Integração em mosaico e geração sintética de dados multiómicos para a descoberta de medicinas de precisão para o cancro

New cancer therapies face significant hurdles: drug resistance and clinical trial high costs, long approval times, and low success rates. SYNTHESIS proposes innovative deep learning approaches to prioritize drug targets for clinical validation that are strongly supported by molecular and phenotypic data and are validated in tumor samples and in vivo models. We will develop variational autoencoders (VAEs) that integrate heterogeneous, large and sparse cancer pre- and clinical databases. Key technological innovations are: 1. Optimized VAE designs that effectively integrate 7 different data types, i.e. five different omics and two phenotypic (drug and genetic) screens. 2. Mosaic data integration across 3 different cancer model systems (cancer cell lines, organoids and tumors) to find drug targets and molecular mechanisms that validate from preclinical to clinical contexts. 3. Transfer learning using generative models pretrained on preclinical or pan-cancer contexts and fine-tuned in clinical or cancer type specific contexts, to simulate tumor profiles, including drug and genetic screens, which are not experimentally possible. These innovations apply recent deep learning techniques to large-scale cancer databases and enable three main objectives: 1. Build a Synthetic Cancer Dependency Map: Using our VAE models we will integrate all broadly available large-scale multi-omics, drug and genetic screens of cancer cell lines. This will synthetically generate any type of omic and phenotypic screens that are missing, thereby creating the first Synthetic Cancer DepMap. 2. Mosaic Integration and Transfer Learning From Preclinical To Clinical Contexts: Cancer vulnerabilities and drug targets are traditionally either studied in preclinical or clinical models. Therefore, treating clinical validation as a secondary test misses the opportunity to use all data for model training. Mosaic integration and transfer learning will merge preclinical (cell lines and organoids) and clinical (tumor) data to train or fine-tune VAE models, identifying drug targets and molecular signatures consistent across contexts and more likely to achieve clinical validation. 3. Prioritizing Synthetic Vulnerabilities In Triple-Negative Breast Cancer (TNBC): TNBC is notoriously difficult to treat. We will employ our VAE models to synthetically enhance existing large preclinical and clinical multi-omic TNBC datasets. Our goal is to improve the identification of cancer subtypes and drug targets in TNBC through synthetic tumor data, which will be validated experimentally in vivo. We envision SYNTHESIS as a proof-of-concept of how deep learning with multi-omics can identify evidence-based targets for clinical validation, promoting the use of generative machine learning to simulate tumor behavior and treatment outcomes, and setting the stage for in silico clinical trials.