8048 Background: Durvalumab consolidation after platinum-based chemoradiotherapy (CRT) is the standard of care for unresectable stage III non-small cell lung cancer (NSCLC). However, optimal treatment strategies after progression remain undefined, and data guiding post-durvalumab strategies, including immunotherapy rechallenge, are limited. Methods: This global retrospective multicenter study included patients with unresectable stage III NSCLC treated with concurrent or sequential CRT who progressed after ≥1 dose of durvalumab consolidation. Survival outcomes were assessed using Kaplan-Meier and Cox methods. Immunotherapy sensitivity was defined by time from durvalumab initiation to progression: refractory (< 12 months) or sensitive (≥12 months) and correlated with outcomes after immunotherapy rechallenge. Results: Among 319 patients, median age was 66.5 years; 54.2% were male, and 96.7% had ECOG PS 0-1. Concurrent CRT was delivered in 96.5% (median RT dose 60 Gy). Median follow-up was 62.2 months. Progression pattern included locoregional recurrence in 34.8% and distant relapse in 65.2%. Of 293 patients receiving a first subsequent therapy (FST), chemotherapy (CT) ± VEGF inhibition was most common (31.7%), followed by local ablative therapy (LAT, 30.7%), immunotherapy ± CT (18.4%), targeted therapy (TT, 10.9%), and durvalumab beyond progression (BP) ± LAT (6.5%). When comparing FST, immunotherapy rechallenge demonstrated superiority over CT. Compared with CT ± VEGF inhibition, immunotherapy ± CT was associated with a higher objective response rate (ORR: 40.7% vs. 16.9%, p < 0.01) and significantly improved progression-free survival (mPFS2: 8.48 vs. 4.14 months; HR 0.41, p < 0.01) and overall survival (mOS2: 23.7 vs. 12.2 months; HR 0.41, p < 0.01), findings confirmed in multivariable analyses. Patients with actionable oncogenic drivers or indolent, focal relapses achieved the most favorable outcomes with TT or durvalumab BP ± LAT, respectively, reflecting a disease biology amenable to highly effective systemic or localized interventions. Among patients who received immunotherapy rechallenge, those relapsing ≥12 months after durvalumab initiation experienced higher ORR and numerically longer PFS and OS. Additionally, upfront rechallenge was independently associated with superior survival in multivariable models (PFS aHR: 0.51, p < 0.01; OS aHR: 0.46, p < 0.01) and with higher ORR (41.5% vs. 17.6%, p = 0.04), compared with later-line rechallenge. Conclusions: Post-PACIFIC treatment approaches are heterogeneous, but immunotherapy-based strategies provide superior efficacy over CT alone. Patients with prolonged benefit from durvalumab (≥12 months) retain immune sensitivity and derive the greatest benefit from immunotherapy rechallenge, particularly as FST.

