Abstract Background: Immunotherapy offers a promising therapeutic option against pediatric tumors, but it is limited by suppression of the antigen processing machinery (APM), including major histocompatibility complex class I (MHC-I). In this study, we quantify antigen presentation in pediatric tumors and explore the role of transcriptional regulators and epigenetic-targeted therapies to restore antigen presentation and immune recognition in rhabdomyosarcoma (RMS). Methods: Expression of HLA-A/B/C was quantified by RNA sequencing (RNA-seq) in adult tumors (n=657) and pediatric cell lines (n=131). Surface MHC-I expression was quantified by flow cytometry in pediatric cell lines (n=76). RMS cell lines (n=9) were treated with IFN-γ and clinically relevant drugs, decitabine (DAC), mocetinostat, and tazemetostat. Changes in MHC-I and APM gene expression were measured by RNA-seq and flow cytometry. NLR family CARD domain-containing 5 (NLRC5) was induced in RMS cell line and patient derived xenograft (PDX) models using lentiviral overexpression or CRISPR activation. MHC-I and APM expression were measured by RNA-seq, flow cytometry, and western blot. T cell cytotoxicity assays were performed with T cells expressing a PRAME-specific HLA-A*02:01 T cell receptor (TCR). Results: Pediatric tumors exhibited variable MHC-I expression, as determined by flow cytometry, and this expression was significantly lower than adult tumors, with RMS displaying low or absent expression. IFN-γ and pharmacologic treatment increased MHC-I surface expression and immune gene signatures in RMS. NLRC5, a key immune regulator, was found to be most significantly correlated with MHC-I expression and induced by treatment with DAC. Epigenetic priming with DAC and upregulation of NLRC5 was sufficient to restore MHC-I expression and sensitized an RMS PDX to killing by engineered TCR-T cells targeting PRAME. Conclusions: Pediatric tumors show distinct patterns of antigen presentation, such as high MHC-I in alveolar soft part sarcoma and low expression in RMS, though individual tumor subtypes exhibit internal variability. IFN-γ and epigenetic agents restore antigen presentation, including increased NLRC5 expression. Pharmacological treatment and NLRC5 restoration enhanced antigen presentation and sensitized RMS PDX to TCR mediated T-cell cytotoxicity. Compounds that reverse APM silencing will be systematically evaluated for their ability to enhance adoptive TCR therapies. This work will establish a foundation for overcoming immune resistance and expanding the impact of MHC-I dependent cancer immunotherapies. Citation Format: Maya Groff, David Milewski, Hsien-Chao Chou, Vineela Gangalapudi, Alexandra Urbanek, Young Song, Meijie Tian, Yong Yean Kim, Jun Wei, Javed Khan. Enhancement of antigen presentation restores immune recognition in rhabdomyosarcoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 645.

