Synthetic image samples accross models

Independent classifier validation and representative bidirectional counterfactual transformations in a multiclass setting. A, External classifier responses to counterfactual morphing. Counterfactual images were generated at increasing morphing amplitudes (α) with MoPaDi and then encoded with three foundation models (UNI2, CONCH, and Virchow2). Independent classifiers were trained on the corresponding encoders’ features extracted from all TCGA-CRC tiles and then used to predict the target probability P(target) on both the original and counterfactual tiles. The resulting change ΔP(target) reflects how strongly the morphing affected class evidence. Bars show the median ΔP(target) with percentile-based variability across tiles. B, Representative examples of bidirectional counterfactual explanations for MSIL patients. We defined MSIL by fitting a two-component Gaussian mixture model to the log-transformed distribution of total MSI events and using the intersection of the two components as the cutoff separating MSIL from MSIH samples.

Abstract Deep learning can extract predictive and prognostic biomarkers from histopathology whole-slide images. However, explainable artificial intelligence approaches widely used in digital pathology, such as attention heatmaps and class activation mapping, provide limited insight into the image features associated with classifier outputs. In this study, we developed Morphing histoPathology Diffusion (MoPaDi), a framework for generating counterfactual explanations for histopathology images that help identify morphologic or stain-related features linked to model predictions. MoPaDi combined diffusion autoencoders with task-specific multiple instance learning classifiers to manipulate images and induce prediction shifts by modifying classifier-associated features. The framework was evaluated on multiple datasets spanning colorectal, breast, liver, and lung cancers, including tasks for tissue type, cancer subtype, and biomarker [microsatellite instability (MSI)] classification. MoPaDi generated perceptually realistic counterfactual histopathology images, enabling pathologists to identify morphologic features associated with changes in model predictions, complementing the conventional inspection of highly attended regions in digital pathology. In the MSI status prediction task, MoPaDi highlighted morphologic features linked to classifier predictions, including mucinous differentiation, altered glandular architecture, and lymphocytic infiltration, consistent with prior literature. Analyses separating stain-related from morphology-related components suggested that in this setting, prediction changes were predominantly associated with morphology-related rather than stain-related alterations. Overall, MoPaDi is a practical framework for counterfactual explanations in computational pathology that supports the evaluation of model-specific decision cues and hypothesis generation. Significance: MoPaDi is a diffusion-based tool for counterfactual image generation in cancer histopathology that reveals features associated with deep learning classifier predictions and supports transparent auditing of computational models in biomedical research.

Tissue-type counterfactual examples for additional classes

Style–morphology decomposition for disentangling structural and staining effects on MSI prediction changes. A, Representative examples of counterfactual manipulation between MSIH and non-MSIH classes. For each original image x and its counterfactual xcf, style-hybrid (xstyle) and morphology-hybrid (xmorph) images were generated using Vahadane stain transfer. Each hybrid isolates the effect of either stain or morphology while controlling for the other. The right-hand bars show the Shapley-style decomposition of the logit change (Δf) into stain (φstyle) and morphology (φmorph) contributions, demonstrating that morphologic differences dominate the model’s predictions. Grad-CAM visualizations below provide region-level attribution under MIL. In contrast, MoPaDi produces class-directed “what-if” edits that offer a complementary view of candidate morphologic and style changes associated with prediction shifts. Scale bar applies to all images within the panel. B, Decomposition results across test-set patients, showing median contributions of φstyle, φmorph, and total (Δf) for manipulations toward (↑) and away from (↓) each class. C, Scatter plot of morphology versus style contributions per patient, illustrating consistent dominance of morphologic effects across both manipulation directions and classes.

