Inputs into cancer prognostic models are primarily structured data such as demographic and clinicopathological features, and lack richer and temporal context often found in unstructured clinical notes. We hypothesize that creating a temporal clinical patient note from structured data that preserves longitudinal and clinical contextual information, and coupling it with a large language models (LLM) that is trained to prognosticate overall survival (OS), may improve model accuracy with an interpretable embedding space. We developed the Patient Chronological Note (PCN), an algorithm that converts structured data elements into textual strings, mirroring physician notes of patient histories. A bidirectional large language model, was pre-trained using PCNs from breast cancer patients (N=580,000), allowing the LLM to learn a representation of the patient journey. The resulting embeddings were fed into a fully-connected 2-layer network that was fine-tuned using Cox survival loss. Fine tuning was performed on PCNs derived from mBC patients (N=28,500), where the model was trained to predict OS from the time of first metastatic diagnosis. A held-out validation dataset of mBC patients (n=28,800) was used to validate survival prediction accuracy. Our LLM-Cox model achieved a prediction performance of 0.66 (concordance index) on the validation cohort, outperforming a standard linear cox model that achieved 0.62. A marginal effect analysis showed that features associated with metastasis, medications, and patient demographics were most important in prediction performance while hormone receptor status and sequencing information were less important. Clustering the LLM embeddings revealed 10 distinct patient groups enriched for key mBC traits, including a high-risk cluster enriched for triple negative status, TP53 mutations, and African American race, a low-risk cluster enriched for ESR1 mutations and CDK46 treatment, and a cluster enriched for low-risk early-onset patients. The different clusters showed different prognostic risk levels. LLM-Cox model learning from PCNs can improve prediction performance over standard models. The internal embedding representation of the LLM was interpretable, and yielded distinct clinical-molecular subtypes that also showed distinct levels of prognostic risk. Furthermore, creating LLM-based clinical-molecular groups of patients with similar journeys and similar prognostic risk presents an opportunity to identify novel stratifications within each group that are associated with treatment-specific responses and not prognostic risk. Raphael Pelossof, Mark Carty, Talal Ahmed, Stanislas Lauly, Alberto Purpura, Erik Mueller, Justin Guinney. Multi-modal large language models for metastatic breast cancer prognosis [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 5006.

Abstract Trastuzumab deruxtecan (TDXD) is approved for HER2 low or positive metastatic breast cancer. HER2-status is assessed through HER IHC and ERBB2 FISH assays. HER2-status has been shown to be correlated to RNA expression. This study aimed to assess if the addition of DNA to an RNA-based model can improve HER2-status prediction performance. Furthermore, to investigate the relationship between model score and outcomes for TDXD patients according to their overall survival (OS). Training and test sets of breast cancer samples were collected for HER2 status prediction (n = 1275, n = 397 respectively). HER2 status was reported positive: 11%, 12%, low: 61%, 60%, negative: 28%, 29% for each set respectively. HER2 testing and NGS were performed on samples with the same collection date. A TDXD discovery cohort of 284 patients included patients who received TDXD after RNA/DNA collection and had at least a 30-day follow-up. Metastatic disease was reported for 98% of the TDXD cohort. Receptor status was HR+/HER2- 39%, HR-/HER2- 17%, HR+/HER2+ 16%, HR-/HER2+ 9%, unknown 19%, and HER2-status was negative 11%, low 42%, and positive 27%, unknown 20%. OS was measured from medication start date. We trained a 12-gene linear model based on RNA expression and DNA copy number to predict HER2-positivity. The genes were selected by stepwise-selection on RNA and DNA features when predicting HER2-positivity using a random forest model. A baseline model consisting only of ERBB2 RNA expression was compared. Setting a HER2-positivity score threshold and allowing for an indeterminate group of 15% of the validation cohort, the RNA-DNA model achieved 91% PPA and 91% PPV in validation, whereas the RNA-only model achieved 83% PPA and 85% PPA. This indicates that the addition of DNA data improved the RNA-only HER2 predictor. The Concordance index of the RNA-DNA model score and OS for TDXD was 0.64. Characterizing this relationship, we partitioned the TDXD cohort to three equal sized sub-cohorts (33 1/3%) ordered by model score: low, intermediate, and high. Median OS in months for groups was: low 13.6, intermediate 18.5, and high 20.4. High-group patients had significantly better OS on TDXD compared to the low-group (HR = 0.39, log-rank p-value < 0.002). For patients with reported status, HER2-positivity rate in each group respectively was 6%, 16%, 85% (n=78), and HR-positivity in each group was 70%, 68%, 67% (n=156). Together, these findings indicate that HER2 status prediction improved by adding DNA to an RNA base predictor, and the predicted score was positively correlated with OS for TDXD. Citation Format: Kaveri Nadhamuni, Talal Ahmed, Mark Carty, Ben Terdich, Whitney L. Hensing, Timothy Taxter, Calvin Chao, Raphael Pelossof. Identification of poor responders to trastuzumab-deruxtecan with a multi-modal HER2-status predictor [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 2502.

