Pancreatic ductal adenocarcinoma (PDAC) is a highly aggressive malignancy with poor survival, even among patients with potentially resectable disease. Outcomes are influenced in part by the tumor immune microenvironment, particularly CD8⁺ T-cell infiltration and regulatory signaling pathways such as focal adhesion kinase (FAK) a non-receptor tyrosine kinase that regulates tumor–stromal interactions, fibrosis, and immune cell exclusion. Surgically resected tumor specimens from 19 patients treated with neoadjuvant (preoperative) chemotherapy were analyzed using multiplex immunohistochemistry, with a focus on CD8 and FAK protein expression. Quantitative image-based analysis was performed, and comparisons between groups were evaluated using unpaired t-tests. The marker panel included CD8 and phosphorylated FAK. Patients with higher FAK expression demonstrated lower CD8⁺ T-cell infiltration, whereas significantly increased T-cell infiltration was observed in tumors with low levels of phosphorylated FAK (P = 0.012). These findings demonstrate an inverse association between FAK signaling and cytotoxic T-cell infiltration, supporting a role for FAK in promoting an immunosuppressive tumor microenvironment in PDAC. Targeting FAK-mediated pathways may represent a rational strategy to enhance antitumor immunity and improve the efficacy of immunotherapeutic approaches in this treatment-resistant disease.

Large language models are increasingly used in mental health applications, yet their reliability in detecting suicide risk remains unclear. This study investigates whether roleplay prompting affects the ability of large language models to classify mental health risk accurately. Using a benchmark dataset of annotated dialogues, we evaluate models on two tasks: Concern Type and Risk Level classification. We adapt the dataset into a user-only setting to isolate how prompt framing influences model behavior. Models are evaluated under multiple prompt conditions, including neutral instructions and roleplay scenarios such as therapist and romantic partner. Performance is measured using accuracy, macro F1, recall on high-risk cases, and false negative rate for severe cases. This work highlights the importance of prompt design in safety-sensitive applications and demonstrates that roleplay prompting may introduce hidden risks in mental health assessment systems using large language models.

T-cell epitope vaccines utilize immunogenic peptide sequences to trigger targeted, adaptive immune responses against pathogens. Considered as the next-generation vaccine platform, these peptide vaccines have shown significant promise in clinical trials against bacteria and viruses, fueling interest for personalized cancer immunotherapy. However, progress in cancer immunotherapy is remains stifled by reliance on whole-exome sequencing and by algorithmic assumptions equating MHC binding affinity with cancer immunogenicity. While this assumption works well for foreign pathogens, cancer treads the line between self and non-self. In this project, we seek to open up the search space by developing a novel diffusion machine learning model that outputs diverse, biologically meaningful, and physicochemically viable epitopes that is patient-context and tumor-type aware. However, generation of these epitopes does not guarantee that these epitopes will trigger a T-cell response. Therefore, we design a multi-step and multi-signal filtering process that holistically validates the novel epitope candidates that considers its immunogenicity against the tumor, probability of binding affinity, and the physical structure between the peptide and the Human Leukocyte Antigen employing multiple ensembles and orthogonal deep learning architectures. By moving beyond the binding-affinity-based framework, this work seeks to expand the possibilities for in-vitro and in-vivo cancer epitope discovery and, ultimately, advance cancer vaccine therapy.

Zika virus (ZIKV) is a member of the Flavivirus genus that can cause congenital Zika syndrome, including microcephaly, and currently lacks approved treatments or vaccines. Previously, a screen of 2,750 small-molecule kinase inhibitors was conducted to determine their effects on ZIKV replication. Three compounds (Tamnorzatinib, Cabozantinib S-malate, and Merestinib) were identified as effective in mitigating ZIKV-induced cytopathic effects (CPE), reducing viral activity in retinal pigment epithelial (RPE) cells, and inhibiting viral entry into host cells. Following this drug screen, RNA sequencing (RNA-seq) analysis was performed to assess the specific upregulatory and downregulatory effects of the three kinase inhibitors. Comparisons were performed across infected treated, infected untreated, uninfected treated, and uninfected untreated conditions. The results show that all three drugs significantly reduced differential gene expression in infected treated cells compared to infected untreated cells, suggesting a reduction in virus-associated host transcriptional responses. Furthermore, treated infected cells exhibited reduced expression of genes associated with antiviral responses, consistent with altered host responses to infection. Overall, these findings suggest that all three kinase inhibitors have potential as candidates for inhibiting ZIKV infection.

This abstract has been withheld from publication.