survival benefits for biomarker-positive versus biomarker-negative patients in the DESTINY-Gastric01 T-DXd arm (Cox regression: P < .0001). These signatures were validated in the DESTINY-Gastric02 cohort (log-rank P < .0001).CONCLUSIONOur findings indicate that cSPS and DPC are predictive of T-DXd clinical outcomes, supporting the potential future use of QCS as a quantitative approach to identify patients with GC who might benefit most from T-DXd therapy. INTRODUCTION Human epidermal growth factor receptor 2 (HER2) overexpression occurs in 10%-20% of gastric cancers (GCs). Recently, the outcome of HER2-positive (HER2+) breast cancer defined by an immunohistochemistry (IHC) score of 3+, or an IHC score of 2+ and HER2 gene amplification by (fluorescent) in situ hybridization ((F)ISH) testing, has been significantly improved since the approval of several anti-HER2 therapies including antibodies directed at HER2 receptors (eg, trastuzumab, pertuzumab, and margetuximab), antibody-drug conjugates (ADCs; eg, ado-trastuzumab emtansine [T-DM1]), and small molecule HER2 receptor tyrosine kinase inhibitors (eg, lapatinib).1 More recently, the HER2-targeted ADC trastuzumab deruxtecan (T-DXd; fam-T-DXd-nxki) has been approved for adults with previously treated, locally advanced or metastatic HER2-positive gastric or gastroesophageal junction adenocarcinoma2. T-DXd is composed of a humanized anti-HER2 antibody, an enzymatically cleavable peptide linker, and a novel topoisomerase I inhibitor (DXd).3-5 Due to the membrane-permeable nature of DXd, molecules of DXd that are released into the extracellular space by the dead cells can also enter and kill neighboring cancer cells. The capacity of T-DXd to kill surrounding cells regardless of their HER2 expression has been described as the bystander effect.6 This effect potentially contributes to the high antitumor activity of T-DXd not only in patients with high HER2 expression but also some with low levels of expression. The observed antitumor activity of T-DXd in patients with low levels of HER2 expression urges the need of accurate scoring of the HER2-low population.7 Manual scoring of IHC images by pathologists has demonstrated clinical utility in identifying patients with HER2+ breast cancer8 and GC.9 However, current scoring guidelines have been developed to detect HER2+ overexpressing patients, and scoring in the lower end might be challenging.8 CONTEXT Key Objective Key Objective Develop a reproducible, quantitative method to measure human epidermal growth factor receptor 2 (HER2) membrane staining intensity (SI) in gastric cancer. Develop a reproducible, quantitative method to measure human epidermal growth factor receptor 2 (HER2) membrane staining intensity (SI) in gastric cancer. Knowledge Generated Knowledge Generated We developed the artificial intelligence–based quantitative continuous scoring (QCS) framework and derived readouts that quantify membrane SI and spatial patterns independent of current manual HER2 scoring. These QCS metrics were applied in the DG01 clinical study and independently validated in DG02, demonstrating robustness and clinical relevance. We developed the artificial intelligence–based quantitative continuous scoring (QCS) framework and derived readouts that quantify membrane SI and spatial patterns independent of current manual HER2 scoring. These QCS metrics were applied in the DG01 clinical study and independently validated in DG02, demonstrating robustness and clinical relevance. Relevance Relevance Artificial intelligence–based quantitative assessment of membrane marker staining can enhance pharmacodynamic understanding and mechanistic exploration of antibody-drug conjugates across tumor indications, supporting biomarker-driven development and response prediction. Artificial intelligence–based quantitative assessment of membrane marker staining can enhance pharmacodynamic understanding and mechanistic exploration of antibody-drug conjugates across tumor indications, supporting biomarker-driven development and response prediction. Computational pathology, particularly deep learning–based digital image analysis, can quantify staining intensity (SI) distributions and intratumoral heterogeneity beyond human reproducibility, enabling novel biomarker definitions.10 For therapeutics such as T-DXd, which include bystander killing potential, the ability of computational pathology to