Abstract Background: The prognostic and predictive value of the PAM50 intrinsic subtypes, namely Luminal A, Luminal B, Her2, and Basal-like subtypes, is well-studied in primary as well as metastatic breast cancer settings. Prosigna has emerged as a rapid PAM50 subtype predictor based on the NanoString nCounter assay. However, assay reproducibility across various RNASeq or qRT-PCR platforms can be challenging, especially when applying the predictor on metastatic breast cancer tumors. Here, we used SpinAdapt to create an intrinsic subtype predictor that works on RNA sequencing data, and validates on multiple tumor-sites. We evaluate real-world outcomes for our intrinsic subtype predictions across various immunohistochemical (IHC) labels and metastatic sites. Methods: We trained the subtype predictor on a cohort of 2,497 breast cancer patients using the PAM50 genes, profiled using Nanostring RNA nCounter assay (GSE148426). Approximately 5,423 de-identified records of breast cancer patients sequenced using whole-exome capture RNA-seq were included in our reference dataset. The reference dataset contained samples collected from various sites including breast (n=2440), liver (n=936), lymph node (n=577), lung (n=540), and bone (n=304). For subtype classification, we first batch-corrected the external dataset to Tempus RNA-seq reference dataset using SpinAdapt, then trained a Support-Vector Classifier (SVC) on the corrected data. A 10-fold CV experiment was performed on the corrected dataset to analytically validate the intrinsic subtype predictions. We retrospectively analyzed 7,021 de-identified breast cancer patients with known hormone receptor (HR) or HER2 status and a matched RNASeq sample. The concordance between HR/HER2 status and PAM50 prediction was analyzed, and these patients were excluded from training. Real-world overall survival (rwOS) was evaluated from the time of first diagnosis. The outcomes across intrinsic subtypes were further assessed according to tumor collection site and HR/HER2 IHC status. Results: The 10-fold CV experiment on the Tempus-adapted GSE148426 dataset achieved F-1 scores of 0.97, 0.86, 0.94, and 0.87 on Basal, HER2-like, Luminal A, and Luminal B PAM50 subtypes, respectively. On the Tempus evaluation dataset, 85.3% of HR+/HER2- patients (n=4,366), 65.4% of HR-/HER2+ patients (n=240), and 75.4% of HR-/HER2- patients (n=1,930) were classified as Luminal, HER2-like, and Basal, respectively. Evaluating outcomes on Tempus patients for each PAM50 group, the rwOS for the basal group was significantly shorter than patients not predicted to be basal (n=5,845, p< 2e-90). The rwOS for the predicted PAM50 basal patients remained significantly shorter than non-basal patients even when stratified by site of metastasis: breast (n=2,405; p< 1e-27), lymph node (n=577, p< 1e-7), liver (n=936; p< 1e-25), lung (n=540, p< 1e-13), and bone (n=304, p< 1e-3). Interestingly, within both HR+/HER2- (n=3,653) and HR-/HER2- (n=1,664) IHC cohorts with available outcomes data, the predicted PAM50 basal subtype could further stratify each of these populations with basal-subtype showing significantly worse prognosis than the non-basal subtype (p< 1e-27 and p< 1e-7, respectively). Conclusions: We retrospectively analyzed Tempus multimodal RWD to validate an in-house breast intrinsic subtype predictor that is agnostic to the site of metastasis. The prognostic value of the basal subtype was significant for breast cancer patients across various sites of metastasis including lymph node, liver, lung, and bones and IHC groups. For patients in each of the HR+/HER2- and triple negative IHC groups, the intrinsic molecular subtypes provided an additional level of prognostic detail with statistical significance. These data emphasize the importance of combining molecular subtypes with IHC-based diagnostics to fully characterize clinically relevant subpopulations and risk. Citation Format: Talal Ahmed, Mark Carty, Kaveri Nadhamuni, Raphael Pelossof. Breast cancer intrinsic subtypes predict outcomes in primary and metastatic samples [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO5-24-03.