Abstract Introduction: Rhabdomyosarcoma (RMS) is a highly malignant soft tissue sarcoma. An aggressive subtype, fusion-positive RMS (FP-RMS), which is driven by PAX3/7::FOXO1 translocations, has a dismal 5-year overall survival rate of 13% for patients with metastatic disease. We hypothesized that targeting the oncogenic driver PAX3::FOXO1 with small molecules would be an effective treatment. To test this, we developed cell lines with endogenous PAX3::FOXO1 tagged with HiBiT epitope and performed a drug screen to identify drugs that downregulate PAX3::FOXO1 protein. Study Design: Using CRISPR-Cas9, we endogenously tagged PAX3::FOXO1 with HiBiT in two RMS cell lines (RH4 and SCMC), enabling the monitoring of PAX3::FOXO1 protein levels. We performed a drug screen using the Mechanism Interrogation Plate (MIPE) library of 2,480 compounds, of which 53% are FDA-approved or in clinical trials. NanoGlo Luciferase assays monitored levels of HiBiT-tagged PAX3::FOXO1, while CellTiterGlo measured cell viability at 24 hours. We selected hits that showed a difference in area under the curve (AUC) between the two readouts of ≥ 90 for drugs that preferentially reduce the fusion protein level over general cytotoxicity. We investigated whether inhibitors led to nuclear accumulation and reduced total protein levels using Western blot and immunofluorescent imaging. Candidates were validated in the parental cells and in vivo studies. Results and Conclusions: The screen identified 183 hits, including Eltanexor, an XPO1 inhibitor. XPO1 exports over 200 proteins from the nucleus by recognizing their nuclear export sequences (NESs). Since the fusion gene retains the NES of FOXO1, a known XPO1 target, we tested whether PAX3::FOXO1 is a substrate. We observed that Eltanexor enhanced PAX3::FOXO1 nuclear accumulation at 6 hours in RH4, and at 2 hours in SCMC, followed by protein downregulation at 24 hours by Western blotting. Furthermore, Eltanexor induced p53 nuclear accumulation, detectable at 6 hours in SCMC, suggesting that early accumulation of PAX3::FOXO1 may drive cytotoxicity. At 24 hours, RNA-seq in Eltanexor-treated cell lines demonstrated downregulation of PAX3::FOXO1 and MYCN signatures, components of the core regulatory network in FP-RMS. Preliminary in vivo studies also showed Eltanexor induces delays in tumor progression in an RMS xenograft model. Furthermore, Eltanexor in combination with Mivebresib, a validated BRD4 inhibitor, demonstrated significant synergy against FP-RMS cells. We will perform site-directed mutagenesis studies to disrupt PAX3::FOXO1’s NES and validate combination with Mivebresib in vivo. In conclusion, we identified Eltanexor, an XPO1 inhibitor, as a novel therapeutic agent that suppressed PAX3::FOXO1 activity and levels, induced nuclear accumulation and led to cytotoxicity in incurable FP-RMS. Citation Format: Soumili Dey, Yong Y. Kim, Katrina Jia, Mehal Churiwal, Michele Ceribelli, Teresa S. Hawley, Raj Chari, David Milewski, Young K. Song, Xinyu Wen, Hsien-Chao Chou, Vineela Gangalapudi, Jun S. Wei, Craig Thomas, Robert G. Hawley, Javed Khan. Small-molecule screening of HiBiT-tagged PAX3::FOXO1 rhabdomyosarcoma cell lines identifies eltanexor as a potent therapeutic agent against fusion-positive rhabdomyosarcoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 6405.

Abstract Prenatal cell-free (cf) DNA sequencing for fetal aneuploidy incidentally detects circulating tumor DNA in the plasma of asymptomatic pregnant women. Distinguishing the subset of women with malignant tumors detected by prenatal sequencing from those with benign conditions, such as uterine fibroids, is critical to maternal medical management. In this pilot study, prospectively collected blood samples from 65 pregnant or postpartum women with and without occult cancers were analyzed blindly for cfDNA somatic mutations in 275 cancer-associated genes. Somatic variants were common among all 65 women, however, when stringent mutation analysis criteria were applied, these data could independently differentiate women with cancer with a sensitivity of 80.6% and specificity of 100%. Mutation profiling complements radiographic imaging by clarifying tumor origin, evaluating malignancy in indeterminate cases, identifying actionable genomic alterations, and flagging high-risk patients for urgent intervention. These findings provide preliminary evidence that cfDNA somatic mutations could serve as an additional noninvasive biomarker of malignancy potentially aiding the management of women with prenatal cfDNA findings suspicious for cancer. Citation Format: Zhigang Kang, Amy E. Turriff, Yuelin Jack Zhu, Erica Pehrsson, Hsein-Chao Chou, Jun Wei, Kerstin Heselmeyer-Haddad, Paul S. Meltzer, Javed Khan, Liang Cao, Diana W. Bianchi. Evaluation of somatic mutations in cell-free DNA as noninvasive biomarkers of cancer in asymptomatic pregnant women [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 7840.

PURPOSE:Medullary thyroid carcinoma (MTC) is often driven by activating mutations in the RET receptor tyrosine kinase. Multikinase and selective inhibitors targeting RET are highly effective for RET-mutant MTC, but acquired resistance is commonly observed, limiting clinical efficacy. MATERIALS AND METHODS:We performed a comprehensive genomic and pharmacological analysis of acquired resistance in a previously described in vitro model of RET-mutant MTC generated from long-term treatment with the RET inhibitor vandetanib. Molecular studies using spectral karyotyping, multiplex interphase fluorescence in situ hybridization, whole-exome sequencing, and RNA sequencing revealed several mechanisms of acquired resistance. Whole-genome CRISPR knockout screening was performed to identify potential genes mediating intrinsic resistance. High-throughput drug screening was used to identify additional therapeutic targets. The combination of RET and MEK inhibitors was evaluated in preclinical animal models. RESULTS:Genomic profiling revealed that resistant MTC cells acquired RET copy-number gain and the clinically observed secondary RET mutation p.G810S. Whole-genome CRISPR knockout screening on MTC cells treated with two different RET inhibitors highlighted that NF1 deletion and subsequent RAS/MAPK activation were sufficient to establish resistance to RET inhibition. High-throughput drug screening also indicated that MTC cells are sensitive to RAS/MAPK inhibition, particularly in combination with RET inhibitors. The combination of RET and MEK inhibitors was synergistic in both RET-inhibitor-naïve and resistant MTC in mouse xenograft models. CONCLUSION:Resistance to RET inhibitors can be acquired through RET copy-number gain and secondary mutations as well as NF1 loss-mediated MAPK pathway activation. This mechanism of resistance can be overcome with dual inhibition of RET and downstream RAS/MAPK signaling, demonstrating clinical potential in RET-mutant MTC.