define novel scoring strategies holds promise. This led us to develop a computational pathology solution, the quantitative continuous scoring (QCS), as a new device with diagnostic potential. In this study, exploration of the potential clinical value of HER2-QCS biomarkers (BMs) in GC was performed using data and samples from the DESTINY-Gastric 01 (DG-01) clinical trial.2 DG-01 was a randomized, Phase 2 trial involving patients with HER2-expressing, locally advanced or metastatic gastric or gastroesophageal junction cancer. HER2 levels were documented as high (score of 3+ on immunohistochemical analysis or score of 2+ FISH+) in the primary cohort or low (score of 2+ FISH–) in an exploratory cohort. Here, exploratory QCS analysis of the DG-01 primary cohort suggested that various predictive BMs defined by HER2-QCS have the potential to enrich patient population with greater benefit from T-DXd treatment compared with standard-of-care treatment. These BMs using QCS were also shown to be prognostic for response in DESTINY-Gastric 02 (DG-02)11, an independent cohort of patients with HER2+ GC. METHODS Data Sets DESTINY-Gastric 01 (DG-01, NCT03329690) is an open-label, randomized, Phase 2 trial that evaluated T-DXd as compared with chemotherapy in patients with HER2+ advanced GC.2 The cohort analyzed by QCS comprised 123 patients with HER2+ GC and 20 patients with HER2– GC, all treated with T-DXd. The control cohort consisted of 59 HER2+ patients treated with chemotherapy. DESTINY-Gastric 02 (DG-02, NCT04014075) is an open-label, randomized, Phase 2 trial that evaluates T-DXd in patients who have progressed on or after a trastuzumab-containing regimen.11 The DG-02 cohort analyzed by QCS comprised 72 patients who were all treated with T-DXd. Image Analysis Pipeline The DG-01 (N = 202) and DG-02 (N = 72) GC tissue samples were stained using conventional HER2 Ventana 4B5 assay. Slides were centrally scored by pathologists using recent ASCO/CAP HER2 Testing Guidelines. The slides were digitized using an Aperio scanner at 20× magnification and underwent a strict quality check. Digitized slides were processed with HER2-QCS, a fully automated pipeline that detects invasive tumor regions (ART) and segments nuclei, cytoplasm, and membranes (subcellular segmentation of tumor cell), generating sample-level quantitative scores. Pathologists visually inspected image analysis outputs and adjusted tumor areas, if needed. See Supplementary Methods for details. RESULTS HER2-QCS Signature Discovery With DG-01 QCS image analysis was performed and IHC SI-based measurements, such as the mean SI per subcellular compartment, were obtained (Fig 1A), resulting in center coordinates and HER2 expression values for the tumor cells in the tissue (see Methods).10 HER2 expression heterogeneity in GC. (A) Image analysis and bioinformatics pipeline to calculate QCS readouts. (B) Intratumoral HER2 heterogeneity as determined by QCS. Tissue samples with unimodal versus bimodal HER2 expression. (C) HER2 median membrane SI per sample. (D) HER2 intratumoral heterogeneity as determined by the median absolute deviation of the HER2 SI distribution. Outliers are marked in red circles. ART, automated region segmentation of tumor; GC, gastric cancer; HER2, human epidermal growth factor receptor 2; PR, partial response; QCS, quantitative continuous scoring; RECIST, Response Evaluation Criteria in Solid Tumors; SD, stable disease; SI, staining intensity; SOC, standard of care; SSTC, subcellular segmentation of tumor cells; T-DXd, trastuzumab deruxtecan. The median membrane SI as measured by QCS provides a quantitative measurement of HER2 expression in the sample. We found a high variability of sample-wise median HER2 expression values in the QCS analysis set of the DG-01 (T-DXd: median 66, IQR, 80; SOC: median 71, IQR, 76, Fig 1C). Furthermore, we found intratumoral variabilities of 22.7 (T-DXd) and 22.0 (SOC), measured as the median absolute deviation in HER2 expression per sample (Figs 1B and 1D). We compared median SI by sample type and found a trend toward higher values in biopsies (T-DXd: P = .07; SOC: P = .024; Data Supplement, Fig S1), possibly reflecting greater sampling heterogeneity in resections. Median SI did not vary meaningfully by anatomic site, although interpretation is limited by few lymph node and liver specimens (Data Supplement, Fig S2). Correlation analysis of objective