Citations

Abstract Understanding how immune checkpoint inhibitors (ICIs) reshape human T-cell responses requires models that move beyond static transcriptomic snapshots and discrete cell-state classifications. Here, we present an integrative AI-enabled virtual cell framework that represents human CD8⁺ T-cell responses as dynamic and computable systems during ICI therapy. By integrating single-cell RNA sequencing with paired T-cell receptor sequencing within the C2S-scale foundation model, we construct a virtual representation of individual T cells, in which each cell is encoded by a unique functional identity that captures its transcriptional, signaling, and clonal characteristics. Using this framework, we identify a previously unrecognized dynamic latent effector state of CD8⁺ T cells characterized by intermediate expression of effector genes, distinct signaling activity, and ongoing clonal expansion. Across independent patient cohorts, the virtual cell model consistently indicates that ICI therapy mainly acts by unmasking pre-existing effector potential rather than inducing de novo effector differentiation. Notably, this latent effector population remains transcriptionally restrained despite active signaling and clonal expansion, revealing a hidden reservoir of antitumor immune capacity. More broadly, our study demonstrates how AI-enabled virtual cell modeling can reconstruct latent cellular states and their dynamic transitions from multidimensional single-cell data. By incorporating functional identity into virtual cell model, this framework uncovers biologically meaningful yet non-obvious T-cell effector program during cancer immunotherapy and provides a generalizable approach for studying immune dynamics in human disease.

TPS160 Background: Up to 50% of patients with gastric cancer (GC) may develop peritoneal metastasis (PM) that is associated with resistance to systemic therapy and poor prognosis. Although cytoreductive surgery and intraperitoneal (IP) chemotherapy were shown to modestly improve the survival of patients with PM from gastric GC (PM-G), these clinical trials were performed in the pre-immunotherapy era. The significant heterogeneity in trial design and patient selection criteria further limit the applicability of these trial results to current clinical practice where anti-PD-1/L1 immunotherapy and/or targeted therapy in combination with chemotherapy have become a common, standard first-line therapy for the majority of patients with metastatic GC, raising the question of whether chemotherapy such as paclitaxel remains the most effective IP therapy. IP therapy using novel immunotherapy such as bispecific antibodies and chimeric antigen receptor T-cell therapy have demonstrated safety and promising antitumor activities in PM-G. We are conducting a phase 1b clinical trial that evaluates the efficacy of IP interleukin-2 (IL-2) in patients with PM-G with stable disease on first-line chemo-immunotherapy based on the hypothesis that IP IL-2 may synergize with intravenous (IV) chemo-immunotherapy to improve the treatment of PM-G. Multi-omics studies using patient blood, peritoneal fluid, and tissue specimens serially collected from this trial will be performed to characterize the changes in the tumor immune microenvironment of PM-G and identify new therapeutic targets. Methods: Patients with PM only from gastric adenocarcinoma are eligible for this phase 1b trial. All participants will continue IV fluorouracil-containing chemotherapy with nivolumab every 2 weeks and concurrently receive up to 16 doses of IP IL-2 (aldesleukin 6 × 10 5 IU/m2) weekly through an indwelling peritoneal catheter (Bard). Diagnostic laparoscopy (DL) will be performed before initiation of IP IL-2 and after 8 doses of IP therapy to assess treatment response based on the peritoneal cancer index (PCI), which is the primary endpoint of the trial. Patients with at least stable radiographic disease will continue to receive another 8 doses of IP IL-2 and undergo a third DL. Secondary endpoints include safety, peritoneal regression grading score, overall survival, and progression-free survival. Major inclusion criteria include mismatch repair-proficient or microsatellite-stable GC, PCI of 3-24, absence of distant metastasis other than PM, stable disease on first-line IV chemo-immunotherapy, tolerability of anti-PD-1 therapy, and adequate organ function. Enrollment began in 2024 and remains active. To date, seven of planned 15 patients have received study treatment without significant treatment-related adverse events and three of these patients underwent curative-intent cytoreductive surgery. Clinical trial information: NCT05802056 .

