Abstract Background: Spatial cellular ecosystems in the tumor microenvironment (TME) form dynamic signaling hubs that critically influence cancer disease progression and response to therapy, including response to immune checkpoint inhibitors (ICIs). However, tumor sampling bias and the impracticality of acquiring serial tumor biopsies have made it challenging to profile the TME clinically. To address this challenge, we developed Liquid EcoTyper, a multi-analyte AI framework to quantify TME spatial cellular ecosystems—termed spatial ecotypes (SEs) —noninvasively from cell-free DNA (cfDNA) methylation data. Methods: In previous work, we discovered nine spatial ecotypes, including SEs associated with ICI response (e. g. , SE7 with benefit, SE4 with resistance), and quantified them from 1, 249 bulk tumor RNA-seq profiles of melanoma, non-small cell lung cancer (NSCLC), and bladder cancer (BC) patients (Cancer Res (2025) 85 (8Supplement₁): 153). Here we trained Liquid EcoTyper, an interpretable deep learning model, to transfer SE profiling to cell-free DNA using CpG methylation profiles. Performance was quantified using simulated and real data, including methylation profiles (EM-seq) of plasma cfDNA paired with tumor biopsy-derived SE levels determined by EM-seq (n=20 pairs) or 10x Visium (n=15 pairs). Clinical outcome prediction was evaluated using pretreatment plasma from patients treated with ICI alone (n=78 melanoma pts), chemotherapy and ICI (n=25 NSCLC pts), and enfortumab vedotin and ICI (n=10 localized muscle-invasive BC (MIBC) pts). Tumor mutational burden (TMB), tumor PD-L1, or circulating tumor DNA (ctDNA) were profiled as a comparator where possible. Results: We observed striking concordance between plasma-derived SE levels and tumor biopsy-confirmed SE levels from the same patients, demonstrating that liquid biopsy analysis of cfDNA methylomes can recapitulate TME spatial biology. Underscoring specificity, no reliable SE signal was detectable from peripheral blood mononuclear cells. In pre-ICI plasma, liquid SE levels were significantly associated with ICI-based response (SE7) versus resistance (SE4) in patients with melanoma, NSCLC, and MIBC, with a mean cross-cancer area under the curve (AUC) of 0. 87 for SE7 (range 0. 8-0. 97) and 0. 81 for SE4 (range 0. 76-0. 85) —far surpassing TMB, tumor PD-L1, and ctDNA in evaluable patients, both in binary analyses of response and multivariable models of survival. Our results also showed potential to predict pathologic complete response in the neoadjuvant setting. Conclusion: Liquid EcoTyper is a scalable noninvasive AI-based framework for spatiotemporal assessment of the TME. By enabling TME-informed risk stratification without surgical biopsies, Liquid EcoTyper has potential to enable more precise and actionable decision making in patients with solid tumors. Citation Format: Erin L. Brown, Wubing Zhang, Abul Usmani, Noah Earland, Ayesha Hashmi, Chibuzor Olelewe, Anushka Viswanathan, Pradeep S. Chauhan, Minji Kang, Chloé B. Steen, Hyun Soo Jeon, Susanna Avagyan, Irfan Alahi, Nicholas P. Semenkovich, Matteo Bergsagel, Janella C. Schwab, Chloe M. Sachs, Faridi Qaium, Peter K. Harris, Antonella Bacchiocchi, Qingyuan Cai, Andrew J. Gentles, Rondell P. Graham, Peter C. Lucas, Ryan C. Fields, Jacob J. Orme, Aaron S. Mansfield, Mario Sznol, Ruth Halaban, David Y. Chen, Aaron M. Newman, Aadel A. Chaudhuri. Liquid biopsy profiling of the tumor microenvironment to determine response to immunotherapy regimens across solid tumors 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 94.
Brown et al. (2026) studied this question.