Biological aging occurs heterogeneously across individuals and organs. However, current measures of biological age incompletely capture organ-specific differences in health and disease risk. Because chest CT visualizes multiple thoracic organs, it offers an opportunity to quantify structural aging across organ systems. Here, we developed MOSAIC-Age, a framework characterizing eight organ-specific aging clocks on chest CT. The clocks were developed and validated using 9,971 CT scans from CT-RATE and MIDRC, and subsequently locked and applied to two independent prospective cohorts with 35,293 participants from the National Lung Screening Trial and Genetic Epidemiology of COPD study. CT-derived biological age gaps (BAGs) were examined in relation to lifestyle and socioeconomic factors, prevalent comorbidities, incident chronic diseases, and all-cause and cause-specific mortality. Higher BAGs, indicating organs that appeared older on CT than expected for their chronological age, were broadly associated with adverse health characteristics, chronic disease burden, and increased mortality risk. Multiple disease outcomes were associated with aging across several organs, whereas in multivariable analyses including all eight organ-specific BAGs, the remaining associations were more organ specific. A greater number of markedly older-appearing organs and a faster pace of aging were each associated with higher mortality. Together, these findings demonstrate that routine chest CT captures both shared and organ-specific patterns of biological aging and establish CT-derived organ aging as a quantitative imaging biomarker for assessing multi-organ health and long-term disease risk. ### Competing Interest Statement The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: N. I. V. receives consulting fees from Regeneron, Amgen, Xencor, AstraZeneca, Tempus, Pfizer, Summit, OncoHost, Guardant, ImmunityBio, and research funding from EMD Serono, IDEAYA, Amgen, Summit, Regeneron, Sanofi, BMS, and OncoHost, outside the submitted work. M.C.B.G. has received research funding from Siemens Healthcare. T. C. reports speaker fees/honoraria (including travel/meeting expenses) from ASCO Post, AstraZeneca, Bio Ascend, Bristol Myers Squibb, Clinical Care Options, IDEOlogy Health, Medical Educator Consortium, Medscape, OncLive, PeerView, Physicians' Education Resource, Targeted Oncology; advisory role/consulting fees (including travel/meeting expenses) from AstraZeneca, Bristol Myers Squibb, Daiichi Sankyo, Genentech, Johnson & Johnson, Merck, Nuvalent, oNKo-innate, Pfizer, and RAPT Therapeutics; and institutional research funding from AstraZeneca, Bristol Myers Squibb and Merck. X. L. reports receiving consultant and advisory fees from Eli Lilly, AstraZeneca, EMD Serono, Daiichi Sankyo, Spectrum Therapeutics, Boehringer Ingelheim, Hengrui Therapeutics, Novartis, and research funding from Eli Lilly, Boehringer Ingelheim, all outside of the submitted work. M. A. reports research funding from Genentech, Nektar Therapeutics, Merck, GlaxoSmithKline, Novartis, Jounce Therapeutics, Bristol Myers Squibb, Eli Lilly, Adaptimmune, Shattuck Labs, Gilead, Verismo Therapeutics, and Lyell; advisory board roles for GlaxoSmithKline, Shattuck Labs, Bristol Myers Squibb, AstraZeneca, Insightec, Regeneron, Genprex, and Lyell; speaker fees from AstraZeneca, Nektar Therapeutics, SITC, and Regeneron; and participation on a safety review committee for Nanobiotix-MDA Alliance, Henlius, all outside of the submitted work. A.A.S. reports serving on the advisory board of DELFI Diagnostics and as an advisor to Droplet Biosciences, all outside of the submitted work. D.E.G. reports research funding from AstraZeneca, Karyopharm, and Novocure; stock ownership in Gilead and Medtronic; stock options in Early Marker, Inc. and OncoSeer Diagnostics, Inc.; consulting and advisory roles for AbbVie, AstraZeneca, Bayer, Catalyst Pharmaceuticals, EMD Serono, and GSK; service on data and safety monitoring boards for Daiichi Sankyo, Summit Therapeutics, and Taiho Oncology; royalties from Oxford University Press; and roles as co-founder and Chief Medical Officer of OncoSeer Diagnostics, Inc., all outside of the submitted work. J. Y. C. has received travel sponsorship from Accuray and Varian Medical Systems, and grants from Varian Medical Systems, outside the submitted work. D. L. G. reports honoraria for scientific advisory boards from AstraZeneca, Sanofi, Alethia Biotherapeutics, Menarini, Eli Lilly, 4D Pharma and Onconova, and research support from Janssen, Takeda, Astellas, Ribon Therapeutics, NGM Biopharmaceuticals, Boehringer Ingelheim, Mirati Therapeutics and AstraZeneca, all outside of the submitted work. C. C. W. reports research support from Medical Imaging and Data Resource Center from NIBIB/University of Chicago and royalties from Elsevier, outside of the submitted work. J. V. H. reports receiving advisory/consulting fees from AstraZeneca, Boehringer Ingeheim, Catalyst, Genentech, GlaxoSmithKline, Guardant Health, Foundation Medicine, Hengrui Therapeutics, Eli Lilly, Novartis, Spectrum, Sanofi, Takeda Pharmaceuticals, Mirati Therapeutics, Bristiol Myers Squibb, BrightPath Biotherapeutics, Janssen Global Services, Nexus Health Systems, EMD Serono, Pneuma Respiratory, Kairos Venture Investments, Leads Biolabs, RefleXion, and research funding from GlaxoSmithKline, AstraZeneca, Spectrum, all outside of the submitted work. J. Z. reports grants from Merck, Novartis, Johnson and Johnson; and personal fees from BMS, AZ, Novartis, Johnson and Johnson, GenePlus, Hengrui, Innovent, outside the submitted work. J. W. reports research funding from Siemens Healthcare. The remaining authors declare that they have no competing interests. ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: The Institutional Review Board of The University of Texas MD Anderson Cancer Center gave ethical approval for this work. I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes CT-RATE imaging data are available through the Hugging Face repository under the repository's access conditions for academic, research, and educational use (https://huggingface.co/datasets/ibrahimhamamci/CT-RATE). MIDRC imaging data are available through the MIDRC Data Commons to registered users under the applicable MIDRC data use agreement (https://data.midrc.org/). The NLST imaging data are available through The Cancer Imaging Archive (https://www.cancerimagingarchive.net/collection/nlst/), and NLST clinical data are available through the NCI Cancer Data Access System (https://cdas.cancer.gov/nlst/). The COPDGene phenotype and imaging data used in this study are available through controlled access via dbGaP under study accession numbers phs000179.v7.p2 and phs004023.v1.p2, respectively. The paired CT/PET-CT data are not publicly available owing to patient privacy considerations but are available for research purposes from the corresponding author upon reasonable request. The pretrained organ-clock model weights are also available from the corresponding author upon reasonable request. National Institutes of Health, R01CA262425, R01CA276178 Cancer Prevention and Research Institute of Texas, RP240117 Victory Houston Permanent Health Funds QIAC Partnership in Research Grant Rexanna's Foundation for Fighting Lung Cancer

Foundation models have emerged as powerful feature extractors in computational pathology. However, they typically omit mechanisms for leveraging the global spatial structure of tissues and the local contextual relationships among diagnostically relevant regions—key elements for understanding the tumor microenvironment. Multiple instance learning (MIL) remains an essential next step following the foundation model, designing a framework to aggregate patch-level features into slide-level predictions. We present EAGLE-Net, a structure-preserving, attention-guided MIL architecture designed to augment prediction and interpretability. EAGLE-Net integrates multi-scale absolute spatial encoding to capture global tissue architecture, a top-K neighborhood-aware loss to focus attention on local microenvironments, and background suppression loss to minimize false positives. We benchmarked EAGLE-Net on large pan-cancer datasets, including three cancer types for classification tasks (10,701 slides) and seven cancer types for survival prediction (4,172 slides), using three distinct histology foundation backbones (REMEDIES, Uni-V1, Uni2-h). Across tasks, EAGLE-Net achieved up to 3% higher classification accuracy and the top concordance indices in 6 of 7 cancer types, producing smooth, biologically coherent attention maps that aligned with expert annotations and highlighted invasive fronts, necrosis, and immune infiltration. These results position EAGLE-Net as a generalizable, interpretable framework that complements foundation models, enabling improved prognostic and diagnostic performance.

