Fig. 1

Schematic workflow for predicting cancer-specific vulnerability in clinical samples. Transcriptomes and matching dependency maps constructed for cancer cell lines are used for training. Tumor and matched normal transcriptomes of clinical samples are used as input for prediction. Cancer-specific vulnerability can be identified by comparing the prediction outcomes for tumor and normal samples. The prediction model consists of in silico CRISPR/RNAi and machine learning. For a given transcription profile derived from a cell line or clinical sample, virtual repression can be performed by adjusting the expression level of each target gene and its downstream genes. The perturbed expression profiles are fed into neural networks for training or prediction. The output of the model is the probability that the perturbed transcriptome is associated with cell death or that the survival of the given sample is dependent on the inactivated gene (see the “Materials and methods” section)