Immune checkpoint inhibitors (ICIs), a class of immunotherapy, offer promising benefits but face challenges such as low response rates to be used as a broadly effective treatment for all patients. In this study, we use a set of ordinary different equation (ODE) models and bladder cancer in vivo data as a case study to outline a biologically informed, data-driven framework for formulating, calibrating and validating immunotherapy models, and thus ensuring their predictive reliability. We consider multiple treatment scenarios and distinct immune cell-mediated killing mechanisms for tumor cells of different antigenicity. By integrating sensitivity analysis and identifiability analysis with targeted experimental design, we demonstrate how mathematical models can move beyond qualitative insight to quantitative prediction. We generate virtual cohorts to show that insufficient data integration leads to systematically overestimated therapeutic benefits of ICIs. We also explore dosing schedules that enhance survival or reduce dosage without compromising survival. ### Competing Interest Statement The authors have declared no competing interest. NIH/NCI, U01CA243075, K08, CA234392

Counterfactual image examples generated for the lung and breast cancer–type classifiers. A, Representative examples of LUSC tile transitioning to its counterfactual lung adenocarcinoma (LUAD) image and vice versa. B, Morphologic feature prevalence in original and counterfactual image pairs (N = 32 transitions; 16 tiles for each class). Horizontal bars show the percentage of image pairs in which at least one of three raters (three board-certified pathologists) identified each morphologic feature as present in the original (dark gray) or counterfactual (light gray) image. Right, Mean pairwise inter-rater agreement (Cohen κ) per feature, computed across all pairs and both directions combined. C, Counterfactual image generation effectiveness, measured as the percentage of generated images predicted as the opposite class across varying manipulation amplitudes. D, Representative examples of ILC tile transitioning to its counterfactual IDC image and vice versa. Difference maps display pixelwise differences between the original and the synthetic tile. Scale bar applies to all images within the panel unless otherwise indicated.

6075 Background: The rising incidence of oropharyngeal squamous cell carcinoma (OPSCC) is largely attributable to human papillomavirus associated (HPV+) disease, which accounts for ~70% of OPSCC cases. Circulating tumor (ct)DNA has the potential to enable more accurate treatment response assessment, guide response-adaptive management, and detect minimal residual disease to indicate persistence or recurrence. Both mutation-based tumor-informed ctDNA and ctHPV-DNA testing have demonstrated utility in HPV+ disease, but prospective intrapatient evaluations remain limited. A direct comparison of these approaches is essential to determine redundancy versus complementarity and to guide optimal integration into OPSCC patient management. Methods: In an ongoing prospective study, serial plasma samples were obtained from patients with stages I-IV OPSCC undergoing curative intent treatment. Up to 50 patient specific somatic variants were selected based on tumor whole exome sequencing to develop a personalized tumor-informed next generation sequencing (NGS) ctDNA assay (Haystack MRD) for plasma analysis. In patients with HPV+ disease (determined via ISH, IHC, and/or NGS), plasma was also analyzed using an NGS-based assay interrogating 13 high-risk HPV strains (Haystack HPV). Paired intrapatient samples were analyzed using percent agreement with 95% confidence intervals and Cohen’s kappa; concordance of dynamic changes was assessed using Spearman’s correlation. Results: As of January 2026, ctDNA results were available for 111 serial timepoints from 26 patients. The median number of timepoints per patient was 4 (range 1-9). Seventeen patients (65%) had HPV+, and 9 (35%) had HPV− disease. In HPV+ patients, across 85 longitudinal samples collected during multimodal treatment and post-treatment surveillance, mutation-based ctDNA and ctHPV demonstrated high concordance (91%; 95% CI, 82.5–95.2; κ=0.80). Of 30 ctDNA+ samples, 28 were ctHPV+ (93%; 95% CI, 78.7–98.2%), while 49 of 55 ctDNA- samples were ctHPV- (89%; 95% CI, 78.2–94.9%). Discordance was infrequent (8/85, 9.4%), predominantly ctHPV+/ctDNA- (6/85, 7.1%). All ctHPV+/ctDNA- cases occurred during neoadjuvant treatment monitoring and reflected earlier clearance of ctDNA, with ctHPV clearance lagging by several weeks to months. Two low-level (<100 parts per million) ctDNA+/ctHPV- cases were observed in the adjuvant setting. When both analytes were present, dynamic changes in ctDNA and ctHPV levels were highly concordant (Spearman’s ρ=0.94), although ctHPV was consistently detected at higher absolute levels. Conclusions: In HPV-driven OPSCC, tumor-informed ctDNA and ctHPV show high longitudinal concordance and distinct clearance kinetics, with earlier ctDNA clearance. Ongoing analyses will define how these assays can be optimally integrated into response assessment, treatment adaptation, and surveillance strategies.