Abstract Background: Pediatric tumors often co-opt normal developmental gene-regulatory programs, with errors in lineage-restricted progenitors that halt or reverse differentiation. Because these cancers arise within restricted developmental windows, display fetal-like programs, and carry relatively few driver mutations compared to adult tumors, we hypothesized that a pan-pediatric, transcriptome-inferred gene-regulatory network (GRN) analysis will discover lineage-specific regulons that anchor each tumor to a developmentally arrested state, which would identify actionable biomarkers and therapeutic targets. Methods: We analyzed 2541 bulk RNA-seq from 35 pediatric cranial and extracranial solid-tumor samples, after batch correction. We inferred a pan-pediatric GRN from gene expression data, integrating networks inferred by ARACNe-AP and GENIE3 into a consensus GRN across all tumor types. We used a one-vs-rest strategy to identify tumor-specific differentially expressed genes (DEGs) within the regulons. Using hypergeometric tests, we quantified transcription factor (TF) activity and their regulons across tumors by assessing the enrichment of tumor-specific DEGs within each regulon. To map genes to drugs, we queried drug libraries, including Mechanistic Interrogation PlatE, Profiling Relative Inhibition Simultaneously in Mixtures, ChEMBL, DrugBank, and DrugCentral. We filtered druggable genes among TFs, their regulon members, and their interactors using log fold change and adjusted p-values, and ranked candidates in 19 tumors with DepMap data by using effect size. Results: We identified 281 enriched TFs across tumors. The functional enrichment analyses showed that TF programs are usually restricted to specific tumor classes, mirroring their developmental cell-of-origin and highlighting candidate tumor-specific biomarkers. Examples include neurodevelopmental and neural-crest-related TFs (e.g., PHOX2B, ASCL1, and SOX10) in neuroblastoma (NB) and muscle-lineage TFs (e.g., MYOG, MYOD1, and PAX3/7) in fusion-positive rhabdomyosarcomas (FP-RMS). Our analysis suggests that TFs behave as robust, tumor-type-specific expression signatures and can distinguish tumors that may be histologically similar but arise from different developmental lineages. Furthermore, we used our TF-centric approach to identify known and new drug targets, such as SIX1, RRM2, AURKA, and BIRC5 in FP-RMS, and ACVR2B & BMPR1B in NB. Conclusions and Future Directions: A unified GRN framework analysis of pan pediatric solid tumors resolves lineage-specific regulons associated with tumorigenesis and yields a ranked set of druggable genetic dependencies. In vitro and in vivo validation studies are currently underway. Citation Format: Daniel Lee, Abid A. Reza, Syed A. Bukhari, Jun S. Wei, Hsein-Chao Chou, Xinyu Wen, Andrew S. Brohl, Javed Khan. A pan-pediatric gene-regulatory network analysis reveals druggable dependencies across pediatric solid tumors [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 4096.