PURPOSE Gastroesophageal carcinoma (GEC) is increasingly managed using biomarker-guided therapies. Although both tissue DNA (tDNA) and circulating tumor DNA (ctDNA) testing are used clinically, the concordance of genomic alterations between primary tumors, metastatic sites, and liquid biopsy remains incompletely understood. METHODS We retrospectively analyzed 345 patients with GEC who underwent ctDNA testing at Mayo Clinic. tDNA next-generation sequencing and immunohistochemistry (IHC) for human epidermal growth factor receptor 2 (HER2)/erb-b2 receptor tyrosine kinase 2 ( ERBB2 ) and mismatch repair (MMR) were collected. Concordance was evaluated between ctDNA and tDNA, ctDNA and IHC, and across all three modalities, using 60-day and 30-day intervals. RESULTS Among 345 patients with ctDNA testing, 186 had tDNA testing, and 310 underwent IHC. ctDNA testing showed a shorter turnaround time than tDNA (median 6 v 13 days). The most frequent alteration in ctDNA, primary tDNA, and metastatic tDNA was TP53. Concordance between ctDNA and primary tDNA was ≥78.8% for key genes; between ctDNA and metastatic tDNA, it was ≥84.8%. IHC-ctDNA concordance was higher in metastatic tumors, with 91.1% for HER2/ ERBB2 and 97.7% for mismatch repair (MMR)/microsatellite instability (MSI) compared with 81.1% and 95.8%, respectively, in primary tumors. Triplet concordance for MSI/MMR was 97.6% in primary tumors and 100% in metastatic tumors, while ERBB2 /HER2 showed 84.4% and 93.9% concordance, respectively. ctDNA temporal heterogeneity analysis revealed 14 patients with fluctuating ERBB2 status and 3 with fluctuating MSI-H status. CONCLUSION ctDNA testing demonstrates high concordance with tDNA and IHC and offers faster turnaround. Given its minimally invasive nature and strong concordance with tissue-based testing, ctDNA is a reliable tool for genomic profiling and longitudinal monitoring in GEC.

837 Background: A recent randomized trial suggested that CR may influence ICI activity in patients with lung cancer. However, the same has not been established in GI cancers, where reports have shown mixed results with variable definitions of 'early' vs 'late' IT. We evaluated multiple approaches to assess the impact of CR on first-line ICI in GI malignancies. Methods: This retrospective study included patients from three Mayo Clinic sites (N=455). Both previously reported IT cutoffs and a new cutoff derived from ROC analysis were evaluated. Patients were categorized by the median of all ITs, the median of the first four ITs, and whether ≥20% of ITs occurred before the identified cutoff. Univariate and multivariate Cox regression (adjusting for age, MSI status, type of cancer, metastasis status, number of cycles, and receipt of chemotherapy with ICI) assessed associations with OS. KM analyses were performed to illustrate survival differences. The primary endpoint was OS. Results: When patients were categorized based on whether ≥20% of infusions occurred before 14:00 (ROC-derived cutoff), a significant association with OS was observed. On univariate analysis, early infusion timing was associated with improved survival (HR 2.30, 95% CI 1.58–3.32, p <0.001). This remained significant after multivariable adjustment (HR 2.11, 95% CI 1.41–3.16, p < 0.001). Median OS was 14.2 months in the early group vs 3.8 months in the late group (log-rank p < 0.001). Using previously defined cutoffs, categorization by the median IT of all infusions or the median of the first four infusions did not demonstrate a significant association with OS in either univariate or multivariate Cox regression analyses. Conclusions: These findings highlight the potential role of infusion timing as a chronotherapy factor in GI cancers. Given the retrospective design and heterogeneous population, prospective studies are warranted to validate this signal and clarify its clinical relevance. Baseline demographics of patients by infusion timing (early vs late). Variables Early N=412 (%) LateN=43(%) Total N=455(%) p-value Age <70 249(60.4) 21(48.8) 270 (59.3) 0.1 >70 163(39.6) 22(51.2) 185 (40.7) Gender F 118(28.6) 16(37.2) 134 (29.5) 0.3 M 294(71.4) 27(62.8) 321 (70.5) Cancer Type Gastroesophageal 109 (26.5) 8 (18.6) 117 (25.7) 0.1 Hepatocellular 159 (38.6) 24 (55.8) 183 (40.2) Biliary 86 (20.9) 4 (9.3) 90 (19.8) MSI-H 58 (14.1) 7 (16.3) 65 (14.3) Metastasis No 150 (36.4) 22 (51.2) 172 (37.8) 0.1 Yes 262 (63.6) 21 (48.8) 283 (62.2) Immunotherapy Cycle Number <6 189 (45.9) 33 (76.7) 222 (48.8) <0.001 >6 223 (54.1) 10 (23.3) 233 (51.2)