BACKGROUND:Identifying oral potentially malignant disorders and oral cavity cancer early can lead to better patient outcomes. The guideline panel evaluated the usefulness of light-based adjuncts for screening adults without mucosal abnormalities and for determining the need for biopsy among adults with mucosal abnormalities in the oral cavity or on the lip. TYPES OF STUDIES REVIEWED:The authors conducted a living systematic review to evaluate evidence on the benefits and harms of light-based adjuncts and a scoping review to assess people and clinician values and preferences regarding the use of light-based adjuncts and biopsy of mucosal abnormalities. The guideline panel used this evidence to formulate recommendations according to the Grading of Recommendations Assessment, Development and Evaluation Evidence to Decision framework. The framework also guided the panel's consideration of required resources, equity, acceptability, and feasibility in shaping the final recommendations. RESULTS:The guideline panel formulated 2 recommendations and 2 good practice statements. For adults with and without mucosal abnormalities, they formulated conditional recommendations against the use of light-based adjuncts on the basis of very low certainty evidence. The good practice statements urge clinicians to perform a clinical oral examination in all adult patients. CONCLUSIONS AND PRACTICAL IMPLICATIONS:Biopsy remains the reference standard for establishing a definitive diagnosis of an oral potentially malignant disorder and oral squamous cell carcinoma. All adults should undergo a clinical oral examination in primary care settings. When implementing or adapting these recommendations, local contexts should be considered to promote equitable access to early detection.

Preclinical studies have evaluated murine double minue 2 (MDM2) inhibitors as a treatment for adenoid cystic carcinoma (ACC), but clinical trials are lacking. This phase I trial (NCT03781986) assesses the safety and antitumor activity of an oral MDM2 inhibitor, alrizomadlin (APG-115), +/- carboplatin in TP53 wild type unresectable recurrent/metastatic salivary gland cancers (R/M SGC) with a planned 1:1 randomization to carboplatin chemotherapy. The co-primary endpoints are determination of dose-limiting toxicity (DLT) and response rate (RR) for alrizomadlin monotherapy +/- carboplatin. Secondary endpoints include safety, survival, and RR by tumor histology. After enrollment of 4 patients to combination therapy, the trial was modified to a single arm study of alrizomadlin monotherapy due to excess toxicity. 1 DLT was seen in the combination arm, all patients had ≥ G3 treatment related adverse events (TRAE). 37 patients were enrolled to alrizomadlin monotherapy. 3 DLTs were encountered, 67% of patients had ≥ G3 TRAE. The RR was 15% with median progression free survival 10.5 months. These findings demonstrate encouraging tolerability of alrizomadlin monotherapy with antitumor activity in patients with TP53 wild type SGC, especially ACC.

Purpose:To develop and validate a multimodal recurrence-risk model integrating histology, genomic testing, and clinical variables. Methods:We developed AI-Path, a whole-slide image biomarker for recurrence prediction trained in CALGB 9344, and validated it in three independent cohorts: TAILORx, a multi-site Chicago cohort, and the MDX-BRCA cohort. We then integrated AI-Path with Oncotype DX Recurrence Score (RS), tumor size, and nodal status into a Cox model, PathClinRS, fit using 60% of cases from TAILORx, with the remaining 40% held out for validation. The primary end point was distant recurrence-free interval. Performance was assessed using Harrell's concordance index (C-index) and Kaplan-Meier analyses. Results:A total of 12,418 patients were included. In TAILORx, AI-Path outperformed RS for distant recurrence (C-index, 0.682 vs 0.647; P = .038), driven by superior prediction of late recurrence (0.656 vs 0.567; P < .001). In node-negative disease, PathClinRS outperformed RSClin in the TAILORx fitting (0.72 vs 0.70; P = .016) and validation sets (0.74 vs 0.70; P = .004). In node-positive disease, PathClinRS outperformed RSClinN+ in Chicago (0.94 vs 0.74; P < .001) and MDX-BRCA (0.71 vs 0.66; P = .004) cohorts. Compared with NATALEE eligibility, PathClinRS identified nearly twice as many high-risk node-negative patients while maintaining a comparable 10-year distant recurrence risk (16.7% vs 16.6% per NATALEE eligibility in TAILORx fitting; 21.0% vs 19.4% in TAILORx validation). PathClinRS identified 68% of intermediate risk premenopausal patients as low-risk with no evidence of chemotherapy benefit, compared to only 36% identified as low risk by standard clinicopathologic criteria. Conclusion:Digital histopathology provides prognostic information complementary to genomic assays and has the potential to personalize therapy beyond existing clinicogenomic tools.