response rate (ORR) in DG-01 against the median SI across all tumor cells per sample showed a significant difference in HER2 expression between responders and nonresponders in patients treated with T-DXd (Wilcoxon test P = .00017; Figs 2A and 2C). Furthermore, a strong continuous association between ORR and increased HER2 expression across the entire range of the assay was observed (Pearson correlation r = 0.87; Fig 2B), in that each incremental increase in SI results in an increase in response rate. However, despite the statistical significance, there is still a large overlap in HER2 expression between responders and nonresponders in patients treated with T-DXd (Fig 2A). Association of HER2 median membrane SI with response and HER2 IHC status. (A) HER2 median membrane SI for responders and nonresponders. Response is defined as partial or complete response. (B) Relationship of HER2 median membrane SI with the ORR in the positive group. The blue line depicts the corresponding regression line. The QCS positive group at a given SI value was defined as every patient with a median membrane SI greater than that value. Correlation is given by Pearson. (C) Association of HER2 median membrane SI with response. (D) Association of HER2 median membrane SI with IHC status. CR, complete response; HER2, human epidermal growth factor receptor 2; IHC, immunohistochemistry; ISH, in situ hybridization; NE, not evaluable; ORR, objective response rate; PD, progressive disease; PR, partial response; QCS, quantitative continuous scoring; RECIST, Response Evaluation Criteria in Solid Tumors; SD, stable disease; SI, staining intensity; SOC, standard of care; T-DXd, trastuzumab deruxtecan. We examined the correlation of HER2 IHC status as determined by pathologist scoring against the median SI across all cells per sample (Fig 2D). We noticed a trend toward higher median SI values with increasing IHC HER2 status but still observed an overlap between both categories, which could be attributed to the HER2 pathologist scoring algorithm itself and the lack of its granularity and ability to provide continuous scores. We evaluated whether IHC 3+ patients with relatively low membrane SI exhibit greater heterogeneity in membrane SI. To quantify heterogeneity, we computed the bimodal separation coefficient12 (Data Supplement, Fig S3A), an index of distribution bimodality. Among IHC 3+ patients, lower median SI was associated with higher bimodal separation coefficients (Spearman correlation ρ = −0.62; Data Supplement, Fig S3B), indicating that reduced median SI corresponds to more pronounced bimodality. In the analyzed cohort of DG-01, patients had a median progression-free survival (mPFS) of 5.55 months with T-DXd and 2.8 months with SOC (Fig 3A). In the T-DXd cohort, HER2 IHC 3+ patients had a mPFS of 5.8 months and an ORR of 46.8% versus 4.3 months and 20.7%, respectively, for HER2 2+ patients (Fig 3B). QCS signature discovery in DG-01. (A) PFS in treatment and SOC arms. (B) PFS based on stratification by HER2 IHC scoring. (C) Association of the cSPS with PFS. Stratification is based on the cutoff that results in the best interaction P value in the Cox regression. (D) Association of DPC with PFS. (E, F) Subgroup comparisons. All forest plots are based on the Cox regression with interaction term. (G, H) Overlap of patient stratification based on the cSPS and the DPC signature with HER2 IHC scoring. The dashed lines indicate the cut-offs that binarize the data into BM+ and BM–. BM, biomarker; cSPS, continuous spatial proximity score; DPC, density of HER2+ cells in the epithelium; HR, hazard ratio; HER2, human epidermal growth factor receptor 2; IHC, immunohistochemistry; ORR, objective response rate; PFS, progression-free survival; SOC, standard of care; T-DXd, trastuzumab deruxtecan. We aimed to determine a QCS-based biomarker that could better stratify patients than the conventional IHC HER2 scoring using PFS as an end point. Based on SI-derived cellular target expression measurements, we calculated a range of features from seven different feature classes that capture distinct aspects of tumor biology and serve as the basis for later biomarker identification. Features were (1) the quantiles of the SI distribution, (2) the ratio between cellular expression in membrane and cytoplasm, (3) the percentage of positive cells, (4) the continuous spatial proximity scores, (5) the binary