830 Background: Immune checkpoint inhibitors (ICIs) are standard therapy for MSI-H GI cancers. Lynch syndrome, caused by germline mutations in mismatch repair (MMR) genes, leads to MSI-H tumors and robust ICI responses, but its effect on irAEs remains poorly defined. Clarifying this risk may inform treatment selection and toxicity monitoring. Methods: We retrospectively reviewed MSI-H or MMR-deficient (dMMR) GI cancer patients (pts) treated with ICIs across Mayo Clinic (AZ, MN, FL) from January 2017 to June 2025. Eligible pts were identified using Mayo Data Explorer. Demographics, treatment, and irAEs (graded per CTCAE v5.0) were collected. Lynch status was determined by germline testing when available. Logistic regression estimated odds ratios (ORs) for any grade and grade ≥3 irAEs, adjusting for age, cancer type, prior chemotherapy, and treatment regimen. Results: A total of 209 pts with dMMR/MSI-H GI cancer were included: 61 with Lynch syndrome and 148 without. Lynch pts were younger (median 58 vs 76 years, p < 0.001) and more likely to have received prior chemotherapy (62% vs 43%, p = 0.009). Colorectal cancer was the predominant cancer type (74%), but Lynch pts had proportionally more non-colorectal GI cancers (15% vs 2.7%, p = 0.007). Overall, 29% of pts experienced an irAE, with rates significantly higher in Lynch vs non-Lynch pts (43% vs 24%, p = 0.005). The most frequent irAEs were gastrointestinal, endocrine, and dermatologic, followed by hepatic and pulmonary events. Grade ≥3 irAEs occurred in 16% overall, with no significant difference between groups (p = 0.124). In adjusted models, Lynch syndrome was independently associated with a higher risk of any grade irAE (OR 2.14, 95% CI 1.09 - 4.19, p = 0.026), while risk of severe irAEs was not increased (OR 1.55, 95% CI 0.64 - 3.59, p = 0.316). Lynch syndrome pts were more likely to have combination therapy (17% vs 6.8%, p = 0.029). However, there were no differences observed in any grade irAE between monotherapy and combination immunotherapy (OR 1.36, 95% CI 0.49 - 3.54, p = 0.535). Conclusions: Lynch syndrome is associated with a higher prevalence of any-grade irAEs among dMMR/MSI-H GI cancer pts treated with ICIs, though severe events were not increased. Incorporating Lynch status into immunotherapy planning may improve risk stratification and guide regimen selection.

399 Background: The standard of care for resectable GEA adenocarcinomas is perioperative chemotherapy with the FLOT (fluorouracil, leucovorin, oxaliplatin, docetaxel) regimen. Recent studies investigating the addition of checkpoint inhibitors (CPIs) to perioperative chemotherapy have shown mixed results. An NMA can leverage both direct and indirect comparisons to evaluate the relative efficacies of available regimens in this setting. Methods: MEDLINE, EMBASE, Scopus, Web of Science and CENTRAL were searched until August 27th, 2025. Eligible studies included phase 3 randomized trials evaluating perioperative or neoadjuvant chemotherapy +/- CPIs ± radiation in resectable GEA. The primary outcomes were disease-free survival (DFS) and overall survival (OS). Hazard ratios and 95% confidence intervals (CIs) were estimated using a frequentist NMA framework. Results: Fifteen trials including 8072 patients were analyzed. For DFS, perioperative durvalumab + FLOT (dFLOT) ranked highest (P-score = 0.99), followed by FLOT (P-score = 0.84) (HR for dFLOT vs. FLOT was 0.71, 95% CI 0.58-0.86). dFLOT also ranked higher than perioperative pembrolizumab + cisplatin and fluorouracil/capecitabine (KN-585), though this comparison was not statistically significant (HR 0.65, 95% CI 0.40-1.04). Compared with CROSS, dFLOT showed superior DFS (HR of 0.48, 95% CI 0.36-0.63). For OS, dFLOT ranked the highest (P-score=0.95) and was superior to CROSS (HR 0.57, 95% CI 0.42-0.77). While dFLOT appeared favorable versus FLOT (HR 0.78, 95% CI 0.62-0.98), this finding was primarily driven by MATTERHORN, which did not meet its prespecified OS significance boundary (P = 0.03 vs threshold P < 0.0001), and the proportional hazards assumption was violated. Similarly, no significant OS difference was observed between dFLOT and the KN-585 arms (HR 0.75, 95% CI 0.45-1.23). Conclusions: In this NMA, perioperative FLOT remained the benchmark regimen for resectable GEA. dFLOT has shown promising early results, ranking the highest for DFS and showing favorable, but not statistically definitive OS trends in comparison to FLOT. Importantly, the network enabled indirect comparisons not tested in head-to-head trials, demonstrating that dFLOT ranked above pembrolizumab+CF/CX (KN-585) without reaching statistical significance, and also outperformed CROSS. However, these findings should be interpreted with caution, as they are primarily driven MATTERHORN data in which OS did not cross the prespecified threshold and proportional hazards assumptions were violated. Perioperative immunotherapy combinations appear promising, but final MATTERHORN results and further confirmatory studies are needed before dFLOT can be considered as a new standard of care.

