Cancerformer: A CRISPR Screen-benchmarked Multimodal AI Platform for Predication of Cancer Dependencies in Patient-derived Organoids

Dissection of cancer dependencies is the central topic of cancer research. Recent advance in artificial intelligence (AI) has provided the opportunies of rapid predication of cancer essential genes. However, these AI models are often limited by the incapability of leveraging multimodal information or insufficient benckmarks, leading to low success rate in physiologically relevant practice. Here, we developed Cancerformer, a multimodal deep learning framework that integrates single-cell RNA sequencing (scRNA-seq), TCGA transcriptomic profiles and protein-protein interaction (PPI) networks to predict cancer gene essentiality. By employing a Transformer architecture to capture gene functional context and Graph Neural Networks to embed topological structures of PPI networks, Cancerformer overcomes the generalization limitations of existing methods. Using the experimental results of CRISPR screen from multiple cancer cell lines HeLa, A549 and U-87MG, we demonstrated that Cancerformer consistently outperformed state-of-the-art baseline models under both gene-wise and sample-wise cross-validation splits as measured by multiple evaluation metrics. In subsequent applications, Cancerformer demonstrated strong generalization ability achieving a 90% experimental verification rate for top candidates in colorectal cancer HCT116 cells. Most importantly, without pre-training on patient-derived organoids (PDOs) data, Cancerformer successfully captured inter-patient heterogeneity in PDOs and revealed a context-specific metabolic dependency on oxidative phosphorylation pathways in 3D culture compared to 2D cell lines. Functional assays on top predicted targets confirmed their essentiality in PDO growth. This study established Cancerformer as a rigorously benchmarked multimodal AI model for predicting cancer dependencies with physiological relevance.