spatial proximity scores, (6) the density of SI-positive tumor cells in tumor epithelium, and (7) the shape of the SI distribution (see Supplementary Methods).10 Each of those feature classes comes with a specific parameter set. Evaluation of all parameters for all features resulted in a set of 278 readouts per whole-slide image, although many of them were highly correlated. After readout consolidation (see Supplementary Methods), a set of 98 unique readouts remained for further analysis. To stratify patients into BM+ and BM– subgroups, a cutoff was needed to dichotomize the continuous readouts. For each feature, the optimal combination of parameters and cutoff was selected based on the predictive effect in DG-01 as determined by the most significant treatment-by-biomarker interaction in a Cox proportional hazards model. We found three scores with an interaction P value <.0001 (Data Supplement, Table S1). Two scores were density scores for HER2+ cells in the tumor epithelium (DPC), with SI positivity threshold of 40 and 50. Cutoffs for these features, binarizing patients into BM+ and BM−, were identified at DPC values of 4,779.4615 and 5,310.9171, respectively. The third score was the so-called continuous Spatial Proximity Score (cSPS)10 that was designed to model the bystander activity of T-DXd drugs (see the Data Supplement, Methods). The best configuration for this score was found at the 90% quantile and using a distance radius of 75 μm (see Supplementary Methods) and a cutoff at 3,104.4097. Although both DPC scores shared a large Spearman correlation (ρ = 0.983), they were less correlated with the cSPS (ρ ≤ 0.87, Data Supplement, Fig S4). Due to the large correlation of DPC (SI 50) with DPC (SI 40), which indicates a redundant feature, we excluded DPC (SI 40) from additional analysis and only investigated DPC (SI 50), and cSPS further. QCS-derived cSPS scores showed significant survival benefit for biomarker-positive (BM+) versus biomarker-negative (BM–) in the DG-01 T-DXd arm (P < .0001; hazard ratio [HR], 0.37 [0.24, 0.58], Fig 3C) and longer mPFS for BM+ (8.3 months, ORR 52%) v BM− (3.9 months, ORR 24%). Within the SOC arm, this signature showed the inverse prognosis (P = .034; HR, 2.17 [1.06, 4.44]; mPFS 2.8 months, ORR 9.1% for BM+ v 4.9 months, ORR 11.5% for BM−, Fig 3E), strongly suggesting the signature is predictive of T-DXd response. Furthermore, QCS-derived density of HER2+ cells (SI 50) in the epithelium (DPC) showed significant survival benefit for BM+ versus BM− in the DG-01 T-DXd arm (P < .0001; HR, 0.28 [0.18, 0.45], Fig 3D) and longer mPFS and ORR for BM+ (9.8 months, ORR 57.4%) versus BM− (3.65 months, ORR, 20%). Similar to the cSPS signature, the density score showed the inverse prognosis (P = .065; HR, 2.03 [0.96, 4.29]; mPFS 2.8 months, ORR 14.7% for BM+ v 4.9 months, ORR 4% for BM−, Fig 3F) within the SOC arm, again suggesting the signature is predictive of T-DXd response. For both cSPS and DPC, the majority of patients in the BM+ group were IHC 3+ (cSPS: 91.8%, DPC: 92.6%, Figs 3G and 3H), whereas in the biomarker-negative group, IHC3+ patients were less frequent (cSPS: 54%, DPC: 56.4%, Figs 3G and 3H). We tested if the two signatures add additional information to IHC scoring in terms of survival prediction. Therefore, we first applied Cox regression for IHC scoring and then added the signature and tested by using the likelihood ratio (LR) test whether the model including the QCS signature predicts PFS significantly better. This was the case for both signatures (cSPS: P = .001957, DPC: P = .0004163, Data Supplement, Fig S5), whereas a combination of both signatures did not improve the model (Data Supplement, Fig S5). To assess the reliability of the signatures, we visualized their interaction P value distribution for different stratification cutoffs and parameters (Data Supplement, Fig S6). Both signatures showed a high stability around the selected cutoff and parameters. We selected the cSPS score and the DPC as candidate signatures for further verification in the independent DG-02 data set. HER2-QCS Signature Validation in DG-02 We validated the two candidate signatures in the QCS analysis set of the DG-02 cohort. The patients in this cohort showed a median PFS of 6.4 months (Fig 4A). HER2 IHC 3+ patients had a significant survival benefit in this cohort (P = .0007, mPFS IHC 3+, 7.1 months, mPFS IHC 2+, 1.6 months, HR, 