INTRODUCTION:Thyroid transcription factor-1 (TTF-1) expression, routinely assessed through immunohistochemistry in the diagnostic evaluation of lung adenocarcinomas (LUADs), is negative (TTF-1Neg) in approximately 15% to 20% of cases. Although worse outcomes have been reported for these tumors compared with TTF-1-positive (TTF-1Pos) LUAD, a comprehensive characterization of TTF-1 negativity is currently lacking. METHODS:Patients with LUAD and available TTF-1 immunohistochemistry from five institutions, The Cancer Genome Atlas, the Stand Up To Cancer-Mark Foundation, and the POPLAR/OAK data sets, were included. Features and outcomes were analyzed according to TTF-1 expression. RESULTS:Among 3297 patients, TTF-1Neg (15%, n = 496), compared with TTF-1Pos (85%, n = 2801), was associated with a more frequent tobacco use history and lower PD-L1 expression. TTF-1Neg LUAD was enriched for STK11, KEAP1, SMARCA4, NKX2-1, CDKN2A, and KRAS mutations (q < 0.05). Patients with metastatic TTF-1Neg LUAD treated with immune checkpoint inhibitors (n = 233), compared with TTF-1Pos cases (n = 1179), had worse objective response rates (ORR, 17% versus 28%, p = 0.001), median progression-free survival (mPFS, 2.5 versus 4.4 mo, p < 0.0001), and median overall survival (mOS, 9.6 versus 20.2 mo, p < 0.0001). Similarly, TTF-1Neg cases had worse outcomes to chemoimmunotherapy (ORR, 26% versus 41%, p < 0.0001; mPFS, 4.6 versus 8.2 mo, p < 0.0001; mOS, 11.2 versus 23.4 mo, p < 0.0001), durvalumab after chemoradiation for unresectable stage III disease (mPFS, 8.0 versus 24.8 mo, p = 0.016; mOS, 20.0 mo versus not reached, p = 0.004), and KRASG12C inhibitors in KRASG12C-mutant LUAD (ORR, 13% versus 36%, p = 0.03; mPFS, 2.7 versus 5.9 mo, p < 0.0001; mOS, 4.4 versus 12.1 mo, p < 0.0001). CONCLUSIONS:TTF-1 negativity identifies a subset of LUAD with worse outcomes to immunotherapy, chemoimmunotherapy, and KRASG12C inhibitors.

8590 Background: Clinical trials of immune checkpoint inhibitors (ICIs) in metastatic non-small cell lung cancer (mNSCLC) limited treatment duration to 2 years despite the absence of evidence defining the optimal time on therapy. In routine clinical practice, however, many patients receive ICIs for more than 2 years. We evaluated survival outcomes in patients treated with 2 years versus > 2 years of therapy. Methods: We conducted a global, retrospective cohort study of patients with mNSCLC treated with ICIs for at least 2 years. Patients were classified as receiving fixed-duration ICI (discontinued at 2 years) or indefinite ICI therapy. Outcomes included progression-free survival (PFS), and overall survival (OS) from ICI start. Multivariable Cox regression, propensity score matching, and 2-year landmark analyses with inverse probability weighting (IPW) were used to address time-dependent biases and adjust for relevant covariates. Results: Among 889 patients who received at least 24 months of ICI therapy, median age was 65.5 years; 45.6% were women; 92.8% had a smoking history; 80.2% had adenocarcinoma histology; 70% of patients received ICI as first-line therapy and median PD-L1 tumor proportion score was 60%. Overall, 58.8% received indefinite ICI and 41.2% discontinued at 2 years. Median PFS and OS for the overall cohort were 5.7 years and 8.1 years, respectively. Baseline clinicopathologic characteristics, including age, sex, histology, PD-L1 expression, ECOG performance status, and smoking history, were well balanced between groups. Indefinite ICI therapy was associated with significantly longer PFS (HR 0.75, p < 0.01) and OS (HR 0.62, p < 0.001) compared to 2 years of therapy. These findings were consistent in a propensity-matched analyses (PFS HR 0.62, p < 0.01; OS HR 0.60, p < 0.001) and multivariable models (adjusted PFS HR 0.75, p = 0.02; adjusted OS HR 0.57, p < 0.001). In 24-month landmark analyses with IPW, indefinite ICI remained associated with improved PFS (HR 0.74, p = 0.02) and OS (HR 0.56, p < 0.001). Benefit with indefinite therapy was observed across all predefined subgroups, including age, sex, histology, PD-L1 strata, ECOG performance status, smoking history. Among 98 patients who received ICI rechallenge at disease progression, the objective response rate was 45%, with a median PFS of 12.9 months and a median OS of 30.2 months. Conclusions: Continuation of ICI therapy beyond 2 years was consistently associated with superior PFS and OS compared with treatment limited to 2 years across subgroups and sensitivity analyses. These findings highlight the limitations of arbitrary treatment caps in clinical trial design, support individualized duration decisions, and underscore the need for prospective trials to define the optimal duration of ICI therapy in NSCLC.