While clinical oncology societies such as ESMO have recommended the use of Comprehensive Genomic Profiling (CGP) to identify patients eligible for targeted treatment, full utilization of the potential benefits of CGP has not yet occurred in routine clinical practice. Here we assess the compliance to ESMO targeted therapy recommendations and associated outcomes for CGP-based ALK, RET, ROS1, and NTRK fusions detected in a large real-world, observational dataset of advanced NSCLC patients. We retrospectively analyzed de-identified stage IV or metastatic NSCLC records from the Tempus database which encompasses molecular and clinical data from hundreds of clinics across the United States sequenced with the Tempus xT assay (DNA and whole-exome capture RNA NGS) from 2018-2022. Therapeutic adoption was analyzed in cases with ≥30 days of medication data post-sequencing. Real-world overall survival (rwOS) was defined as the interval from start of medication prescribed after sequencing to date of death, censored on the last known physician encounter. A cox proportional hazards model was fit to evaluate the relationship between matched targeted therapy compliance and rwOS in fusion-positive patients. Among 1,950 patients that met study inclusion criteria, N= 65 (3.3%) had a fusion detected by CGP (N=38 ALK fusions, N=15 RET fusions, N=11 ROS1 fusions and N= 1 NTRK fusion). The overall compliance rate of targeted therapy was 82% (N=53). The median time from sequencing to start of targeted therapy was < 1 month. Fusion-positive patients receiving matched therapy had significantly longer rwOS than those that did not receive matched therapy (HR=0.13, p<0.001). ALK-positive patients receiving matched therapy had significantly longer rwOS than those that did not receive matched therapy, (HR=0.12, p<0.05). This study demonstrates that in a real-world, retrospective cohort, most oncologists utilized CGP to timely treat patients with ESMO-recommended targeted therapy for fusion-positive advanced NSCLC. More importantly, CGP-matched guideline-recommended treatment is associated with improved rwOS. Future studies are needed to understand the gap in compliance with matched targeted therapy.

Supplementary Figure from KRAS Mutants Upregulate Integrin β4 to Promote Invasion and Metastasis in Colorectal Cancer