0.26, Fig 4B). Again, we observed an association between median membrane SI, as measured by QCS, and response. The lowest 25% of patients according to their median SI showed an ORR of 28%, whereas the highest 25% of patients showed a response rate of 72% (Fig 4C). Furthermore, we observed a continuous association between median SI and overall response (Pearson correlation r = 0.97, Fig 4D). Verification of QCS signatures in DG-02. (A) Progression-free survival. (B) Patient stratification based on HER2 IHC scoring. (C) Association of HER2 median membrane SI with patient response. ORRs are given per quartile. (D) Relationship of HER2 median membrane SI with ORR in the positive group. The blue line depicts the corresponding regression line. The QCS-positive group at a given SI value is defined as every patient with a median membrane SI greater than that value. (E) Patient stratification based on the cSPS signature. (F) Patient stratification by DPC. (G, H) Overlap of patient stratification based on the cSPS and the DPC signature with HER2 IHC scoring. The dashed lines indicate the cutoffs that binarize the data into BM+ and BM–. BM, biomarker; cSPS, continuous spatial proximity score; DPC, density of HER2+ cells in the epithelium; HER2, human epidermal growth factor receptor 2; HR, hazard ratio; IHC, immunohistochemistry; mPFS, median progression-free survival; ORR, objective response rate; QCS, quantitative continuous scoring; SI, staining intensity. Better patient stratification was observed for both candidate signatures compared with IHC manual scoring for PFS in DG-02. Patients from BM+ group based on cSPS signature had a median PFS of 9.7 months for BM+ patients versus 2.8 months for BM– patients (HR, 0.23; CI [0.13 to 0.44]; P < .0001; Fig 4E). Similarly, patients from BM+ group based on DPC signature had a median PFS of 9.7 months for BM+ patients versus 2.8 months for BM– patients (HR, 0.23 [0.12, 0.44]; P < .0001; Fig 4F). However, patients from IHC 3+ group had a median PFS of 7.1 months versus 1.6 month for IHC 2+ group (HR, 0.26; CI [0.11 to 0.6]; P = .0007; Fig 4B). Patients from the BM+ group based on the cSPS signature had an ORR of 52.2% versus 23.1% in the BM– group. Patients in the BM+ group based on the DPC signature had an ORR of 53.8% versus 27.3% in the BM– group. A total of 37 patients were identified as BM+ based on both signatures. Most BM+ patients were scored as IHC 3+ (cSPS: 45/46, DPC: 39/39, for one patient IHC was not centrally evaluated), whereas the BM– group was found to be a mixture of IHC 2+ and IHC 3+ patients (Figs 4G and 4H). We have observed that both signatures added additional information to IHC scoring in terms of survival prediction. The combined model of IHC scores and the respective QCS signature significantly improved survival prediction in comparison with IHC scoring alone (cSPS: P = .0054, DPC: P = .0008526, Data Supplement, Fig S7). DISCUSSION HER2-directed therapies are effective in multiple cancer indications as demonstrated by T-DXd in pan-cancer trials.5 Current semiquantitative measurements of HER2 levels by IHC are not adequate to accurately quantify HER2 levels and particularly tumor heterogeneity. HER2 overexpression in GCs is well known to be heterogeneous, and scoring guidelines reflect this complexity. Computational pathology approaches allow quantitation of HER2 across a spectrum of expression levels and quantitation of tumor heterogeneity more precisely than pathologist visual scoring. In this study, we show the improved prediction of response using a computation pathology approach termed QCS and its potential for exploratory and scientific applications. QCS-based BMs extend the current standard scoring concepts that rely on estimating the percentage of positive cells by (1) providing a drug efficacy–driven definition of cellular positivity and (2) adding a spatial dimension using the spatial proximity scores. The latter enables the assessment of intratumoral heterogeneity of HER2 expression (Fig 1) and the inclusion of a potential bystander activity in the scoring algorithm. Due to their quantitative nature, the two detected QCS signatures, cSPS as well as DPC, allow a more specific stratification of patients. This is apparent when comparing the QCS results with IHC scoring. Both QCS signatures outperform manual scoring in terms of median PFS (Table 1). In DG-01, the patients in the