Immune checkpoint inhibitors (ICIs) benefit only a subset of patients with metastatic non-small cell lung cancer (NSCLC), but current selection relies on tissue PD-L1 immunohistochemistry (IHC), which is invasive and prone to sampling bias. We developed and validated SCENT (Scalable Ensemble Transformer), a CT-based deep learning model for noninvasive prediction of PD-L1 status and immunotherapy outcomes. In this retrospective study, 972 stage IV NSCLC patients treated with ICIs at MD Anderson were analyzed; SCENT was developed and validated in 640 patients with paired CT and PD-L1 IHC, and clinical applicability was assessed in an additional 332 CT-only patients. Generalizability was evaluated in independent cohorts from Mayo Clinic (n = 72) and the phase III LONESTAR trial (n = 116), where paired baseline and 3-month CT enabled longitudinal assessment. SCENT classified PD-L1 status (50% or higher vs lower) in the MD Anderson cohort with AUC 0.84 (95% CI 0.785 to 0.887), specificity 83.9%, and sensitivity 85.3%, outperforming clinical and radiomics models; external validation achieved AUC 0.80 (Mayo) and 0.78 (LONESTAR). SCENT-derived PD-L1 stratified progression-free survival (HR 1.49, p < 0.001) and overall survival (HR 1.40, p = 0.009), comparable to IHC, and provided complementary prognostic value when combined with IHC, with concordant low-low patients showing the poorest survival (OS HR 1.45, p = 0.008). In LONESTAR, serial SCENT-inferred PD-L1 status showed a borderline association with 3-month progression without paired post-treatment tissue confirmation. SCENT is a generalizable CT-based virtual biopsy for baseline PD-L1 prediction and complementary tissue IHC stratification, with longitudinal use requiring prospective validation.

Distinct mutations in chromatin regulators and aberrant expression of transposable elements (TEs), have been associated with clinical benefit to immunotherapy (IO) in specific clear cell renal cell carcinoma (ccRCC) clinical contexts. However, the relationship between mutations in chromatin regulators and TE expression, and their effect on clinical outcomes, are incompletely understood. Here, we identified TEs expressed in distinct mutational subtypes of ccRCC, with endogenous retroviruses (ERVs) comprising the majority of TEs observed. Of these, ERVs 544 and 2014 were upregulated in PBRM1 mutant samples. Patients with high expression of these ERVs and somatic PBRM1 mutations had improved progression-free survival with IO monotherapy, but not targeted therapy, and their upregulation associated with expression of innate immune pathways. Chromatin accessibility increased at ERV 544 and 2014 loci in PBRM1-deficient ccRCC cells, and ERV 544 and 2014 were upregulated upon in vitro PBRM1 knockout in ccRCC cell line clones. Broadly, our study supports a link between PBRM1 mutations, subsequent chromatin accessibility changes, and aberrant but immunoresponsive ERVs in ccRCC.

The complexity of tumor biology, driven by heterogeneous and dynamic tumor–immune–stromal interactions, remains insufficiently captured by current analysis pipelines and static biomarkers. Triple-negative breast cancer (TNBC) represents a compelling setting for ecosystem-based modeling because it exhibits marked spatial heterogeneity and variable responses to neoadjuvant chemotherapy and chemoimmunotherapy. Here, we present TEMPO (Tumor Ecosystem Modeling for Predictive Outcomes), a spatial proteomic digital twin framework that integrates longitudinal single-cell Imaging Mass Cytometry (IMC), mechanistic modeling, and machine learning to reconstruct and predict patient-specific treatment trajectories. Applied to a phase III TNBC cohort (n=279; >2.5 million cells across serial biopsies), TEMPO addresses a key gap by linking spatial organization and dynamics of the tumor microenvironment to therapeutic response through patient-specific ordinary differential equation (ODE) models of tumor–immune–stromal interactions. We show that response is driven by coordinated ecosystem reprogramming: pathological complete response (pCR) is marked by sustained effector T cell expansion, reduced regulatory T cell activity, and stromal disruption, whereas residual disease (RD) exhibits persistent tumor growth reinforced by immunosuppressive and stromal feedback. Sensitivity analysis identified key parameters, tumor proliferation, effector T cell activity, and stromal inhibition, as central regulatory axes influencing cellular interactions and treatment outcomes. Model-derived dynamic features accurately predict response (AUROC > 0.91), outperforming static approaches, and reveal therapy-specific drivers—stromal–macrophage programs in chemotherapy and immune regulatory balance in chemoimmunotherapy. In silico perturbations further demonstrate that reprogramming key interactions can shift RD-like systems toward pCR-like states, highlighting actionable strategies targeting immune activation and stromal remodeling. Together, TEMPO establishes spatial proteomic digital twins as a mechanistic and predictive framework for precision oncology in TNBC.