326 Background: HER2-low tumors, defined as a score of 1+ on immunohistochemical (IHC) analysis or as an IHC score of 2+ and negative results on in situ hybridization (ISH), are not well characterized among gastric or gastroesophageal junction adenocarcinomas. More importantly, this patient population is currently treated as HER2-negative (HER2-low and HER2-zero [IHC score of 0]) without anti-HER-2 directed therapy. Herein, we describe a real world outcome analysis of this distinct genomically defined underrepresented population. Methods: De-identified, multimodal real-world data (RWD) of 429 advanced gastric, esophageal, and GEJ adenocarcinomas from the Tempus Lens database was analyzed. Inclusion criteria was a diagnosis of gastric, esophageal, or GEJ adenocarcinoma between January 2017 to May 2021 and treatment with first-line therapy. Median overall survival (mOS) was estimated using Kaplan-Meier methods. Results: HER2-low was identified in 24.9% (107/429) of the total cohort. Histologies within HER2 low group were 83.2% (89/107) adenocarcinoma (NOS), 7.5% (8/107) signet ring cell carcinoma, 5.6% (6/107) diffuse-type, and 3.7% (4/107) mucinous-type. PDL-1 TPS status was available in only 50.5% (54/107) of the HER2-low group. TPS distribution was: 74.1% (40/54) for TPS <1, 22.2% (12/54) for TPS >=1-<10, and TPS >= 10: 3.7% (2/54). ERBB2 amplification or focal gain was detected in 1.1% (1/88) of patients in the HER-2 low group that had undergone tumor genomic sequencing, and 0% (0/28) of patients that had circulating tumor DNA (ctDNA) sequenced. Median followup time was 20.7 (8.5, 35.6) months. Median OS (mOS, months) was 9.8 (9.0, 12.7) for HER2-low, 13.2 (8.9, 27.0) for HER2(-), and 14.1 (11.0, 19.2) for HER2(+). Survival differences were also compared with variations across ERBB2 RNA expression and copy number variation (CNV). HER2 groups were defined on RNA expression and copy number estimates based on thresholds determined from the proportion of samples in each of the HER2 groups from IHC/ISH data. For RNA expression analysis, the mOS was 10.8 (8.8, 13.5) months for HER2-low, 10.9 (7.2, 17.4) months for HER2(-), and 13.0 (11.4, 20.5) months for HER2(+). For CNV data, the mOS(months) was: HER2-low: 10.9 (8.5, 14.1), HER2(-): 10.2 (7.0, 15.7), HER2(+): 14.2 (11.5, 21.4). Conclusions: Our real world, biomarker outcome analysis suggests that advanced gastric and GEJ patients with HER2-low expression identified by IHC and RNA expression have a poor prognosis with shortest survival. Furthermore, our data suggest that there is a heterogeneity of PDL-1 enrichment within this group. Hence, targeting expression of low levels of HER2 similar to HER2+ group alone or in combination with immunotherapy should be evaluated in future clinical trials to improve efficacy outcomes.

Abstract KRAS mutation in colorectal cancer is associated with aggressive tumor behavior through increased invasiveness and higher rates of lung metastases, but the biological mechanisms behind these features are not fully understood. In this study, we show that KRAS-mutant colorectal cancer upregulates integrin α6β4 through ERK/MEK signaling. Knocking-out integrin β4 (ITGB4) specifically depleted the expression of integrin α6β4 and this resulted in a reduction in the invasion and migration ability of the cancer cells. We also observed a reduction in the number and area of lung metastatic foci in mice that were injected with ITGB4 knockout KRAS-mutant colorectal cancer cells compared with the mice injected with ITGB4 wild-type KRAS-mutant colorectal cancer cells, while no difference was observed in liver metastases. Inhibiting integrin α6β4 in KRAS-mutant colorectal cancer could be a potential therapeutic target to diminish the KRAS-invasive phenotype and associated pulmonary metastasis rate. Implications: Knocking-out ITGB4, which is overexpressed in KRAS-mutant colorectal cancer and promotes tumor aggressiveness, diminishes local invasiveness and rates of pulmonary metastasis.

The incidence of rectal cancer is increasing in patients younger than 50 years. Locally advanced rectal cancer is still treated with neoadjuvant radiation, chemotherapy and surgery, but recent evidence suggests that patients with a complete response can avoid surgery permanently. To define correlates of response to neoadjuvant therapy, we analyzed genomic and transcriptomic profiles of 738 untreated rectal cancers. APC mutations were less frequent in the lower than in the middle and upper rectum, which could explain the more aggressive behavior of distal tumors. No somatic alterations had significant associations with response to neoadjuvant therapy in a treatment-agnostic manner, but KRAS mutations were associated with faster relapse in patients treated with neoadjuvant chemoradiation followed by consolidative chemotherapy. Overexpression of IGF2 and L1CAM was associated with decreased response to neoadjuvant therapy. RNA-sequencing estimates of immune infiltration identified a subset of microsatellite-stable immune hot tumors with increased response and prolonged disease-free survival.