Abstract Background: Rhabdomyosarcoma (RMS) is a highly malignant pediatric soft-tissue sarcoma where molecular subtyping, particularly PAX3/7::FOXO1 fusion status, drives prognosis and treatment. However, histology-based diagnostic approaches remain limited by subjectivity and the scarcity of comprehensive molecular annotations. To overcome these challenges, we improved our previously reported convolutional neural network learning models that predict PAX3/7::FOXO1 fusion status from whole-slide images (WSIs) while additionally trained models to infer gene expression profiles from histology, thereby linking morphology to transcriptomic signatures. Methods: A total of 826 independent WSIs from three sources [Children’s Oncology Group (COG) biobanking protocols = 322, Kids First (KIDS) = 252, Childhood Cancer Data Initiative/Molecular Characterization Initiative (CCDI/MCI) = 252] were used to train and evaluate an Attention-Based Multiple Instance Learning (ABMIL) model using UNI2-h foundation features for fusion classification. For gene expression prediction, 135 RMS WSIs paired with bulk RNA-seq data were used to fine-tune a SEQUOIA transformer model, which was trained on TCGA UCEC/COAD datasets. Model performance was evaluated using the Matthews Correlation Coefficient (MCC), AUC, and Pearson's r correlation, with biological validation through pathway enrichment analysis. Results: The fusion detection model achieved robust and generalizable performance across independent test cohorts (MCC ≥ 0.80, AUC ≥ 0.94), with multi-institutional training improving external generalization (MCC up to 0.84). The gene expression model reliably predicted bulk transcriptomic profiles from WSIs (mean r > 0.6, p < 0.05), identifying biologically meaningful pathways including cell cycle and muscle development. Together, these models demonstrate the feasibility of integrating morphological imaging data to gain molecular insights that would not be possible with histology alone. Conclusions: This work presents the first large-scale validated deep learning framework for simultaneous molecular subtyping and transcriptomic inference in RMS. By combining digital pathology with molecular prediction, our approach offers a scalable, tissue-sparing, and generalizable tool for advancing precision oncology in RMS. Citation Format: Dorsa Ziaei, Hyun Jung, Philip J. Lupo, Pagna Sok, Jack F. Shern, Corinne M. Linardic, Syed Abbas Bukhari, Hsein-Chao Chou, Jun S. Wei, Curtis Lisle, Uma Mudunuri, Javed Khan. PAX3/7::FOXO1 fusion detection and transcriptomic prediction from whole-slide images of rhabdomyosarcoma using attention-based deep learning frameworks: A multi-institutional study [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 2758.

Compared with chimeric antigen receptor (CAR) T cell therapy, antibody-drug conjugates (ADCs) offer distinct advantages. Here, we report on two FGFR4-targeted ADCs, using the high-affinity monoclonal antibody 3A11, the same binder used in a CAR T cell format that is being evaluated at the NCI (NCT06865664) for patients with relapsed/refractory rhabdomyosarcoma (RMS). These ADCs, conjugated to monomethyl auristatin E (MMAE) or an exatecan derivative, are rapidly internalized, causing potent fibroblast growth factor receptor 4 (FGFR4)-dependent cytotoxicity in vitro. In subcutaneous RMS xenograft models, both ADCs demonstrated robust anti-tumor activities, significantly prolonging survival. Notably, exatecan-ADC showed better efficacy, with durable tumor control, in aggressive fusion-negative (FN) RMS559 and fusion-positive (FP) RH4 RMS cell-line-derived xenografts (CDXs) and a patient-derived RMS xenograft (PDX). Furthermore, exatecan-ADC effectively controls tumors in an FGFR4-expressing MDA-MB-453 breast cancer mouse model, eradicating relapsed tumors with retreatment. These findings highlight FGFR4-targeted ADCs as potent therapeutic agents against aggressive FGFR4-expressing malignancies, supporting their further clinical development.