BM+ groups had an mPFS of 8.3 (cSPS) and 9.8 (DPC) months compared with an mPFS of 5.8 months in IHC 3+ patients. Similarly in DG-02, BM+ patients of both QCS signatures had an mPFS of 9.7 months compared with an mPFS of 7.1 months in patients who were scored IHC 3+. In the BM+ groups, most patients were IHC 3+, whereas in the BM– groups, a mixture IHC 3+ and IHC 2+ patients was observed, again supporting the assumption that the QCS signatures select patients in a more specific manner. TABLE 1.Results of Selected HER2 QCS Candidates in DG-01 and DG-02Analysis Population and End PointDESTINY-Gastric01DESTINY-Gastric02T-DXdSOCT-DXdOverall cohort, No.1235972 mPFS (months)5.552.86.4Stratification according to IHC IHC 3+, No.944663 IHC 3+, mPFS (months)5.82.87.1 IHC 3+, ORR (%)44/94 (46.8)4/46 (8.7)29/63 (46) IHC 2+, No.29138 IHC 2+, mPFS (months)4.34.91.6 IHC 2+, ORR (%)6/29 (20.7)2/13 (15.4)1/8 (12.5) Log-rank P value: IHC3+ v IHC2+0.00410.490.00072 IHC3+ v IHC2+ HR (95% CI)0.49 (0.3 to 0.81)1.35 (0.56 to 3.27)0.26 (0.11 to 0.60)Stratification according to cSPS (BM+ v BM– status) BM+, No.733346 BM+, mPFS (months)8.32.89.7 BM+, ORR (%)38/73 (52.1)3/33 (9.1)24/46 (52.2) BM–, No.502626 BM–, mPFS (months)3.94.92.8 BM–, ORR12/50 (24)3/26 (11.5)6/26 (23.1) P value: BM+ v BM– Based on Cox regression with interaction term1.9E-05.034 Log-rank P value1e-05.047.89E-07 BM+ v BM– HR (95% CI)0.37 (0.24 to 0.58)2.17 (1.06 to 4.44)0.23 (0.13 to 0.44)Stratification according to density of positive cells (BM+ v BM– status) BM+, No.683439 BM+, mPFS (months)9.82.89.7 BM+, ORR (%)39/68 (57.4)5/34 (14.7)21/39 (53.8) BM–, No.552533 BM–, mPFS (months)3.654.92.8 BM–, ORR (%)11/55 (20)1/25 (4)9/33 (27.3) P value: BM+ v BM– Based on Cox regression with interaction term1.5E-07.065 Log-rank P value4e-08.081.8e-06 BM+ v BM– HR (95% CI)0.28 (0.18 to 0.45)2.03 (0.96 to 4.29)0.23 (0.12 to 0.44)NOTE. P values and HR for IHC are restricted to IHC 2+ and IHC 3+ cases only. HRs for QCS candidates in DG-01 were estimated using a Cox proportional hazards model with an interaction term.Abbreviations: BM, biomarker; cSPS, continuous spatial proximity score; HER2, human epidermal growth factor receptor 2; HR, hazard ratio; IHC, immunohistochemistry; mPFS, median progression-free survival; ORR, objective response rate; QCS, quantitative continuous scoring; T-DXd, trastuzumab deruxtecan. Results of Selected HER2 QCS Candidates in DG-01 and DG-02 NOTE. P values and HR for IHC are restricted to IHC 2+ and IHC 3+ cases only. HRs for QCS candidates in DG-01 were estimated using a Cox proportional hazards model with an interaction term. Abbreviations: BM, biomarker; cSPS, continuous spatial proximity score; HER2, human epidermal growth factor receptor 2; HR, hazard ratio; IHC, immunohistochemistry; mPFS, median progression-free survival; ORR, objective response rate; QCS, quantitative continuous scoring; T-DXd, trastuzumab deruxtecan. In both cohorts, the prevalence of IHC 3+ patients is higher (DG-01: 76.42%; DG-02: 87.5%) than the prevalence of BM+ patients (cSPS: DG-01: BM+ 59.35%; DG-02: 63.89%; DPC: DG-01: 55.28% DG-02: 54.17%). In combination with the longer median PFS times that we observe for BM+ patients, this suggests that QCS subselects HER2 3+ patients with long-term response. We assume that in this setting, QCS rather quantifies the extent of HER2 expression, whereas traditional manual IHC scoring emphasizes intensity. Although a manual approach that quantifies extent (eg, an H-score) could be informative for predicting T-DXd efficacy, the inherent limitations of manual scoring remain: lower reproducibility and the lack of spatial quantification. Our data suggest that spatial information—although it seems not to be dominant in this indication—adds measurable value and could be even more important for other targets or disease settings. We envision that this method could be further extended to other tumor types where there is no approved HER2-directed therapy, with the aim to help extend the patient population benefiting from T-DXd treatment. More generally, QCS-based analysis can be applied to any digitized IHC slide stained with a membrane-specific marker. Deployment of such a digital pathology system to the clinical practice will require the standardization of the tissue preparation protocols (eg, section thickness), IHC staining, and slide digitization. For the DG-01 and DG-02 