Reproducibility of results obtained using ribonucleic acid (RNA) data across labs remains a major hurdle in cancer research. Often, molecular predictors trained on one dataset cannot be applied to another due to differences in RNA library preparation and quantification, which inhibits the validation of predictors across labs. While current RNA correction algorithms reduce these differences, they require simultaneous access to patient-level data from all datasets, which necessitates the sharing of training data for predictors when sharing predictors. Here, we describe SpinAdapt, an unsupervised RNA correction algorithm that enables the transfer of molecular models without requiring access to patient-level data. It computes data corrections only via aggregate statistics of each dataset, thereby maintaining patient data privacy. Despite an inherent trade-off between privacy and performance, SpinAdapt outperforms current correction methods, like Seurat and ComBat, on publicly available cancer studies, including TCGA and ICGC. Furthermore, SpinAdapt can correct new samples, thereby enabling unbiased evaluation on validation cohorts. We expect this novel correction paradigm to enhance research reproducibility and to preserve patient privacy.

Reproducibility of results obtained using RNA data across labs remains a major hurdle in cancer research. Often, molecular predictors trained on one dataset cannot be applied to another due to differences in RNA library preparation and quantification. While current RNA correction algorithms may overcome these differences, they require access to all patient-level data, which necessitates the sharing of training data for predictors when sharing predictors. Here, we describe SpinAdapt, an unsupervised RNA correction algorithm that enables the transfer of molecular models without requiring access to patient-level data. It computes data corrections only via aggregate statistics of each dataset, thereby maintaining patient data privacy. Furthermore, SpinAdapt can correct new samples, thereby enabling evaluation of validation cohorts. Despite an inherent tradeoff between privacy and performance, SpinAdapt outperforms current correction methods that require patient-level data access. We expect this novel correction paradigm to enhance research reproducibility and patient privacy. Finally, SpinAdapt lays a mathematical framework that can be extended to other -omics modalities.

Somatic mutations in the KRAS oncogene are associated with poor outcomes in locally advanced rectal cancer but the underlying biologic mechanisms are not fully understood. We profiled mRNA in 76 locally advanced rectal adenocarcinomas from patients that were enrolled in a prospective clinical trial and investigated differences in gene expression between KRAS mutant (KRAS‐mt) and KRAS ‐wild‐type (KRAS‐wt) patients. We found that KRAS‐mt tumors display lower expression of genes related to the tumor stroma and remodeling of the extracellular matrix. We validated our findings using samples from The Cancer Genome Atlas (TCGA) and also by performing immunohistochemistry (IHC) and immunofluorescence (IF) in orthogonal cohorts. Using in vitro and in vivo models, we show that oncogenic KRAS signaling within the epithelial cancer cells modulates the activity of the surrounding fibroblasts in the tumor microenvironment.

e13507 Background: Recent advances in transcriptomics have resulted in the emergence of several publicly available breast cancer RNA-Seq datasets, such as TCGA, SCAN-B, and METABRIC. However, molecular predictors cannot be applied across datasets without the correction of batch differences. In this study, we demonstrate a homogenization algorithm that allows the transfer of molecular subtype predictors from one RNA-Seq cohort to another. The algorithm only uses cohort-level RNA-Seq summary statistics, and therefore, does not require joint normalization of both datasets nor the transfer of patient information. Using this approach, we transferred a breast cancer subtype (Luminal A, Luminal B, HER2+, Basal) predictor trained on SCAN-B data to accurately predict subtypes from TCGA. Methods: First, we randomly split the TCGA cohort (n = 481 Luminal A, n = 189 Luminal B, n = 73 Her2+, n = 168 Basal) into two sets: TCGA-train and held-out TCGA-test (n = 455 and n = 456, respectively). Second, the SCAN-B cohort (n = 837) was homogenized with the TCGA-train set. Third, a molecular subtype predictor, based on a logistic regression model, was trained on homogenized SCAN-B RNA-Seq samples and used to predict the subtypes of TCGA-test RNA-Seq samples. For baseline comparison, a similar predictor trained on the non-homogenized SCAN-B cohort was tested on the TCGA-test set. The experimental framework was iterated 250 times. Reported P-values reflect a paired one-sided t-test. Results: To quantify model performance, we measured the average F1 score for each tumor subtype prediction from the held-out TCGA test set with and without cohort homogenization. The average F1 scores with vs. without homogenization were: Luminal A, 0.88 vs. 0.85 ( P< 1e-69); Luminal B, 0.74 vs. 0.51 ( P< 1e-183); Her2+, 0.73 vs. 0.53 ( P< 1e-99); Basal, 0.98 vs. 0.97 ( P< 1e-53). Overall, homogenization significantly outperformed no homogenization. Conclusions: We developed a novel homogenization algorithm that accurately transfers subtype predictors across diverse, independent breast cancer cohorts.

