cfSort
cfSort – Deep-Learning-Based Tissue Deconvolution of cfDNA
Technologies
cfSort
Plasma cell-free DNA (cfDNA) is released by dying cells throughout the body and therefore provides a noninvasive window into tissue injury and cellular turnover. Determining the tissue of origin of cfDNA may reveal abnormal cell death associated with disease, creating opportunities for disease detection, treatment monitoring, and assessment of therapy-related organ injury.
We developed cfSort, the first supervised, deep-learning-based approach for sensitive and accurate cfDNA tissue deconvolution. In benchmarking studies, cfSort achieved greater sensitivity and accuracy than existing deconvolution methods and remained robust to interindividual epigenetic variation, differences in tissue composition, and experimental noise.
The potential clinical utility of cfSort was evaluated using plasma samples from healthy individuals and patients with cirrhosis or cancer, including liver, lung, colorectal, and stomach cancers. The estimated fractions of affected-tissue-derived cfDNA distinguished several disease groups from healthy controls. The study also demonstrated the potential use of cfSort for longitudinally monitoring treatment-associated tissue injury.
Software availability
cfSort is implemented in Python using TensorFlow. Source code, preprocessing tools, model-training and prediction scripts, and links to the pretrained models are available through the cfSort GitHub repository (https://github.com/jasminezhoulab/cfSort). For details, please refer to the full paper. Li S. et al., Proc Natl Acad Sci U S A. 2023 Jul 11;120(28):e2305236120.
(A) Generation of in silico cfDNA mixtures. Simulated cfDNA samples were generated by computationally mixing RRBS data from selected tissue samples. For each simulated sample, source tissues were randomly selected and assigned a tissue composition in which white blood cells (WBCs) were the predominant contributor. DNA fragments were then sampled from each tissue’s RRBS data in proportion to its assigned fraction and pooled to create the simulated cfDNA sample. The predefined tissue composition served as the ground truth for model training and evaluation.
(B) Architecture of cfSort. cfSort is an ensemble of two component deep neural networks. Each network contains three fully connected hidden layers with rectified linear unit (ReLU) activation. A batch-normalization layer precedes each hidden layer, and a dropout layer follows it. The softmax output layer contains 29 nodes representing the tissue categories included in the deconvolution. The final cfSort estimate is obtained by averaging the tissue-fraction predictions from the two component networks.