8557 Background: PD-L1 low and STK11 mutation associate with immune checkpoint inhibitor (ICI) resistance in non-small cell lung cancer (NSCLC), but appear differentially expressed among racial ethnic groups. This has not been validated in large, diverse studies. Methods: We retrospectively studied NSCLC patients 18 or older, without targetable EGFR or ALK alterations, treated with frontline ICI between January 2014 and February 2020 at MD Anderson Cancer Center and University of Illinois Chicago. We analyzed clinicogenomic and survival characteristics by race/ethnicity. Differences in clinicogenomic predictors were assessed through log-rank and chi-squared comparison of proportions tests. Survival differences were estimated via Kaplan-Meier method. Results: 1694 patients met inclusion criteria, 383 (22.6%) were minorities. Poor performance status (PS 2-3) was most frequent in African American (AA) (32.5%) and Native Alaskan/Hawaiian or American Indian (NAHAI) (27.3%) patients (p=0.005) (Table). Heavy smoking was more frequent in White (50.3%) patients. PD-L1 <1% was most prevalent among Asian (31.3%) and least prevalent among Hispanic/Latino (HL) (15.8%) patients (p=0.001). STK11 mutation rate was most prevalent in AA (14.7%) and least prevalent in HL (6.9%) and Asian (2.5%) patients (p=0.056). Median overall survival (OS) was lower (21.3, 23.5, and 24.3 months) for HL, NAHAI, and AA patients, and higher (25.4 and 30.6 months) for White and Asian patients, respectively (p<0.01). Conclusions: Our dual center study with 22% minority patient representation shows self-reported race can result in biological differences, because of ancestry and/or environmental differences. AA, HL, and NAHAI patients had lower rates of heavy smoking and different clinicogenomic patterns than White patients. AA tended towards higher prevalence of STK11 mutation—while not statistically significant, it was overall not highly represented in the sample. In line with these poor prognostic factors, AA had statistically significant lower median overall survival (OS) than their White and Asian counterparts. Asians had the lowest rates of heavy smoking and STK11 mutation, highest rates of PD-L1<1%, and highest median OS. HL had the lowest OS, with low prevalence of PD-L1<1% and STK11 mutation – future studies should evaluate other clinicogenomic factors to determine potential culprits. White Black or African American (AA) Hispanic or Latino (HL) Asian Native Alaskan/Hawaiian or American Indian (NAHAI) p-value Total cohort n=1694, No. (%) 1311 (77.4) 191 (11.3) 101 (6.1) 80 (4.7) 11 (0.5) ECOG PS 2-3 at ICI start 267 (20.4) 62 (32.5) 23 (22.8) 20 (25.0) 3 (27.3) 0.005 20+ Pack Years Smoking 659 (50.3) 77 (40.3) 26 (25.7) 19 (23.8) 5 (45.5) 4.9 ×10 -9 PD-L1 <1% 317 (24.2) 47 (24.6) 16 (15.8) 25 (31.3) 1 (9.1) 0.001 STK11 mut 129 (9.8) 22 (14.7) 7 (6.9) 2 (2.5) 1 (9.1) 0.056 Median OS (mo) 25.4 24.3 21.3 30.6 23.5 < 0.01

Cancer research is increasingly driven by the integration of diverse data modalities, spanning from genomics and proteomics to imaging and clinical factors. However, extracting actionable insights from these vast and heterogeneous datasets remains a key challenge. The rise of foundation models (FMs) large deep-learning models pretrained on extensive amounts of data serving as a backbone for a wide range of downstream tasks—offers new avenues for discovering biomarkers, improving diagnosis, and personalizing treatment. This paper presents a comprehensive review of widely adopted integration strategies of multimodal data to assist advance the computational approaches for data-driven discoveries in oncology. We examine emerging trends in machine learning (ML) and deep learning (DL), including methodological frameworks, validation protocols, and open-source resources targeting cancer subtype classification, biomarker discovery, treatment guidance, and outcome prediction. This study also comprehensively covers the shift from traditional ML to FMs for multimodal integration. We present a holistic view of recent FMs advancements and challenges faced during the integration of multi-omics with advanced imaging data. We identify state-of-the-art FMs, publicly available multi-modal repositories, and advanced tools and methods for data integration. We argue that current state-of-the-art integration methods provide the essential groundwork for developing the next generation of large-scale, pre-trained models poised to further revolutionize oncology. To the best of our knowledge, this is the first review to systematically map the transition from conventional ML to advanced FM for multimodal data integration in oncology, while also framing these developments as foundational for the forthcoming era of large-scale AI models in cancer research. The GitHub repo of this project available at https://github.com/WuLabMDA/Medical-Foundation-Models.

e20593 Background: Lung squamous cell carcinoma (LUSC) is associated with inferior outcomes to immune-checkpoint inhibitors (ICIs), compared to lung adenocarcinoma. However, the underlying mechanisms of ICI resistance in LUSC remain poorly understood. Methods: The GEMINI database was queried to identify advanced or metastatic LUSC patients who were treated with ICI or in combination with chemotherapy (ICI-chemo). Clinical pathologic, genomic, and outcome data were extracted. PD-L1, TMB, standard of care genomic, and Xenium data on patient tumor samples were analyzed. Results: A total of 569 patients with LUSC treated with ICIs were identified, including 421 in the first-line setting. Overall, 66.5% were male, and 91% were smokers. At baseline, 290 (51%) patients had intrathoracic disease, while 279 (49%) had distant metastasis, including 90 with liver involvement and 71 with brain metastases. The overall response rate was 34%, with a median PFS of 6.5 months (95%CI, 5.9 – 7.3) and a median OS of 19.3 months (95%CI, 16.5 – 22.3). Median PFS was 3.5 months (95%CI, 2.5 – 4.9) for cases with liver metastases, and 5.1 months (95%CI 4.0 – 12.2) for those with brain metastases. Among the 420 cases with progression disease, 256 were limited to baseline organs, 55 involved new organ sites excluding the liver or brain, and 109 had progression in the liver or brain. Patients with liver or brain progression had the worst PFS and OS (p < 0.0001) compared to other progression patterns. The addition of CTLA4 inhibitor or chemotherapy did not enhance outcomes of ICI monotherapy in LUSC patients with low PD-L1 or with liver and brain metastases. At baseline, the median TPS level of PD-L1 expression was negative in non-responders versus 20% in responders (complete or partial response) (p < 0.001). There is no significant difference of TMB between responders and non-responders (7.5 vs 8, p = 0.337). LUSC tumors were enriched for TP53 (84%), CDKN2A (18%), PIK3CA (14%), and NFE2L2 (12%). KMT2C mutations were associated with favorable response (p = 0.01). In contrast, MYC amplification was associated with de-novo progression disease (p = 0.05). From preliminary Xenium analysis of human tissue samples, we observed that the patient who achieved a complete response to pembrolizumab monotherapy exhibited a higher density of immune cell populations and stronger immune signals compared with the 3 patients who demonstrated de-novo resistance to ICI. Conclusions: In LUSC patients, liver metastases are associate with a shorter time to acquired resistance, and progression in the liver or brain correlates with poorer outcomes. Adding a CTLA-4 inhibitor or chemotherapy did not improve outcomes over ICI monotherapy. A more immune enriched tumor microenvironment at baseline may be associated with clinical response to ICIs. Further analysis of the Xenium data is ongoing to explore the resistance mechanism of ICI therapy in LUSC.