3081 Background: Tumors of unknown origin occur in approximately 5% of newly diagnosed cancers and are difficult to treat without establishing the tissue type from which they derive. Establishing tumor origin guides standard of care treatment for several NCCN targeted therapy guidelines. Leveraging tissue specificity in gene expression profiles, classification models based on RNA expression offer a promising approach to identify the likely primary cancer site in tumors of unknown origin. Methods: In this study, we developed a transcriptome-based cancer type classifier trained on over 10,000 tissue samples annotated by pathologists and sequenced for RNA expression to identify conserved patterns of expression characteristic of 30 tumor types across primary and metastatic tissue sites. The classifier probabilistically ranks cancer of origin. Results: Overall, the accuracy of the most probable cancer prediction was 85%, 88% within primary tumors and 77% within metastatic tumors. The top three cancers types with the highest accuracy were colorectal (accuracy in metastatic: 93%, accuracy in primary tumors: 99%), breast (95%, 96%) and lung (87%, 94%). Classifier performance was lower in low-purity metastatic tumors where the surrounding normal tissue obscures the tumor transcriptional profile, though the classifier still achieves 71% accuracy on metastatic tumors with less than 50% purity. Conclusions: We present a novel method to probabilistically predict tumor type for cancers of unknown origin using RNA-Seq. Our method achieves robust classification that is applicable to primary and metastatic tumors and demonstrates the value of utilizing RNA-Seq to aid cancer diagnosis and treatment decisions.

Rectal cancer (RC) is a challenging disease to treat that requires chemotherapy, radiation and surgery to optimize outcomes for individual patients. No accurate model of RC exists to answer fundamental research questions relevant to patients. We established a biorepository of 65 patient-derived RC organoid cultures (tumoroids) from patients with primary, metastatic or recurrent disease. RC tumoroids retained molecular features of the tumors from which they were derived, and their ex vivo responses to clinically relevant chemotherapy and radiation treatment correlated with the clinical responses noted in individual patients' tumors. Upon engraftment into murine rectal mucosa, human RC tumoroids gave rise to invasive RC followed by metastasis to lung and liver. Importantly, engrafted tumors displayed the heterogenous sensitivity to chemotherapy observed clinically. Thus, the biology and drug sensitivity of RC clinical isolates can be efficiently interrogated using an organoid-based, ex vivo platform coupled with in vivo endoluminal propagation in animals.

Rectal cancer (RC) is a challenging disease to treat that requires chemotherapy, radiation, and surgery to optimize outcomes for individual patients. No accurate model of RC exists to answer fundamental research questions relevant to individual patients. We established a biorepository of 32 patient-derived RC organoid cultures (tumoroids) from patients with primary, metastatic, or recurrent disease. RC tumoroids retained molecular features of the tumors from which they were derived, and their ex vivo responses to clinically relevant chemotherapy and radiation treatment correlate well with responses noted in individual patients’ tumors. Upon engraftment into murine rectal mucosa, human RC tumoroids gave rise to invasive rectal cancer followed by metastasis to lung and liver. Importantly, engrafted tumors closely reflected the heterogenous sensitivity to chemotherapy observed clinically. Thus, the biology and drug sensitivity of RC clinical isolates can be efficiently interrogated using an organoid-based, in vitro platform coupled with endoluminal propagation in animals.

