Data-Driven Soft Labeling Scales DNA Read Classification to Whole-Body Cell-Type Deconvolution

Cell-type deconvolution is a fundamental task in computational biology, aiming to estimate the proportions of different cell types within a heterogeneous sample. Traditional methods often aggregate methylation data, losing the granular information contained in individual DNA reads. While some read-level approaches exist, they struggle to scale to large cell-type panels because hard labels (assigning a read to a single cell type) conflict with the biological reality that methylation patterns are often shared across multiple cell types. To address this, the authors developed Syto, a modular framework that utilizes data-driven soft labels to estimate the conditional distribution of cell types for each read, allowing for more nuanced classification.

Syto consists of three main, independently improvable components: a read-level classifier, a deconvolver, and a post-hoc calibrator. The authors propose a novel soft-labeling scheme that acts as a maximum-likelihood estimator of the true conditional categorical distribution, effectively handling the many-to-many mapping between methylation patterns and cell types. To handle class imbalance and low per-signature read coverage, the framework incorporates a pooling mechanism that aggregates counts from similar signatures. The framework also introduces a simple, hyperparameter-free linear calibration method to correct biases in the deconvolver's output, ensuring the predicted proportions align with the probability simplex.

In experiments using a 39-cell-type reference atlas, Syto consistently outperformed existing state-of-the-art methods, achieving a 2.56-fold reduction in mean squared error (MSE). The framework also demonstrated strong transferability to an out-of-distribution dataset spanning 16 tissues, where it achieved high tissue concordance scores. By enabling the use of larger, more complex reference panels, Syto provides a scalable foundation for future applications in tumor classification, purity estimation, and the analysis of complex immune microenvironments.