Driscoll, Julia

Abstract The use of natural killer (NK) cells for the treatment of solid tumors such as cholangiocarcinoma has been hindered by challenges related to cell production, persistence and trafficking, as well as by reduced efficacy within an immunosuppressive tumor microenvironment. Antitumor immunity can be reduced by the interaction of NK cell expressed programmed cell death protein 1 (PD1) with ligands on tumor and other cells within the tumor microenvironment. The antitumor effect of NK cells is enhanced in cells engineered to express a truncated PD1 without an intracellular domain. The cytolytic potential and therapeutic efficacy of extracellular vesicles (EVs) derived from human NK cells expressing truncated PD1 was evaluated as a cell-free therapeutic to overcome some limitations faced by cell therapies. We demonstrate the feasibility of efficient expansion, production and isolation of therapeutic vesicles from human NK cell lines and from engineered NK92 cells expressing truncated PD1. The cytotoxic efficacy of these therapeutic vesicles was validated in both monolayer tumor cell cultures and in multicellular tumor spheroids. Concomitant administration of gemcitabine upregulated NKG2D ligands on tumor cells and increased susceptibility to NK cell and to NK cell derived EV mediated killing. These findings position engineered NK cell derived-EVs as a scalable and promising cell-free immunotherapeutic for cholangiocarcinoma.

Abstract Cholangiocarcinoma (CCA) is an aggressive biliary malignancy with poor prognosis and limited treatment options, underscoring the need for targeted RNA-based therapeutic strategies. Though DNA- aptamers could be one of the promising targeted ligands, conventional enrichment-driven selection often fails to reliably identify high-affinity binders due to PCR bias, stochastic drift, and sequencing errors. In this study, we combined cell-SELEX with a data-driven artificial intelligence framework to identify DNA aptamers targeting CCA cells. Eleven rounds of selection were performed on a library exceeding 6×1014 sequences, with iterative negative selection against normal hepatocytes and cholangiocytes to enhance tumor specificity. Given the limitations of enrichment-based ranking in prioritizing functional binders, we developed and evaluated four deep learning architectures—MLP, CNN, BiLSTM, and a hybrid CNN–BiLSTM—using aptamer pools derived from the final SELEX round. These models were built by integrating one-hot sequential features and DNA structural descriptors of the dataset along with their enrichment patterns across SELEX rounds. Among them, the hybrid CNN–BiLSTM model demonstrated superior performance by capturing both local sequence motifs and long-range dependencies. Model performance was evaluated using AUC-ROC and complementary classification metrics on an independent test set. The hybrid CNN–BiLSTM model achieved the best overall performance with an AUC-ROC of 0.97 and a precision of 68%, indicating strong discriminative power under class imbalance (strong: non-binder=1:68). The model exhibited high specificity (99.5%; 29,855/30,005 true negatives) and moderate sensitivity (71.7%; 312/435 true positives), reflecting its ability to reliably filter non-binding sequences while retaining true functional binders. Model-prioritized candidates were experimentally validated using qPCR-based aptamer–cell binding assays. Notably, two low-abundance/high-prediction aptamers (Read counts: 4) exhibited 2–3-fold higher binding than the reference aptamer (Positive Control) derived from traditional SELEX enrichment. A high-read/high-prediction candidate showed ∼2-fold greater binding than the positive control, highlighting the model’s ability to recover overlooked functional binders. To assess translational relevance, selected aptamers were used to functionalize the lactosome-based nanovesicles for siRNA delivery targeting PD-L1, resulting in measurable gene silencing in HuCCT1 cells after 48 hours. Collectively, this integrative framework enables accurate identification and prioritization of high-affinity, functionally relevant aptamers beyond enrichment alone. This approach enhances SELEX efficiency, supports rational ligand design, and provides a scalable strategy for developing aptamer-guided RNA therapeutics for any target. Citation Format: Sourav Pal, Julia Driscoll, Brandon Wilbanks, Ayano Kabashima, Rory Smoot, Tushar Patel. Integrating cell-SELEX and deep learning method for discovery of high-affinity aptamers targeting cholangiocarcinoma [abstract]. In: Proceedings of AACR Drug Discovery and Development (AACR D3) Conference; 2026 Jul 21-24; Boston, MA. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(14_Suppl):Abstract nr A033.

Abstract Background and Objectives Cholangiocarcinoma (CCA) is an aggressive, highly lethal cancer of the biliary tract epithelium for which tumor resection is the only curative treatment. However, approximately half of patients present with nonresectable disease and median survival of these patients is just 8-12 months. A significant need exists for tumor-targeted therapeutics for patients with advanced disease who currently have limited treatment options. We aimed to address this need by developing an aptamer-targeted therapy for delivery of siRNAs inhibiting undruggable components of the Hippo pathway, a key driver of cholangiocarcinoma. Method We applied cell-SELEX against human CCA model cell line HuCCTl with negative selection against normal hepatocytes and cholangiocytes to identify an aptamer that selectively bound tumor cells. This aptamer, termed MC7122, was chemically modified to include fluorescent labels that enabled whole-animal biodistribution studies to assess in vivo selectivity. MC7122 was then deployed onto the surface of siRNA-loaded nanovesicles to drive tumor-targeted nucleic acid delivery. Tumor-bearing mice were treated with siRNAs three times per week for two weeks alone or in combination with standard of care chemotherapy. Results Combined treatment with siRNAs and standard of care chemotherapy reduced tumor burden in mice versus vehicle controls and slowed the spread of metastasis. Importantly, we found tumor-specific target gene knockdown without corresponding off-target knockdown in healthy surrounding liver tissue. Further, we demonstrated that sites in the lung, the primary location of metastasis in this model, were reduced not because of global target knockdown in lung but due to tumor-specific knockdown at the site of liver implantation. Conclusions Aptamer-guided vesicles represent a promising platform for targeted therapy delivery for previously undruggable targets. With this platform, tumor-selective therapies can be designed to exhibit exceptionally potent cytotoxicity, as healthy tissues remain unaffected by off-target effects. Future work will focus on refinement of therapeutic siRNA target selection and modification of nanovesicle components for enhanced levels of tumor specificity.