e20609 Background: KEAP1 and STK11 are associated with resistance to immune checkpoint inhibitors (ICIs) in NSCLC in KRAS -mutant disease, and are used to guide treatment selection. In contrast, biomarkers to guide treatment selection in KRAS wt NSCLC are poorly defined. We aimed to develop predictive models to guide immunotherapy selection and to identify predictive clinicogenomic biomarkers of response to immunotherapy in KRAS wt NSCLC. Methods: We analyzed patients with metastatic KRAS wt non-squamous NSCLC (nsNSCLC) treated with ICIs at multiple centers, with external validation using the US-based de-identified Flatiron Health–Foundation Medicine NSCLC Clinico-Genomic Database. XGBoost models were trained to predict progression-free survival (PFS) > 6 months following first-line anti-PD-1 monotherapy (ICI-mono) or chemo-immunotherapy (ICI-chemo). Feature importance was assessed using SHAP values. The academic cohort was split 80/20 into training and test sets, and frontline-treated patients in Flatiron served as an external validation cohort. PD-L1 ≥50% (ICI-mono) and PD-L1 ≥1% (ICI-chemo) were used as baseline comparators. Results: The academic and flatiron cohorts included 1,183 and 4087 patients, respectively. The ICI-mono model demonstrated strong discrimination (test AUC = 0.71 vs PD-L1 ≥50% AUC = 0.59) and validated externally (Flatiron real-world time to next treatment [rwTTNT] HR = 0.60, p < 0.001; rwOS HR = 0.60, p < 0.001). The ICI-chemo model did not perform as well internally (test AUC = 0.62 vs PD-L1 ≥1% AUC = 0.53) or externally (Flatiron rwTTNT HR = 0.80, p = 0.039; rwOS HR = 0.80, p = 0.026). In both models, higher tumor mutational burden and PD-L1 were associated with improved outcomes, whereas ECOG and liver metastases were associated with worse outcomes. Genomic features were weakly predictive and did not validate in Flatiron. None of the top predictive genomic features validated in Flatiron, underscoring the importance of external validation. Given prior associations in KRAS -mutant NSCLC, KEAP1 and STK11 were evaluated in KRASwt patients: compared with double–wildtype tumors, KEAP1 -only tumors treated with ICI-mono had improved outcomes (PFS HR = 0.80, p = 0.023; Flatiron rwTTNT HR = 0.80, p = 0.031), STK11-only tumors showed no significant difference, while KEAP1 / STK11 co-mutation was associated with worse outcomes (PFS HR = 1.3, p = 0.048; Flatiron rwTTNT HR = 1.3, p = 0.013). Conclusions: In integrated modeling of KRAS wt nsNSCLC treated with ICIs, no genomic alterations were consistently predictive of response. However, univariate analyses suggest that KEAP1 and STK11 mutations exert distinct and non-additive effects in KRAS wt disease, contrasting with their role in KRAS -mutant NSCLC. These findings have potential implications for treatment intensification strategies, including selective use of anti-CTLA-4 therapy in KRAS wt NSCLC.

e20635 Background: Mutations in STK11 and KEAP1 are consistently associated with poor prognosis in NSCLC. In the lung squamous cell carcinoma (LSCC) subtype, KEAP1 and NFE2L2 mutations converge on constitutive activation of NRF2, a key oncogenic driver of tumor progression. Although KEAP1 and NFE2L2 are among the most frequently mutated genes in LSCC, their combined prognostic significance along with STK11 in LSCC remains poorly characterized. Methods: This multicenter retrospective study enrolled patients with LSCC treated at Dana-Farber Cancer Institute, Memorial Sloan Ketting Cancer Center, and MD Anderson Cancer Center. Genomic and clinicopathologic data were collected for all patients. Eligible patients received immunotherapy, and with or without chemotherapy. Baseline characteristics and clinical outcomes were systematically assessed. Progression-free survival (PFS) and overall survival (OS) were estimated using the Kaplan–Meier method and compared between groups using the log-rank test. Results: Across the three combined cohorts, a total of 407 patients were included, 99 (24.3%) harbored mutations in at least one of the three genes of interest. The median age was 68 years; 75.4% (n = 307) were male, and 89.9% (n = 366) had a history of smoking. Among mutation carriers, 42 had STK11 mutations, 33 had KEAP1 mutations, and 32 had NFE2L2 mutations, with 8 patients exhibiting co-occurring mutations. Baseline clinicopathologic features were generally well balanced between patients with versus without STK11/KEAP1/NFE2L2 mutations. However tumor mutational burden was higher among STK11/KEAP1/NFE2L2 mutant compared wild-type cases (harmonized TMB 0.35 vs -0.03, p < 0.01). In the pooled cohort, the presence of STK11/KEAP1/NFE2L2 was not associated with response rate (37.4% vs 26.9%, p = 0.06) progression-free survival (PFS HR 1.22; p = 0.10) or overall survival (OS HR 1.09; p = 0.57). When stratified by individual gene (STK11, KEAP1 and NFE2L2), none demonstrated a statistically significant association with, ORR, PFS or OS. Conclusions: STK11, KEAP1, and NFE2L2 mutations, alone or in combination, were not associated with inferior outcomes in patients treated with immunotherapy with or without chemotherapy in LSCC. Unlike non-squamous NSCLC, these alterations appear to have limited clinical impact in LSCC, with important implications for clinical decision-making and for the interpretation and design of future trials.