334 Background: Sarcopenia, or decreased skeletal muscle mass, has been associated with worse surgical and oncologic outcomes across multiple malignancies. This study investigates the impact of sarcopenia on outcomes after multimodal therapy for esophageal cancer. Methods: Retrospective observational cohort study including adult patients who underwent neoadjuvant therapy followed by esophagectomy for primary esophageal cancer from January 2000 - December 2021. A fully automated abdominal segmentation using deep convolutional neural network was utilized to calculate skeletal muscle index (SMI) from preoperative CT scans performed at our institution. Males with SMI <52.4 cm 2 /m 2 and females with SMI <38.5 cm 2 /m 2 at any point before surgery were classified as sarcopenic. The relationship between sarcopenia and patient demographics, operative details, and complications were explored using chi-square tests. The relationship between sarcopenia and disease-free survival and overall survival were investigated using parsimonious multivariable analysis retaining variables with p-value <0.05 using backward selection method. Results: We identified 200 patients meeting the study criteria with 142 (71%) having sarcopenia at any time prior to surgery. Patients with sarcopenia were significantly older [65.2 (59.6, 71.2) vs 63 (53.6, 67.5), p = .023] and more likely to have had open surgery [129 (90.8%) vs 46 (79.3%), p = .034]. There were no significant differences in sex, history of diabetes, coronary artery disease, renal disease, smoking, cancer type, cancer stage, tumor size, tumor location, or margin status between sarcopenic groups ( p > .05). Sarcopenic patients were significantly more likely to require blood transfusion ≥3 units after esophagectomy [79 (55.6%) vs 8 (13.8%), p < .001] or to experience complications requiring surgical intervention [36 (24.6%) vs 7 (12.1%), p = .048]. Multivariate analysis revealed that sarcopenia was associated with disease-free and overall survival (Table). Conclusions: Sarcopenia is a prognostic factor in patients undergoing multimodal therapy for esophageal cancer. Interventions that aim to prevent decreases in SMI after initiating neoadjuvant therapy should be investigated. Multivariate analysis with sarcopenia predicting outcomes. Disease free survivalHR (95% CI) p-value Overall survivalHR (95% CI) p-value Age 1.04 (1.02, 1.06) < .001 1.04 (1.02, 1.07) < .001 Epicenter Middle Upper 1.88 (1.21, 2.86)3.16 (1.67, 5.98) .001.005< .001 1.96 (1.26, 3.04)4.00 (1.90, 8.44) < .001.003< .001 Stage Stage II Stage III Stage IV 1.88 (0.95, 3.71)2.90 (1.46, 5.75)16.02 (6.12, 41.93) < .001.069.002< .001 1.65 (0.83, 3.25)2.76 (1.37, 5.55)10.14 (4.03, 25.53) < .001.151.005< .001 Cancer type - - 1.76 (1.03, 3.02) .039 Transfusion > 3 pRBC - - 1.54 (0.99, 2.40) .057 BMI prior to cancer diagnosis - - 0.98 (0.94, 1.01 .208 Sarcopenic 1.76 (1.16, 2.67) .007 1.64 (1.04, 2.60) .033