Abstract Originating from the neural crest, neuroblastoma is the most common extracranial solid tumor in children. Infiltrating immune cells contribute to the tumor’s growth and treatment response, as patients with high-risk disease, associated with MYCN amplification, benefit from anti-GD2 antibody immunotherapy. However, MYCN-amplified disease remains lethal in more than half of cases and has been associated with immunosuppression in bulk RNA studies. We thus hypothesize that MYCN activation, in conjunction with other molecular and clinical traits, influences the tumor microenvironment (TME), which can either support or suppress the disease. To comprehensively characterize the neuroblastoma TME, we performed CO-Detection by indEXing (CODEX), a multiplex immunohistochemistry technique, on 5 clinically annotated tissue microarrays containing 371 neuroblastic tumors from 179 patients representing all major disease subgroups and treatment protocols. In a subset of specimens, we also applied Visium HD spatial transcriptomics to identify regional malignant programs and their associated immune infiltrates. In parallel, we developed a novel natural language processing approach to detect generalizable spatial cell networks across these tissues. Interrogating more than 40 tumor-, immune-, and stroma-associated proteins revealed that neuroblastomas, despite downregulating MHC class I (MHC-I), harbor rich TMEs composed of about 20 phenotypically distinct cell populations, including multiple lymphoid- and myeloid-derived subsets. MYCN-amplified tumors are profoundly deficient in infiltrating helper, memory, and cytotoxic T cell lineages, whereas they form prominent tertiary structures in non-amplified disease. By contrast, antigen-presenting suppressor-like myeloid cells dominate the MYCN-amplified microenvironment, where they persist in chemotherapy-resistant tumors. These subtype-specific differences prompted functional studies of intrinsic immune and cytokine programs. RNA-seq of multiple human MYCN-amplified cell lines revealed that the differentiation therapy retinoic acid, while suppressing MYCN, drastically upregulated class I antigen presentation and pro-inflammatory cytokines. To further understand MYCN- and treatment-related changes in the inflammatory secretome, we are currently performing extracellular proteomics on these cell lines. Integrative spatial profiling by CODEX and Visium HD reveals that MYCN amplification fosters a T cell-poor, myeloid-rich microenvironment, in contrast to the organized lymphoid structures characteristic of non-amplified neuroblastoma. Together with the finding that retinoic acid restores MHC-I and pro-inflammatory programs in MYCN-amplified tumors, these results suggest that combining retinoic acid with cellular immunotherapies and myeloid-targeting approaches may significantly improve survival in patients with high-risk neuroblastoma. Citation Format: Joseph Seamus Toker, Katherine Elizabeth Masih, Noemi Kedei, Zahin Islam, Ben J. Somerville, Amir Jassim, Michail Mamalakis, Aysen Yuksel, Daniel R. Catchpoole, Li Zhou, Paul Aiyetan, Yong Yean Kim, David Milewski, Shaoli Das, Xinyu Wen, Yong Song, Jun Wei, Richard J. Gilbertson, Javed Khan. Distinct microenvironments define subtypes of neuroblastoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 639.

A high-throughput screening campaign designed to discover natural product inhibitors of rhabdomyosarcoma oncogene PAX3-FOXO1 gene expression identified partially purified fractions from an organic extract of the plant Gonystylus borneensis to have potent activity in the assay. Bioassay-guided isolation yielded five new 5,6-dihydro-α-pyrone natural products, namely, gonystylones A-E (1-5). Their structures were elucidated using 1D and 2D NMR experiments and their absolute configurations determined using semisynthetic and electronic circular dichroism methods. The gonystylones were found to be cytotoxic to rhabdomyosarcoma cells at low micromolar concentrations (3-19 μM).

Supplemental Description of the Rhabdomyosarcoma Web-based Application

Prediction of RAS pathway mutations using a trained CNN. A, Workflow for deep learning of RAS pathway mutations from FN-RMS WSIs. B and C, Representative (B) H&E images and (C) class activation maps of a RAS pathway wild-type tumor and a tumor with a KRAS p.G12C mutation (VAF = 0.659). D, Confusion matrix for predictions on a test dataset. Micro F1, Macro F1, and Matthew's correlation coefficient shown below. E, Statistics for confusion matrix. F, Average ROC curve from holdout test data.

Supplemental Figure S4. Sample partitioning for training a MYOD1 mutation predictive model using K-fold cross-validation.