Background:Immune checkpoint inhibitors (ICIs) have revolutionized cancer therapy but can cause serious immune-related adverse events (irAEs), with pneumonitis (ICI-P) being among the most severe. Early identification of high-risk patients before ICI initiation is critical to close monitoring, enable timely intervention, and optimize outcomes. Purpose:To develop and validate a deep learning foundation model to predict ICI-P from baseline CT scans in patients with lung cancer. Methods:We designed the Checkpoint-Inhibitor Pneumonitis Hazard EstimatoR (CIPHER), a deep learning-powered foundation model combining contrastive learning with a transformer-based masked autoencoder to predict ICI-P from baseline CT scans in lung cancer patients. Using self-supervised learning, CIPHER was pre-trained on 590,284 CT slices from 2,500 non-small cell lung cancer (NSCLC) patients, to understand heterogeneous lung parenchyma. Following pre-training, the model was fine-tuned on an internal NSCLC cohort for ICI-P risk prediction, with images from 254 patients used for model development and from 93 patients for internal validation. We compared CIPHER with classical radiomic models. We also validated CIPHER on an external NSCLC cohort of 116 patients. Results:In our internal immunotherapy cohort, CIPHER consistently distinguished patients at elevated risk of ICI-P from those without the event, with AUCs ranging from 0.77 to 0.85. In head-to-head benchmarking, CIPHER achieved an AUC of 0.83, outperforming radiomic model. In the external validation cohort, CIPHER maintained high performance (AUC=0.83; balanced accuracy=81.7%), exceeding the radiomic models (Delong p=0.0318) and demonstrating superior specificity without sacrificing sensitivity. By contrast, radiomic model, despite high sensitivity (85.0%), showed markedly lower specificity (45.8%). Confusion matrix analyses confirmed CIPHER's robust classification, correctly identifying 80 of 96 non-ICI-P cases and 16 of 20 ICI-P cases. Conclusions:We developed and externally validated CIPHER for predicting future risk of developing ICI-P from pre-treatment CT scans. With prospective validation, CIPHER can be incorporated into routine patient management to improve outcomes.

INTRODUCTION:Homozygous deletion of 9p21, containing CDKN2A/B and MTAP (CMdel), underlies emerging therapeutic strategies including PRMT5/MAT2A inhibition. The clinicogenomic context and treatment outcomes of CMdel in NSCLC remain incompletely defined. METHODS:We analyzed Tempus Lens records with paired DNA/RNA sequencing. After quality filtering, 16,947 patients were included in the primary dataset and 10,996 in the validation dataset, and stratified by deletion status (CMdel vs CMwt). Interferon status was defined by interferon-kappa (IFNK) expression (log2[TPM+1]) using the 25th percentile of CMwt as threshold. Immune composition was inferred by quanTIseq. First-line regimens were grouped as ICI-mono, ICI-chemo, chemotherapy, or TKI. Best response and time-to-next-treatment with risk-set adjustment were assessed. RESULTS:CMdel occurred in 11.7% (n=1,991) and 15% (n=1,676), and was enriched in mucinous histology, stage IV disease, and never-smokers. CMdel associated with EGFR, TERT, and SMARCA4 alterations, whereas CMwt associated with RB1, SETD2 and ARID2. >50% of CMdel tumors harbored targetable co-alterations, enriched in EGFR/ALK/RET and relatively depleted in KRAS. CMdel tumors exhibited lower PD-L1, mutation burden, and immune infiltration. TTNT did not differ by CM status for ICI-mono or TKI, but was shorter in CMdel for ICI-chemo and chemotherapy. Discordant and broad deletions were less common and had weaker TTNT association. IFNK-low expression marked immune-excluded subgroups with worse treatment outcomes. CONCLUSIONS:CDKN2A/MTAP loss defines a prevalent, genomically distinct NSCLC subset enriched in actionable drivers. CMdel associates with worse outcomes to ICI/chemotherapy-containing regimens. These findings highlight convergent biology at 9p21 and support biomarker-guided development of PRMT5/MAT2A inhibitors in MTAP-deleted NSCLC.