103 Background: Clinicians spend significant time extracting and synthesizing information to risk-stratify and make adjuvant chemotherapy (AC) recommendations for CRC. By automating parts of this process, LLMs could help reduce documentation burden and variability in decision-making. Methods: We selected 144 stage III, pMMR CRC patients treated at Mayo Clinic (2017–2021). HIPAA-compliant Gemini-2.0-Flash-001 and GPT-4o extracted clinical variables (age, neuropathy grade, pathological stage) from EHRs and generated AC recommendations. Two approaches were compared: (1) a rule-based algorithm with nested if/else statements translating variables into NCCN-based recommendations, and (2) a dynamic LLM agent generating recommendations using a structured prompt informed by NCCN guidelines. Concordance with treating physicians was assessed, with disagreements reviewed by a blinded GI oncologist. Results: The dataset included 144 patients with median age 67 (IQR 66–68), 48.6% female. The LLM pipeline extracted neuropathy grade with 100% accuracy and pathological T/N stages with 99.3% accuracy. Concordance with physician recommendations was 93.1% for the rule-based algorithm and 62.5% for the LLM agent. In discordant cases, blinded adjudication by a GI oncologist favored the algorithm and LLM agent over physicians in most cases (Table 1). Conclusions: This exploratory work shows LLM agents can reliably extract variables and generate evidence-based treatment recommendations for stage III CRC. The high rate of expert preference for LLM recommendations in discordant cases is intriguing but limited by small sample size and possible adjudicator bias. Discordant physician recommendations reflect the complexity of decision-making, including patient preferences, performance status, comorbidities, and subtle findings not fully captured in EHRs. Rather than replacing judgment, we envision LLMs as decision support tools—analogous to an advanced trainee reviewing charts and presenting evidence-based recommendations to the attending physician with ultimate authority. Larger multi-institutional studies with multiple adjudicators are needed to validate these findings and clarify AI’s role in oncology. Patient characteristics, LLM extraction accuracy, and concordance with physician recommendations. Metric Result Total patients 144 Median age (IQR) 67 (66-68) Female (%) 48.6% LLM extraction accuracy – Neuropathy grade 100% LLM extraction accuracy – Pathological T and N stages 99.3% Primary Concordance – Rule-based algorithm vs. physician 93.1% Adjudicator preference of rule-based algorithm over treating physician for discordant cases 100% Concordance – LLM agent vs. physician 62.5% Adjudicator preference of LLM agent over treating physician for discordant cases 82.0%

PURPOSE Neoadjuvant immune checkpoint inhibitors (nICIs) have demonstrated high response rates in deficient mismatch repair/microsatellite instability–high (dMMR/MSI-H) gastroesophageal adenocarcinoma (GEA). The NEONIPIGA and INFINITY trials demonstrated high rates of pathologic complete response (pCR) in this patient population. Furthermore, the INFINITY trial explored the feasibility of managing these patients nonoperatively, demonstrating promising results. This study aimed to evaluate clinical outcomes of nICIs in resectable dMMR/MSI-H GEA, with a focus on the feasibility of nonoperative management (NOM). MATERIALS AND METHODS This retrospective cohort study included patients with resectable dMMR/MSI-H GEA and treated with nICIs ± surgery at the Mayo Clinic. Patients were identified from institutional records, and clinical data were retrospectively reviewed. Primary outcomes were clinical complete response (cCR) and pCR. Secondary outcomes included event-free survival (EFS), radiologic complete response (rCR), and immune-related adverse events (irAEs). RESULTS A total of 26 patients treated between April 1, 2017, and July 30, 2025, were identified. Nine patients (34.6%) underwent surgery, of whom six (66.7%) achieved pCR. Seventeen patients (65.4%) pursued NOM, with 10 (71.4%) of 14 evaluable patients achieving cCR and 14 (82.4%) of 17 evaluable achieving rCR. One patient who initially achieved cCR had a local recurrence on surveillance endoscopy and underwent salvage endoscopic resection. At a median follow-up of 19.3 months, 15 (88.2%) of 17 patients in the NOM cohort were alive and metastasis-free, with EFS rates of 87.3% at 12 and 24 months for all patients. irAEs occurred in nine patients (34.6%), with no grade ≥3 toxicities. CONCLUSION In this retrospective cohort study, nICIs led to high cCR and pCR rates in resectable dMMR/MSI-H GEA, supporting the use of immune checkpoint inhibitors in this setting and the feasibility of NOM in select patients.