analyses, all samples were processed in a central laboratory and hence followed the same protocol. However, in a real-world setting with decentralized testing, section thickness and differences in staining protocols might be an issue. Additionally, the present HER2-QCS software program is in development and currently in a research-use-only setting. The HER2-QCS software program would need to be upgraded to software program as a medical device standard or medical product standard before clinics would be able to deploy it. Similar algorithms could be developed for patients with HER2-low GC. We checked the performance of QCS in a small HER2-low cohort (2+ ISH–) of DG-01 (see the Data Supplement, Methods). Although patient numbers were limited (approximately 20 patients), we found QCS signatures that have a continuous association with PFS by Cox regression. Identification of patients in the HER2-low group who could benefit from T-DXd is critical for its use in a patient population with a high unmet need, who would otherwise not be treated with anti-HER2 therapy. Moreover, we could challenge any ISH reflex testing, as QCS could stratify patients based only on their digitized HER2 IHC. Computational pathology methods, such as QCS, may help us to better predict response to ADC therapies and identify potential patients who could receive the most benefit. ACKNOWLEDGMENT The authors thank the patients and their families for their participation and the study site staff for their contributions to the studies. The authors retained full control of the content and made the final decisions for all aspects of this manuscript. Ansh Kapil Employment: AstraZeneca Stock and Other Ownership Interests: AstraZeneca Patents, Royalties, Other Intellectual Property: I am on patents originating from AstraZeneca. I receive no royalties Travel, Accommodations, Expenses: AstraZeneca Henrik Failmezger Employment: AstraZeneca Computational Pathology GmbH Yu Fu Employment: Pierre Fabre Andreas Spitzmüller Employment: AstraZeneca Patents, Royalties, Other Intellectual Property: Pending and provisional patents developed at AstraZeneca Travel, Accommodations, Expenses: AstraZeneca Susanne Haneder Employment: AstraZeneca Stock and Other Ownership Interests: Daiichi Sankyo Anatoliy Shumilov Employment: AstraZeneca Alice Barkell Employment: AstraZeneca Stock and Other Ownership Interests: AstraZeneca Koichiro Inaki Employment: Daiichi Sankyo Co, Ltd Fumitaka Suto Employment: Daiichi Sankyo Stock and Other Ownership Interests: Daiichi Sankyo Travel, Accommodations, Expenses: Daiichi Sankyo John Allard Employment: Daiichi Sankyo Inc Stock and Other Ownership Interests: Daiichi Sankyo Research Funding: Daiichi Sankyo Jeeyun Lee Consulting or Advisory Role: AstraZeneca, Trutino Biosciences, Amgen, Daiichi Sankyo, ImmuneOncia, BMS, AbbVie, Samsung Bioepis (Inst), Samsung Biologics (Inst), Samsung Electronics (Inst) Research Funding: AstraZeneca, Merck Sharp & Dohme, Samsung Bioepis, Samsung Electronics Arthur W. Lambert Employment: AstraZeneca Stock and Other Ownership Interests: AstraZeneca Markus Schick Employment: AstraZeneca J. Carl Barrett Employment: AstraZeneca, Precede Bio Leadership: SAGA Diagnostics, Corista Stock and Other Ownership Interests: AstraZeneca, Precede Bio, Corista, SAGA Diagnostics Consulting or Advisory Role: ExAI, Bio-AI Health, Akoya Biosciences, Leica Biosystems, Agilent, Tristar Technologies, Bain Capital, Zephyr AI, Quanterix, Nexosomes, BPGbio Günter Schmidt Employment: AstraZeneca Stock and Other Ownership Interests: AstraZeneca, Merck KGaA Patents, Royalties, Other Intellectual Property: Patents owned by AstraZeneca, related to computational pathology–based biomarkers Hadassah Sade Employment: AstraZeneca Stock and Other Ownership Interests: AstraZeneca Danielle Carroll Employment: AstraZeneca Stock and Other Ownership Interests: AstraZeneca Travel, Accommodations, Expenses: AstraZeneca Fabiola Cecchi Stock and Other Ownership Interests: AstraZeneca/MedImmune No other potential conflicts of interest were reported. PRIOR PRESENTATION Presented in part as a poster at the ASCO Gastrointestinal Cancers Symposium, San Francisco, CA, January 19-21, 2023. SUPPORT Supported by AstraZeneca Pharmaceuticals. In March 2019, AstraZeneca entered into a global development and commercialization collaboration agreement with Daiichi Sankyo for trastuzumab deruxtecan (T-DXd; DS-8201). CLINICAL TRIAL INFORMATION NCT03329690, NCT04014075 A.K. and H.F. contributed equally to this work. DATA AVAILABILITY STATEMENT The data sets generated and/or analyzed during the current study are not publicly available owing to ongoing work on the data analysis. Please contact Giulia Fabbri (giulia.fabbri@astrazeneca.com) or Max Qian (max.qian@daiichisankyo.com) for any data requests relating to this study. Source code is proprietary information of AstraZeneca and will not be made publicly available. AUTHOR CONTRIBUTIONS Conception and design: Ansh Kapil, Henrik Failmezger, Andreas Spitzmüller, Susanne Haneder, Anatoliy Shumilov, Jeeyun Lee, Markus Schick, J. Carl Barrett, Günter Schmidt, Hadassah Sade, Mark Gustavson, Danielle Carroll, Fabiola Cecchi Administrative support: J. Carl Barrett Collection and assembly of data: Ansh Kapil, Andreas Spitzmüller, Jessica Chan, Susanne Haneder, Anatoliy Shumilov, Alice Barkell, Toshihiro Oguma, Koichiro Inaki, Fumitaka Suto, John Allard, Jeeyun Lee, Markus Schick, J. Carl Barrett, Danielle Carroll Data analysis and interpretation: Ansh Kapil, Henrik Failmezger, Yu Fu, Andreas Spitzmüller, Susanne Haneder, Anatoliy Shumilov, Arthur W. Lambert, Markus Schick, J. Carl Barrett, Mark Gustavson, Danielle Carroll, Fabiola Cecchi Manuscript writing: All authors Final approval of manuscript: All authors Accountable for all aspects of the work: All authors AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/po/author-center. Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments). Ansh Kapil Employment: AstraZeneca Stock and Other Ownership Interests: AstraZeneca Patents, Royalties, Other Intellectual Property: I am on patents originating from AstraZeneca. I receive no royalties Travel, Accommodations, Expenses: AstraZeneca Henrik Failmezger Employment: AstraZeneca Computational Pathology GmbH Yu Fu Employment: Pierre Fabre Andreas Spitzmüller Employment: AstraZeneca Patents, Royalties, Other Intellectual Property: Pending and provisional patents developed at AstraZeneca Travel, Accommodations, Expenses: AstraZeneca Susanne Haneder Employment: AstraZeneca Stock and Other Ownership Interests: Daiichi Sankyo Anatoliy Shumilov Employment: AstraZeneca Alice Barkell Employment: AstraZeneca Stock and Other Ownership Interests: AstraZeneca Koichiro Inaki Employment: Daiichi Sankyo Co, Ltd Fumitaka Suto Employment: Daiichi Sankyo Stock and Other Ownership Interests: Daiichi Sankyo Travel, Accommodations, Expenses: Daiichi Sankyo John Allard Employment: Daiichi Sankyo Inc Stock and Other Ownership Interests: Daiichi Sankyo Research Funding: Daiichi Sankyo Jeeyun Lee Consulting or Advisory Role: AstraZeneca, Trutino Biosciences, Amgen, Daiichi Sankyo, ImmuneOncia, BMS, AbbVie, Samsung Bioepis (Inst), Samsung Biologics (Inst), Samsung Electronics (Inst) Research Funding: AstraZeneca, Merck Sharp & Dohme, Samsung Bioepis, Samsung Electronics Arthur W. Lambert Employment: AstraZeneca Stock and Other Ownership Interests: AstraZeneca Markus Schick Employment: AstraZeneca J. Carl Barrett Employment: AstraZeneca, Precede Bio Leadership: SAGA Diagnostics, Corista Stock and Other Ownership Interests: AstraZeneca, Precede Bio, Corista, SAGA Diagnostics Consulting or Advisory Role: ExAI, Bio-AI Health, Akoya Biosciences, Leica Biosystems, Agilent, Tristar Technologies, Bain Capital, Zephyr AI, Quanterix, Nexosomes, BPGbio Günter Schmidt Employment: AstraZeneca Stock and Other Ownership Interests: AstraZeneca, Merck KGaA Patents, Royalties, Other Intellectual Property: Patents owned by AstraZeneca, related to computational pathology–based biomarkers Hadassah Sade Employment: AstraZeneca Stock and Other Ownership Interests: AstraZeneca Danielle Carroll Employment: AstraZeneca Stock and Other Ownership Interests: AstraZeneca Travel, Accommodations, Expenses: AstraZeneca Fabiola Cecchi Stock and Other Ownership Interests: AstraZeneca/MedImmune No other potential conflicts of interest were reported. REFERENCES Associated Data Data Availability Statement The data sets generated and/or analyzed during the current study are not publicly available owing to ongoing work on the data analysis. Please contact Giulia Fabbri (giulia.fabbri@astrazeneca.com) or Max Qian (max.qian@daiichisankyo.com) for any data requests relating to this study. Source code is proprietary information of AstraZeneca and will not be made publicly available.