OBJECTIVE:To investigate associations between genetic mutations and qualitative as well as quantitative features on MRI in rectal adenocarcinoma at primary staging. METHODS:In this retrospective study, patients with rectal adenocarcinoma, genome sequencing, and pretreatment rectal MRI were included. Statistical analysis was performed to evaluate associations between qualitative features obtained from subjective evaluation of rectal MRI and gene mutations as well as between quantitative textural features and gene mutations. For the qualitative evaluation, Fisher's Exact test was used to analyze categorical associations and Wilcoxon Rank Sum test was used for continuous clinical variables. For the quantitative evaluation, we performed manual segmentation of T2-weighted images for radiomics-based quantitative image analysis. Thirty-four texture features consisting of first order intensity histogram-based features (n = 4), second order Haralick textures (n = 5), and Gabor-edge based Haralick textures were computed at two different orientations. Consensus clustering was performed with 34 computed texture features using the K-means algorithm with Euclidean distance between the texture features. The clusters resulting from the algorithm were then used to enumerate the prevalence of gene mutations in those clusters. RESULTS:In 65 patients, 45 genes were mutated in more than 3/65 patients (5%) and were included in the statistical analysis. Regarding qualitative imaging features, on univariate analysis, tumor location was significantly associated with APC (p = 0.032) and RASA1 mutation (p = 0.032); CRM status was significantly associated with ATM mutation (p = 0.021); and lymph node metastasis was significantly associated with BRCA2 (p = 0.046) mutation. However, these associations were not significant after adjusting for multiple comparisons. Regarding quantitative imaging features, Cluster C1 had tumors with higher mean Gabor edge intensity compared with cluster C2 (θ = 0°, p = 0.018; θ = 45°, p = 0.047; θ = 90°, p = 0.037; cluster C3 (θ = 0°, p = 0.18; θ = 45°, p = 0.1; θ = 90°, p = 0.052), and cluster C4 (θ = 0°, p = 0.016; θ = 45°, p = 0.033; θ = 90°, p = 0.014) suggesting that the cluster C1 had tumors with more distinct edges or heterogeneous appearance compared with other clusters. CONCLUSIONS:Although this preliminary study showed promising associations between quantitative features and genetic mutations, it did not show any correlation between qualitative features and genetic mutations. Further studies with larger sample size are warranted to validate our preliminary data.

609 Background: KRAS-mutant (KRASmut) colorectal cancers (CRCs) are associated with worse prognosis and resistance to therapy. We have previously shown that KRASmut CRCs have different transcriptomic signature of stromal and immune-related genes compared to KRAS-wild type (KRASwt) tumors. Here, we validated the immune-related changes in the tumor microenvironment associated with the KRAS mutation in CRC to guide the design of novel immunotherapy strategies. Methods: The expression of different immune markers (T cells, B cells, macrophages, natural killer cells, and immune check ligands) were assessed using multiplex immunofluorescence (M-IF) technique in both tumor core (TC) and invasive margin (IM). Sequential slides were cut from paraffin blocks of CRC resected at our institute. Each slide was stained with 4 immune markers using M-IF technique. The stained slides were scanned, and quantification of immune cells was done using ImageJ software. Student’s t test was used for statistical analyses. DNA was extracted from each tumor and profiled for 420 cancer genes using targeted exome-capture sequencing (MSK-IMPACT assay). DNA mismatch-repair (MMR) proteins deficiency were analyzed by immunohistochemistry. Only MMR-proficient (pMMR) tumors were included. Results: A total of 39 patients with pMMR CRC were included. AJCC stages (I-III) were not different between KRASmut (n = 15) and KRASwt (n = 25) tumors. M2-macrophages (CD68+CD163+ cells) and IL-17-producing cells (IL17+ cells) were significantly higher (p = 0.002, and 2.9e-6 respectively), while T-helper cells (CD3+CD4+cells) were significantly lower (p = 3.9e-4) in TC of KRASmut tumors compared to KRASwt. Treg (CD3+CD4+FOXP3+cells) were significantly higher in IM of KRASmut tumors (p = 0.01). KRASmut tumors had significantly higher ratios of Treg:T-helper cells, and Treg:T cytotoxic cell (p = 0.008, and p = 0.04; respectively). Conclusions: KRAS oncogene is associated with more pro-tumorigenic (M2 macrophages, IL17 and Treg) and less anti-tumorigenic (CD4 T-helper) immune cells in CRC. These results can be used to guide further research to design novel immunotherapy strategies against KRASmut CRC.