ABSTRACT Emerging evidence indicates that a subset of cancer cells enriched for stemness-related gene signatures possess distinct immunomodulatory capacities, enabling these tumor-initiating stem cells (tSCs) to more effectively evade or resist anti-tumor immunity. Despite these advances, the tSC-specific molecular circuits orchestrating their specialized immune privilege program are not well defined. Here, in squamous cell carcinomas of the skin and oral cavity, we comprehensively delineate the unique immune-evasive properties of tSCs and dissect the transcriptional regulation shaping their immunomodulatory programs. By integrating transcriptome profiling, chromatin landscape mapping, genetic perturbation, and single-cell RNA sequencing, we found that the tSC-specific immune program is broadly governed by SOX2, a stemness-associated transcription factor. We demonstrate that SOX2 enables tSCs to sustain immature tumor-associated neutrophils (TANs) and subsequently trigger these myeloid cells to foster the development of tumor-associated macrophages (TAMs). This SOX2-directed tSC-TAN-TAM axis establishes a localized immunosuppressive niche for protecting tSC. SIGNIFICANCE Here, we uncover SOX2 as a master regulator that orchestrates conserved immune modulatory circuits in tSCs to sustain pro-tumor myeloid cell states. These findings place tSCs at the apex of immune landscape remodeling, asserting a central role of stemness-associated program in organizing the immunosuppressive tumor microenvironment.

Medical foundation models compress biomedical data into embeddings that support diverse downstream clinical tasks. However, successful model deployment is hampered by performance degradation on external data. It is recognized that embeddings capture acquisition signatures, such as hardware and technical differences, in addition to biology. Effective harmonization must remove the acquisition signature while preserving biological signals, a trade-off that current methods fail to balance adequately. Input-level normalization fails to eliminate acquisition signatures from embeddings, whereas embedding-level methods adjust features in an untargeted manner. We present FEATMAP, a harmonization approach that models acquisition signatures as geometric distortions between manifolds of similarly arranged embeddings. Using paired data that isolate the effect of acquisition signatures, FEATMAP fits a single global affine transformation per foundation model to correct acquisition signatures directly in the embedding space. This targeted, reusable correction aims to preserve biological and demographic variation while harmonizing across acquisition signatures. Across scanner and foundation-model harmonization in digital pathology and field-strength harmonization in brain MRI, FEATMAP improves cross-condition embedding similarity, reduces performance gaps without retraining, and suggests potential for the alignment of disparate embedding spaces.

