{
  "abstract": "Background Precision oncology is increasingly vital in clinical settings, improving treatment outcomes through molecular profiling. However, obtaining transcriptomic and methylation data remains expensive and time-consuming. Here, we present Path2Omics, a deep learning framework capable both gene expression and DNA methylation directly from histopathology slides across 30 cancer types.Methods Unlike existing approaches that rely solely on FFPE slides for training, Path2Omics leverages both FFPE and FF slides by constructing two separate models: one based on the FFPE slides, called ‘ FFPE model’, and another based on FF slides, called ‘FF model’. Our final ‘integrated model’ combines both predictions from the FFPE model and FF model.Results In five-fold cross-validation on the TCGA cohort, FFPE models achieved an average of 3,057 well-predicted genes. The FF models, however, achieved double the predictive coverage, with an average of 6,311 well-predicted genes. To assess model generalizability, we applied the pre-trained models to 7 external datasets comprising of 1,323 slides (1,163 FFPE and 160 FF). Surprisingly, despite 6 of 7 datasets consisting solely of FFPE slides, the FF models still outperformed the FFPE model, achieving an average of 3,691 well-predicted genes compared to 3,404. The integrated model further improved performance, achieving an average of 4,391 well-predicted genes. Importantly, the well-predicted genes are strongly enriched in immune-related pathways across most cancer types, highlighting the potential to develop a predictor of patient response to immunotherapy directly from pathology slides.To demonstrate the potential translational value of the inferred gene expression, we developed models to predict patient overall survival and treatment response to cancer therapies using the inferred gene expression profiles. The results were comparable to those achieved with measured gene expression and outperformed the ‘direct’ modes, which relied solely on pathology slides without inferred gene expression as an intermediate.Conclusions Our results demonstrate that integrating two different types of slide preparations while training gene expression and methylation predictors significantly enhance their prediction accuracy, even when only one slide type is available during inference. Our findings also highlight the model’s utility for downstream clinical tasks such as survival analysis and treatment response prediction, underscoring its potential as a scalable tool for precision oncology.",
  "authors": [
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Danh-Tai Hoang"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Eldad D Shulman"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Saugatorahman R Dhruba"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Nishanth Ulhas Nair"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Ranjan K Barman"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Thomas Cantore"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "H Lalchungnunga"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Omkar Singh"
    },
    {
      "affiliations": [
        "University of Pennsylvania, Philadelphia, PA, USA"
      ],
      "name": "Maclean P Nasrallah"
    },
    {
      "affiliations": [
        "Australian National University, Canberra, ACT, Australia"
      ],
      "name": "Eric A Stone"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Kenneth Aldape"
    },
    {
      "affiliations": [
        "National Cancer Institute, National Institutes of Health, Bethesda, MD, USA"
      ],
      "name": "Eytan Ruppin"
    }
  ],
  "title": "1095 Pan-cancer prediction of transcriptomics and methylation from tumor pathology slides",
  "uid": "4f67c609-5a0e-55db-b6ea-885e4aee900d"
}