Deep Learning models to predict lung transplant rejections using cell-free DNA fragmentomics
Summary
Chronic respiratory diseases impose a significant burden to society, and lung transplant remains the last-resort treatment option. However, long-term survival rates for lung transplant recipients are hindered by a high risk of rejection, leading to Chronic Lung Allograft Dysfunction (CLAD) and eventual death. Traditional methods for diagnosing rejection are invasive and often prove too late. Thus, there is an urgent need for diagnostic methods for timely detection of rejections to improve lung transplant outcomes. The levels of donor-derived cell-free DNA (dd-cfDNA) in the recipient’s blood hold great promise as a biomarker for diagnosing transplant rejections. Current measurement methods involve the selective amplification of dd-cfDNA using donor and recipient SNPs. However, this method is limited by the number of SNP differences between the donor and the recipient. We introduce a novel method for measuring dd-cfDNA levels by employing Deep Learning (DL) models to distinguish between donor- and recipient-derived cfDNA based on their tissue of origin. To this end, three DL models were developed: a feed-forward neural network trained on epigenetic features of cfDNA (extracted using Enformer), a convolutional neural network (CNN) focused on sequence motifs, and a second CNN model utilizing both epigenetic features and sequence motifs for classification. The models, trained and evaluated on a dataset of cfDNA samples collected from 47 lung transplant recipients, achieved only marginal improvement over a random classifier, with the best-performing model achieving an area under the ROC curve of just 0.524. Baseline logistic regression and dimensionality reduction analysis pointed to either highly complex or absent signals in the data. High accuracy achieved by these models on simulations with artificially embedded signals showed that given enough signals, the models could learn to distinguish dd-cfDNA from rd-cfDNA, further underscoring the complexity of the real dataset. Contrary to common consensus, there was a poor correlation between dd-cfDNA percentage and clinical signs of rejection in the training dataset, raising questions about training label accuracy. Poor data quality, inaccuracies in training labels, and incomplete hyperparameter optimization are potential contributing factors to the suboptimal performance of the models. While the models we developed did not perform well enough to be considered reliable for diagnosis, the approach of using dl models to measure dd-cfDNA levels is novel and holds great promise. With higher quality training data, selective sample incorporation into the training set, and a more extensive hyperparameter search, these DL models could serve as highly effective non-invasive tools for transplant rejection diagnosis, with potential applications in areas like prenatal diagnosis and liquid biopsy.