Liao, Zhongxing

Abstract Background: Immune checkpoint inhibitors (ICIs) have improved survival in non-small-cell lung cancer (NSCLC), yet only a subset of patients benefits, and biomarkers like PD-L1 remain limited. Here, we introduce a deep learning-based pathomics framework that utilizes routine H&E-stained slides to predict therapeutic response and survival outcomes in ICI-treated metastatic NSCLC. Methods: The study included 797 ICI-treated NSCLC patients from MD Anderson, with external validation in 280 patients from Mayo Clinic, Gustave Roussy, and the Phase III ICI-naïve LUSC Lung-MAP S1400I trial receiving nivolumab with or without ipilimumab. Path-IO (Pathology-Driven Immunotherapy Optimization) comprised of four major steps: (1) pathologist-verified tissue classifier segmented WSIs into eight compartments—Background, Bronchi, Immune, Lung, Necrosis, Stroma, Tumor, and Vessel—and was validated on TCGA and CPTAC dataset; (2) a survival prediction module generating patient-level risk scores in the MD Anderson cohort and validated across external datasets; (3) integration of Path-IO risk scores with radiomics and clinical features to improve prognostic accuracy; and (4) biological interpretability analyses correlating Path-IO risk immune contexture from multiplex immunofluorescence and transcriptomic signatures from NanoString profiling. Results: Path-IO effectively stratified patients into high and low-risk groups with significant survival differences. In the MD Anderson cohort, it achieved HR = 2.11 (p < 0.001) and HR = 2.51 (p < 0.001) for OS in the discovery and validation sets, and HR = 2.34 (p < 0.001) and HR = 1.87 (p < 0.001) for PFS, respectively. In the multicenter Phase III Lung-MAP S1400I trial, Path-IO achieved HR = 1.78 (p = 0.016) for OS and HR = 2.76 (p = 0.006) for PFS. In external validation, it predicted outcomes in the Gustave Roussy (OS = 1.97, p = 0.003; PFS = 1.51, p = 0.046) and Mayo Clinic (OS = 2.46, p = 0.007; PFS = 2.45, p = 0.027) cohorts. Path-IO outperformed PD-L1 and remained an independent predictor in multivariate analysis (p < 0.001). It enabled data-driven frontline selection between anti-PD-1 monotherapy and chemo-immunotherapy, offering guidance toward more personalized treatment decisions. Integration with radiomics and clinical features further improved predictive accuracy (OS 0.63→0.75; PFS C-index 0.58→0.70), while high Path-IO risk scores correlated with immune-cold phenotypes identified by multiplex immunofluorescence and NanoString transcriptomics. Conclusions: Path-IO demonstrates that deep learning models rooted in histopathologic architecture can generate interpretable and biologically informed survival predictions in NSCLC treated with ICIs. By integrating pathology, radiology, and clinical data, Path-IO provides complementary predictive value beyond established biomarkers such as PD-L1. Citation Format: Rukhmini Bandyopadhyay, Linghzi Hong, Mihaela Aldea, Shenduo Li, Lodovica Zullo, Frank R. Rojas, Maliazurina B. Saad, Maricel C. Marin, Muhammad Waqas, Jiexin Zhang, Eman Showkatian, Claudio A. Arrechedera, Xiaoyu Han, Yuliya Kitsel, Sherif Ismail, Muhammad Aminu, Bo Zhu, Carol C. Wu, Brett W. Carter, Joe Y Chang, Zhongxing Liao, Maria R. Ghigna, Davide Soldato, Hai T. Tran, Xiuning Le, Tina Cascone, Bingnan Zhang, Haniel A. Araujo, Mehmet Altan, Simon Heeke, David Jaffray, Don L. Gibbons, Ara Vaporciyan, J Jack Lee, Neda Kalhor, Cara Haymaker, Ignacio Wistuba, John V. Heymach, Yanyan Lou, Natalie Vokes, Luisa M. Solis Soto, Jianjun Zhang, Jia Wu. Path-IO: A deep learning pathomics framework for personalized immunotherapy selection and outcome prediction in metastatic non-small cell lung cancer [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 4003.

Abstract The complementarity and clinical utility of combining liquid biopsies and radiomic image analysis has not been demonstrated. ctDNA minimal residual disease after chemoradiotherapy (CRT) for non–small cell lung cancer (NSCLC) is highly prognostic, but on-treatment biomarkers are needed to enable response-adapted therapies. In this study, we analyzed 418 patients with NSCLC undergoing CRT to develop and validate a novel dynamic risk model that accurately predicts ultimate progression-free survival during treatment. We optimize tissue-free variant calling from plasma samples to facilitate ctDNA monitoring and demonstrate the importance of accounting for persistent clonal hematopoiesis variants. We show that mid-CRT ctDNA concentration is prognostic for disease progression and integrate additional pre-CRT risk factors, including radiomics, into a combined model that improves outcome prediction. Our results suggest that tumor features, radiomics, and mid-CRT ctDNA analysis are complementary and can identify patients at high and low risk of progression to potentially enable response-adapted therapies. Significance: This study demonstrates that combining tumor features, radiomics, and ctDNA analysis improves outcome prediction in NSCLC treated with CRT therapy. Our integrated model could enable personalized and response-adapted therapies to reduce toxicity and improve outcomes in patients. See related commentary by Anagnostou and Aggarwal, p. 1534

