{
  "abstract": "Background Checkpoint blockade with anti-PD-1 therapy has improved outcomes in advanced melanoma, but identifying which patients are likely to benefit remains a clinical challenge. In this study, we performed an exploratory analysis using the publicly available GSE78220 dataset, which includes transcriptomic profiles from pretreatment melanoma biopsies of patients treated with anti-PD-1. Our goal was to see if TxGemma-2b, a generative language model, could predict treatment response using only a small set of immune genes, without requiring model retraining.Methods To focus on immune-related pathways, we selected 14 genes out of over 25,000 measured per patient, based on prior work, 1–3 which consistently linked these genes to response to immune checkpoint inhibitors. These genes included CD8A, IFNG, GZMB, CXCL9, CXCL10, PRF1, PDCD1, CD274, LAG3, HAVCR2, TIGIT, HLA-A, HLA-B, and HLA-C. Each patient’s gene expression profile was converted into a structured summary describing the expression levels of individual immune markers (e.g., ‘CD8A is high (8.2); IFNG is low (0.5)’), followed by a brief interpretation and classification as Responder or Non-responder. TxGemma-2b was then tested on 23 held-out patients (12 Responder, 11 Non-Responder) in two ways: zero-shot, with no examples provided, and one-shot, using one responder and one non-responder as examples. The one-shot test was repeated five times, and the final prediction was based on majority vote.Results TxGemma achieved 61% accuracy (95% CI: 0.41-0.78) in the zero-shot setting, which improved to 67% (95% CI: 0.47-0.81) using one-shot prompting with a responder and non-responder example. Predicted responders showed higher expression of key immune-related genes, including CD8A, IFNG, and CXCL9, which are associated with cytotoxic T cell activity and interferon signaling. These findings align with established markers of clinical benefit from PD-1 blockade and suggest that even low-shot generative models may capture relevant patterns of response from focused gene signatures.Conclusions This exploratory study shows that TxGemma can predict anti-PD-1 response using just a small set of immune genes and the right prompting strategy, without any retraining. Presenting the data in a way that reflects how clinicians think may help interpret gene expression and support treatment decisions when full model retraining isn’t possible.References Hugo W, Zaretsky JM, Sun LU, Song C, Moreno BH, Hu-Lieskovan S, Lo RS.Genomic and transcriptomic features of response to anti-PD-1 therapy in metastatic melanoma. Cell. 2016;165(1),35–44.Ayers M, Lunceford J, Nebozhyn M, Murphy E, Loboda A, Kaufman DR, McClanahan TK. IFN-γ-related mRNA profile predicts clinical response to PD-1 blockade. The Journal of Clinical Investigation. 2017;127(8):2930–2940.Danaher P,Warren S, Dennis L, D’Amico L, White A, Disis ML, Fling SP. Gene expression markers of tumor infiltrating leukocytes. Journal for Immunotherapy of Cancer. 2017;5:1–15.",
  "authors": [
    {
      "affiliations": [
        "Sentara Albemarle Medical Center, USACS, Elizabeth City, NC, USA"
      ],
      "name": "Mobina Shrestha"
    },
    {
      "affiliations": [
        "The Ohio State University, Columbus, OH, USA"
      ],
      "name": "Salina Dahal"
    }
  ],
  "title": "1119 TxGemma enables low-shot prediction of anti-PD-1 response in melanoma via immune gene expression",
  "uid": "02e89412-6aad-5c6b-9f29-1470efa4af1e"
}