e20583 Background: KEAP1 and STK11 are associated with poor response to immune checkpoint inhibitors (ICIs) in KRAS m NSCLC, but their distinct vs. combined effects are incompletely defined. Methods: We analyzed metastatic KRAS m NSCLC patients treated with ICIs at multiple centers, with external validation in the US-based deidentified Flatiron Health-Foundation Medicine NSCLC Clinico-Genomic Database. PFS/OS and real-world time-to-next-treatment (rwTTNT)/rwOS were the primary outcomes in the academic and Flatiron cohorts, respectively. Patients were stratified as KEAP1 m only (K), STK11 m only (S), co-mutant (KS), or wildtype (WT). To identify their predictive effects independent of other clinicogenomic variables, XGBoost models were trained to predict PFS > 6 months from anti-PD-1 monotherapy (ICI-mono) or chemo-immunotherapy (ICI-chemo). SHAP identified feature importance. The academic cohort was split 80/20 into train/test and Flatiron served as external validation. CD8⁺ T-cell infiltration was assessed by immunofluorescence and validated using MCP-counter-inferred CD8⁺ T-cells from TCGA. Results: The academic and flatiron cohorts included 1019 and 2261 patients, respectively. As compared to WT tumors treated with ICI-mono (n=387), K tumors had no difference in PFS (n=58, HR 1.2, p=0.25), while S (n=56; HR 1.9, p<0.001) and KS tumors (n=75; HR 2.2, p<0.001) had progressively worse PFS. OS mirrored this: K tumors had no difference (HR 1.3, p=0.13); S (HR 1.6, p=0.0032) and KS tumors had progressively worse OS (HR 2.4, p<0.001). As compared to WT tumors treated with ICI-chemo (n=219), K tumors trended toward worse PFS (n=35; HR 1.4, p=0.10), while S (n=60; HR 1.4, p=0.033) and KS tumors (n=89; HR 2.2, p<0.001) had progressively worse PFS. OS was similar for S tumors (HR 1.2, p=0.38) but worse for K (HR 1.8, p<0.001) and KS tumors (HR 1.9, p<0.001). Flatiron validation confirmed these results. As compared to WT tumors (median PD-L1 30%), PD-L1 was lower in K (5%), and even lower in S (0%) and KS (0%) tumors. XGBoost models performed well for ICI-mono (test: AUC=0.70; Flatiron: rwTTNT HR 0.6, p<0.001) and ICI-chemo (test: AUC=0.73; Flatiron: HR 0.6, p<0.001). KS was the most predictive genomic feature in each model; K or S were not predictive. Compared with WT tumors, overall CD8 T-cell density was reduced in K and KS, but not S, tumors, a finding validated in TCGA. In contrast, CD8 T-cell density was reduced in the tumor-compartment in S and KS, but not K, tumors, suggesting that KEAP1 impairs global CD8 infiltration, whereas STK11 restricts intratumoral CD8 localization. Conclusions: In KRAS m NSCLC, KEAP1 and STK11 exhibit varying predictive and prognostic effects and immune-phenotypes depending on whether they occur alone or together. These findings have implications for risk stratification, treatment selection, and the development of therapeutics to overcome ICI resistance in these patients.

BACKGROUND:Patients with BRAFV600E (ie, Val600Glu)-mutated non-small-cell lung cancer (NSCLC) can be treated with BRAF and mitogen-activated protein kinase (MEK) inhibitors, or with immune checkpoint inhibitors (ICIs) with or without chemotherapy. We aimed to investigate which initial systemic treatment should be prioritised in this population. METHODS:In this retrospective cohort study conducted across 17 centres in the USA, Italy, France, and Brazil, clinicopathological data were collected from participants aged 18 years and older with stage IV, treatment-naive, metastatic BRAFV600E-mutated NSCLC and with an Eastern Cooperative Oncology Group performance status of 0-3, who started first-line treatment with ICIs with or without chemotherapy (PD-1 or PD-L1 inhibitors with or without platinum-based chemotherapy) or BRAF and MEK inhibitors (dabrafenib and trametinib or encorafenib and binimetinib) between Jan 2, 2015, and July 11, 2024. The primary endpoint was overall survival with first-line ICIs with or without chemotherapy versus with BRAF and MEK inhibitors. FINDINGS:284 participants were identified for this study, of whom 88 (31%) received ICIs with or without chemotherapy and 196 (69%) received BRAF and MEK inhibitors. The median age of participants was 68 years (IQR 61-74), and 148 (52%) participants were female and 136 (48%) male. Participants in the ICIs with or without chemotherapy group had a higher history of smoking (73 [83%] vs 118 [60%]; p=0·0002) and a higher PD-L1 expression (≥50% in 58 [66%] vs 76 [39%], 1-49% in 16 [18%] vs 67 [34%], and <1% in eight [9%] vs 31 [16%]; p=0·0003) than those in the BRAF and MEK inhibitor group. At a median follow-up time of 45·0 months (95% CI 39·0-55·7), ICIs with or without chemotherapy were associated with improved median overall survival compared with BRAF and MEK inhibitors (40·9 months [95% CI 33·3-not reached] vs 25·2 months [19·9-31·1]; hazard ratio [HR] 0·69 [0·49-0·98], p=0·039). In subgroup analyses, ICIs with or without chemotherapy, compared with BRAF and MEK inhibitors, were associated with longer median overall survival in participants with a history of smoking (HR 0·60 [0·40-0·90], p=0·013), with a PD-L1 tumour proportion score of ≥1% or higher (HR 0·66 [0·45-0·98], p=0·039), aged 70 years or older (HR 0·54 [0·31-0·94], p=0·029), with TP53 co-mutations (HR 0·46 [0·27-0·79], p=0·0048), and without brain metastases (HR 0·66 [0·45-0·99], p=0·045). With BRAF and MEK inhibitors, frequencies of adverse events of any grade and of grade 3 and higher were similar whether administered as first-line therapy or as second-line therapy following ICIs with or without chemotherapy. INTERPRETATION:First-line ICIs with or without chemotherapy were associated with improved overall survival compared with BRAF and MEK inhibitors in participants with metastatic BRAFV600E-mutated NSCLC, particularly among specific subpopulations. These findings, although suggesting potential clinical relevance, remain exploratory and require confirmation from prospective studies. FUNDING:NextGenerationEU.

Supp Figure 3. ATM co-mutation enrichment analyses within cohorts, stratified by KRAS mutation status.

List of 301 genes included in all versions of MSK-IMPACT and DFCI OncoPanel NGS platforms used for all analyses.