131 Background: Trimodality therapy is the standard of care for patients with resectable cancer of the esophagus. However, patients ≥75 years have been underrepresented or excluded from landmark clinical trials to date. We investigated the feasibility and safety of treatment of esophageal cancer with trimodality therapy in patients’ ≥ 75 years. Methods: We performed a retrospective review of all patients ≥75 years who received trimodality therapy for esophageal cancer in 3 high volume tertiary cancer institutions from June 2007 to June 2013. All patients received neoadjuvant radiation with concomitant chemotherapy followed by esophagectomy. Toxicities and clinical outcomes were abstracted from the electronic medical record and partially from a prospectively maintained database. Overall and disease-free survival were estimated using the Kaplan-Meier method. Results: Five hundred seventy patients were treated with trimodality therapy for esophageal cancer from 2007-2013. Of these, 38 patients (7%) were 75 or older at the time of diagnosis. At diagnosis, comorbidities included coronary artery disease (32%), atrial fibrillation (11%) and COPD (13%). The majority of patients (87%) received 50.4Gy/28 fractions. 5-fluorouracil (5-FU)/cisplatin was the most common chemotherapy regimen (37%), followed by 5-FU/docetaxel (24%). A total of 13 patients (34%) developed acute grade ≥3 toxicity associated with neoadjuvant therapy. The most common acute grade 3 toxicities were haematological (10%), nausea (8%), esophagitis (5%) and fatigue (5%). Significant postoperative complications included respiratory (empyema, ARDS, pleural effusion) (39%), arrhythmia (32%), anastomotic leak (5%), and ileus (5%). There were 2 deaths (5%) within 90 days of surgery: one was secondary to empyema, the other developed DIC and sepsis. Median overall survival and disease free survival were 4.4 and 2.3 years respectively. Conclusions: Trimodality treatment is a reasonable approach for management of carefully selected elderly patients with esophageal cancer, with similar rates of cancer outcomes, and treatment related morbidity and mortality as compared to younger patients.

Objective: To discern recurrence risk stratification and investigate its influence on postoperative surveillance in patients with esophageal adenocarcinoma (EAC) after neoadjuvant chemoradiotherapy (CRT). Background: Reports documenting recurrence risk stratification in EAC after neoadjuvant CRT are scarce. Methods: Between 1998 and 2014, 601 patients with EAC who underwent neoadjuvant CRT followed by esophagectomy were included for analysis. The pattern, site, timing, and frequency of the first recurrence and potential prognostic factors for developing recurrences were analyzed. This cohort was used as the training set to propose a recurrence risk stratification system, and the stratification was further validated in another cohort of 172 patients. Results: A total of 150 patients (25.0%) achieved pathologic complete response (pCR) after neoadjuvant CRT and the rest were defined as the non-pCR group (n = 451) in the training cohort. After a median follow-up of 63.6 months, the pCR group demonstrated a significantly lower locoregional (4.7% vs 19.1%) and distant recurrence rate (22.0% vs.44.6%) than the non-pCR group (P < 0.001). Based on independent prognostic factors, patients were stratified into 4 recurrence risk categories: pCR with clinical stage I/II, pCR with clinical stage III, non-pCR with pN0, and non-pCR with pN+, with corresponding 5-year recurrence-free survival rates of 88.7%, 65.8%, 55.3%, and 33.0%, respectively (P < 0.001). The risk stratification was reproducible in the validation cohort. Conclusions: We proposed a recurrence risk stratification system for EAC patients based on pathologic response and pretreatment clinical stage. Risk-based postoperative surveillance strategies could be developed for different risk categories.

The objective of this study was to evaluate and understand the systematic error between the planned three‐dimensional (3D) dose and the delivered dose to patient in scanning beam proton therapy for lung tumors. Single‐field and multifield optimized scanning beam proton therapy plans were generated for ten patients with stage II‐III lung cancer with a mix of tumor motion and size. 3D doses in CT datasets for different respiratory phases and the time‐weighted average CT, as well as the four‐dimensional (4D) doses were computed for both plans. The 3D and 4D dose differences for the targets and different organs at risk were compared using dose‐volume histogram (DVH) and voxel‐based techniques, and correlated with the extent of tumor motion. The gross tumor volume (GTV) dose was maintained in all 3D and 4D doses, using the internal GTV override technique. The DVH and voxel‐based techniques are highly correlated. The mean dose error and the standard deviation of dose error for all target volumes were both less than 1.5% for all but one patient. However, the point dose difference between the 3D and 4D doses was up to 6% for the GTV and greater than 10% for the clinical and planning target volumes. Changes in the 4D and 3D doses were not correlated with tumor motion. The planning technique (single‐field or multifield optimized) did not affect the observed systematic error. In conclusion, the dose error in 3D dose calculation varies from patient to patient and does not correlate with lung tumor motion. Therefore, patient‐specific evaluation of the 4D dose is important for scanning beam proton therapy for lung tumors.PACS number: 87.55.D