Rhabdomyosarcoma (RMS) is the most common pediatric sarcoma, representing 3-4% of childhood and adolescent cancers. While multimodal therapies improved the outcomes for localized disease, 5-year survival for relapsed or metastatic RMS cases remains poor. Antibody-drug conjugates (ADCs) use the specificity of monoclonal antibodies to selectively deliver potent anticancer chemotherapy agents to tumor cells while sparing healthy tissues. FGFR4, a cell-surface receptor tyrosine kinase highly expressed in RMS and other cancers such as some breast cancers, but minimally in normal tissues, is a promising immune target. We hypothesize that FGFR4-targeted ADCs could effectively treat RMS and other FGFR4-positive cancers with limited systemic toxicity. High-affinity human FGFR4-specific binders were developed and evaluated for their internalization in RMS cell lines. The top candidate 3A11, a murine monoclonal antibody, was conjugated to either monomethyl auristatin E (MMAE) via a cathepsin cleavable mcValCit-PABC linker (3A11-MMAE), or an Exatecan derivative via a legumain-cleavable mpGlyAsnAsn linker (3A11-Exatecan). ADCs’ in-vitro efficacy was assessed in FGFR4-positive or FGFR4-negative cells using a live-cell analysis system. Western blots were performed to validate their action mechanisms. Finally, ADCs' in-vivo efficacy was tested in 3 subcutaneous xenograft models: fusion-positive RMS (RH4), fusion-negative RMS with an FGFR4 V550L activating mutation (RMS559), and FGFR4-positive breast cancer (MDA-MB-453). 3A11 was internalized by FGFR4-positive cells and the internalization efficiency was significantly correlated with FGFR4 surface expression levels. Both ADCs selectively killed FGFR4-expressing cells, with their potency correlating with FGFR4 expression and 3A11 internalization. Western blot confirmed that 3A11-MMAE induced specific apoptosis and 3A11-Exatecan induced cell death after DNA damage in FGFR4-expressing RMS cells. In vivo, 3A11-MMAE (3 mg/kg, twice weekly for two weeks) delayed RH4 tumor growth, improving survival by 30%, whereas a single dose of 3A11-Exatecan (10 mg/kg) eradicated RH4 tumors, achieving 100% survival. In the aggressive RMS559 model, 40% of 3A11-MMAE treated mice achieved tumor clearance, with a 70% survival rate. While two doses of 3A11-Exatecan completely eradicated RMS559 tumors. Furthermore, both ADCs effectively controlled MDA-MB-453 breast cancer growth. and Future Directions: Our results demonstrate an unprecedented efficacy of these FGFR4-targeting ADCs specifically against human cancers expressing FGFR4 including aggressive rhabdomyosarcoma and breast cancers. We plan to humanize the 3A11 binder and perform pharmacokinetic and toxicology studies in non-human primates for preparation in clinical trials. Meijie Tian, Katrina Jia, Jerry T. Wu, Jun S. Wei, Adam T. Cheuk, Siteng Fang, Victor Ojo, Eleanor G. Pope, Yong Yean Kim, Shyam K. Sharan, Zoe Weaver Ohler, Ludmila Szabova, Simone Difilippantonio, Borys Shor, L. Nathan Tumey, Javed Khan. Development of potent FGFR4-targeted antibody-drug conjugate therapies for rhabdomyosarcoma and other cancers expressing FGFR4 [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 3777.

Deep learning of histologic features from RMS tumor tissue. A, (i) Samples are randomly selected for training, validation, or holdout test groups with k-fold cross validation. Networks were trained to recognize (ii) basic histological characteristics and (iii) features associated with FOXO1 fusion status, or (iv) other relevant RMS mutations. (v) A predictive model was also developed to predict disease risk using only an H&E image. B, Representative H&E images (left), expert pathologist manual annotation (middle), and pixel-level segmentation results of the A.I. algorithm (right). C, Histogram of the weighted IoU scores from holdout test data (n = 29). Samples corresponding to B are indicated. D, Average and weighted intersection over union (IoU) scores of A.I. performance across a 3-fold cross-validation set compared with a pathologist manual annotation.

Background With the development of novel lung cancer treatment modalities, there is an increasing demand for the preoperative diagnosis of lung cancer. Pathological diagnosis is often time-consuming owing to the need for surgical sampling and subsequent processing. Therefore, it is imperative to develop automated and efficient discrimination and prediction models based on CT imaging. Methods In this study, 317 patients with histologically confirmed non-small cell lung cancer (NSCLC) were selected from the TCIA database, and 1834 radiomic features were extracted from the preoperative CT images of each patient. LASSO regression was employed for feature selection, followed by the application of 11 machine learning algorithms to predict the pathological subtypes of NSCLC. The generalization ability of the selected models was evaluated through cross-validation and independent testing. Results During the cross-validation stage, random forest (RF), XGBoost, and extra trees (ET) demonstrated excellent performance, with area under the curve (AUC) values of 0.993, 0.998, and 1.000, respectively, indicating their high accuracy in distinguishing pathological subtypes of NSCLC. During the subsequent independent testing procedure, gradient boosting (GB) achieved the best performance, with an AUC value of 0.644 for the lung adenocarcinoma (LUAD) dataset. XGBoost demonstrated superior results, with an AUC value of 0.618 on the lung squamous cell carcinoma (LUSC) dataset, whereas support vector machine (SVM) achieved an AUC score of 0.638 on the large cell carcinoma (LCC) dataset. Conclusion Our CT-based radiomics approach, leveraging widely available imaging, offers comparable or superior performance to PET/CT and MRI, with potential for integration into clinical workflows to enhance diagnostic efficiency. We developed relatively accurate predictive models for pathological subtypes of NSCLC (LUSC, LUAD, and LCC) using machine learning and radiomic techniques. This method is highly valuable for developing noninvasive diagnostic methods for pathological classification of NSCLC, thereby promoting the early prediction of tumor characteristics and biological behavior.