MIL transformer architecture for patient-level classification

6012 Background: A subset of salivary gland cancers (SGCs) express the androgen receptor (AR). A prior AR+ SGC trial with the anti-androgen enzalutamide alone failed to meet its primary endpoint. This is a multicenter, phase II study evaluating the efficacy and safety of the anti-androgen darolutamide (Bayer) in combination with androgen-deprivation therapy (ADT; leuprolide acetate) in hormone-therapy naïve patients with AR+ SGCs. Methods: Patients with locally advanced/unresectable or recurrent/metastatic AR+ SGCs were enrolled. AR status was determined locally by immunohistochemistry (IHC). Prior AR-targeted therapy was not allowed, unless administered in the neoadjuvant and/or adjuvant setting >6 months before disease recurrence. Darolutamide 600 mg orally twice daily was given with leuprolide acetate intramuscular injections (1 cycle= 28 days). The primary endpoint was best overall response (BOR) rate according to RECIST v1.1 within 1 year of initiating treatment. Secondary endpoints were progression-free survival (PFS), overall survival (OS), and toxicity. Exploratory endpoints were evaluating biomarkers in serially obtained research biopsies and exploring efficacy among patients who had not received prior systemic therapy. A two-stage minimax design was used to detect a 50% BOR rate (vs. 25%) (alpha = 9%; beta = 83%). ≥3 responses in the first 9 patients would trigger accrual to 20; ≥8 responses would be considered promising. Results: Study accrual of 20 patients (pts) with AR+ SGCs was completed on 10/30/25. 15 males, 5 females with a median age of 70.5 years were enrolled. Among the 9 pts in the first stage, 6 confirmed partial responses (PRs) were observed, allowing for full study accrual. With a data cutoff of 1/21/26, the best RECIST v1.1 responses among 20 pts were 8 (40% [19.1% 63.9%]) PRs (7 confirmed, 1 unconfirmed with pt still on treatment), 10 (50% [27.2%, 72.8%]) stable disease (SD), 1 (5% [1%, 24.9%]) progression of disease; 1 pt on treatment has not had radiographic assessments yet. 6 (30% [11.9%, 54.3%]) pts came off trial for disease progression, 1 (5% [1%, 24.9%]) withdrew consent after 6 cycles, and 13 (65%, [40.8%, 84.6%]) remain on treatment. Among 5 female pts, best responses were 1 (5% [1%, 24.9%]) PR, 3 (15% [3.2%, 37.9%]) SD, 1 (5% [1%, 24.9%]) PD. With additional follow-up, PFS, OS, and biomarker data (AR IHC %, HER2 status) will be presented. Conclusions: In this molecularly selected cohort of patients with SGC, darolutamide plus ADT possesses significant clinical activity, validating AR as a relevant therapeutic target in a subset of SGCs. Analysis of serial research biopsies will be performed to identify biomarker and/or drug combination strategies to enhance the efficacy of AR-targeting. Clinical trial information: NCT05669664 .

6097 Background: Neoadjuvant immunotherapy is an emerging strategy in head and neck squamous cell carcinoma to enhance systemic antitumor immunity and enable response-adapted de-escalation. In HPV associated oropharyngeal squamous cell carcinoma (OPSCC), virally encoded oncoproteins represent shared, tumor-specific antigens and a rational immunologic target well suited for the neoadjuvant setting in the presence of intact tumor antigen. We conducted a phase I/II trial evaluating neoadjuvant HPV16-specific viral immunotherapy (HB200; HB201 and HB202 HPV16 therapeutic vaccines) plus chemotherapy followed by response-adapted definitive treatment in non-metastatic HPV16+ OPSCC (NCT05108870). Methods: This investigator-initiated phase I/II trial enrolled patients with previously untreated, non-metastatic HPV16+ OPSCC (N1-3 or T3-4; smokers permitted). All patients received three cycles of neoadjuvant HB200 (HB201 alone or alternating HB202/201) with carboplatin/paclitaxel, followed by radiographic response assessment. Patients with T1-2 tonsil or well-lateralized base of tongue tumors achieving ≥50% tumor shrinkage underwent transoral robotic surgery (TORS) alone. Remaining patients received response and risk adapted radiotherapy (50-70Gy based on risk/response) with or without cisplatin. The primary endpoint was deep response rate (DRR; ≥50% tumor shrinkage). Secondary endpoints included survival and toxicity. Exploratory endpoints included circulating tumor HPV-DNA (ctHPV-DNA), HPV16-specific immunity, and spatial transcriptomics. Results: Thirty-five patients were enrolled (median age 58; 89% male); Twelve patients (34%) received HB201 alone and 23 (66%) received alternating HB202/201. Nineteen patients (54%) were current or former smokers, and 49% had stage II-III (AJCC 8 th edition). The DRR was 87.9% (95% CI, 71.8-96.6). Thirty (86%) received de-escalated definitive therapy. At a median follow-up of 23 months, 2-year PFS and OS were 86% and 100% respectively. Most common AEs during neoadjuvant HB200/chemo were fatigue (97%), nausea (91%), and fever (76%). Detectable ctHPV-DNA following treatment was significantly associated with disease recurrence ( p <0.01). HPV16-specific immune responses and spatial transcriptomic analyses will be presented. Conclusions: Neoadjuvant HB200 combined with chemotherapy resulted in high deep response rates, frequent treatment de-escalation, and excellent survival outcomes in locoregionally advanced HPV16+ OPSCC. These findings support further evaluation of HPV directed immune therapy in neoadjuvant setting. Clinical trial information: NCT05108870 .