Supplemental Figure S3. Sample partitioning for training and testing a RAS pathway mutation predictive model using K-fold cross-validation. Three independent experiments were trained on a random selection of samples for training, validation and testing.

Oncogenic fusion genes are attractive therapeutic targets due to their tumor-specific expression and driver roles in cancer. PAX3::FOXO1 (P3F) is the dominant oncogenic driver of fusion-positive rhabdomyosarcoma (FP-RMS) with no current targeted therapy. HiBiT tag, an 11 amino acid peptide of NanoLuc luciferase, was inserted into the C-terminal end of the endogenous P3F using CRISPR. Western was used for HiBiT tag validation. RNA-seq and ChIP-seq were used to assess transcriptomics and DNA binding of HiBiT-tagged P3F (P3F-HiBiT). High-throughput drug screen was performed using the Mechanism Interrogation PlatE drug library with known mechanisms of action. Cell viability was measured using CellTiter-Glo. Mouse xenograft models were used to investigate in vivo efficacy. We validated the HiBiT tagging of P3F by Western. Both P3F-HiBiT and unmodified P3F activated the same gene sets in fibroblasts by RNA-seq Gene Set Enrichment Analysis (GSEA). ChIP-seq using HiBiT antibody verified that P3F-HiBiT binds to the same sites as P3F. A screen for compounds that downregulate P3F in both RH4 and SCMC identified 182 drugs. Filtering for drugs with ≥ 3 hits for the same target identified 14 drug classes, including HDAC inhibitors, BRD4 inhibitors, and CDK inhibitors. Focusing on CDK inhibitors, we found that FP-RMS was most sensitive to CDK7, CDK9 and multi-CDK inhibitors. TG02, a multi-CDK inhibitor with highest inhibition of CDK9 and currently in human trials, downregulated P3F protein. GSEA showed marked suppression of P3F targets after TG02 treatment. Western validated the inhibition of CDK9 with decreased RNA Pol2 Ser2 phosphorylation (Pol2S2). ChIP-seq for RNA Pol2 showed a decrease in transcription pause-release, indicating inhibition of transcription by TG02. Moreover, analysis of genes ranked by decreased Pol2S2 in the gene body showed significant enrichment for P3F targets (p<0.001). TG02 significantly delayed tumor progression without weight loss in a mouse xenograft model of FP-RMS. Also, we found that Vincristine (VCR) and Irinotecan (IRN) are synergistic with TG02 in vitro. Combinations of TG02 with VCR or IRN showed a significant delay in tumor progression compared to TG02 alone in mouse xenograft models. By HiBiT tagging the fusion oncogene P3F, we identified 182 drugs that suppress P3F levels. One of the top hits, TG02, showed in vivo efficacy, indicating that FP-RMS is susceptible to multi-CDK inhibition. Decreased occupancy of Pol2S2 in the gene body of P3F targets indicates that the mechanism of TG02 is primarily through transcriptional inhibition of P3F and its targets. This indicates that TG02 may be effective in transcriptionally addicted cancers such as FP-RMS. We also found synergy between TG02 with VCR and IRN showing promise for clinical translation in FP-RMS. Yong Yean Kim, Katrina Jia, Mehal Churiwal, Soumili Dey, Teresa S. Hawley, Silvia Pomella, Raj Chari, David Milewski, Ranuka Sinniah, Young K. Song, Hsien-Chao Chou, Xinyu Wen, Craig J. Thomas, Michele Ceribelli, Jun S. Wei, Robert G. Hawley, Javed Khan. Endogenous HiBiT-tagging of PAX3::FOXO1 reveals that CDK inhibitors downregulate the fusion oncogene, and demonstrate synergy effects when combined with vincristine and irinotecan [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 7037.

Supplemental Figure S6. Graphical User Interface for tissue segmentation, MYOD1 mutation prediction, and risk prediction models.

Supplemental Figure S5. Frequency of mutations in COG and A.I. designated risk groups.