Counterfactual results for the lung cancer type classifier

Abstract Background: Despite advances in immunotherapy, median survival for advanced bladder cancer remains under 3 years, and mechanisms governing PD-1/PD-L1 blockade efficacy remain ill defined. Activation of fibroblast growth factor receptor-3 (FGFR3) has been shown to be linked to a non-T cell-inflamed tumor microenvironment (TME) and with resistance to checkpoint blockade. Using human bladder tumors and murine models, we investigated mediators of immunotherapy response and how FGFR3 shapes the immune landscape affecting PD-1/PD-L1 treatment efficacy. Methods: Multiplex immunofluorescence (mIF) and RNAscope were performed on 47 human bladder tumors to assess immune infiltrates and FGFR3 expression. A second cohort of 21 bladder tumors collected before anti-PD-1/PD-L1 therapy was analyzed by mIF and spatial transcriptomics (ST). CD8+ T cell and DC1 clustering was evaluated using an unbiased computational analysis (K-cross, a modified Ripley’s K function). ST was used to investigate the effect of FGFR3 expression on immune cells spatial distribution and transcripts. In vivo, subcutaneous and orthotopic MB49 models expressing FGFR3 G370C activating mutation, FGFR3 kinase-dead mutant (K508M), or control vector were used with/without anti-PD-L1. FTY720 was used to assess the requirement for new T cell entry into the TME for anti-PD-L1 efficacy. Tumor growth and immune phenotypes were analyzed by spectral flow cytometry. Results: FGFR3-activated tumors showed reduced CD86+ DC1s in tumor-draining lymph nodes (tdLN), impaired CD8+ T cell activation, greater exhaustion, and fewer TCF-1+ progenitor-exhausted CD8+ T cell (TPE) in the TME after anti-PD-L1 treatment. Blockade of new T cell entry to the TME abrogated the anti-PD-L1-induced accumulation of TPE in control tumors. FGFR3-kinase-dead mutation resulted in higher TPE frequencies in tumors, tdLN, and spleen, and reduced CD8+ T cell exhaustion. In human tumors, clinical response to PD-1/PD-L1 blockade was associated with CD8+/DC1 clustering rather than absolute cell numbers. FGFR3 expression inversely correlated with CD8+ and DC1 abundance and co-localization. Within tumors, FGFR3+ regions showed fewer DC1s, CD8+ T cells, reduced co-localization, and diminished expression of chemokines that recruit these cells. CD8+ T cells in FGFR3+ areas exhibited lower activation/effector and TPE gene expression and higher exhaustion/inhibitory-related transcripts, compared to those in FGFR3 negative areas. Conclusions: Our results indicate that DC1/CD8+ T cell clustering is essential for PD-1/PD-L1 efficacy in bladder cancer, and FGFR3 activation drives resistance by disrupting this interaction, preventing CD8+ T cells from acquiring phenotypes required for effective immunotherapy. Citation Format: Andrea Ziblat, Ken Hatogai, Anthony A. Fernald, Danny E. Kim, Hyunsik Lee, Alexander T. Pearson, Madeleine S. Torcasso, Randy F. Sweis. FGFR3 impairs DC1s/CD8+ T cell clustering in bladder cancer preventing the PD-1/PD-L1 blockade-induced CD8+ T cell phenotypes associated with immunotherapy efficacy [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 160.