Evaluating single-cell ATAC-seq atlasing technologies using sequence-to-function modeling.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Understanding cis-regulatory logic is crucial for cell identity.
• Single-cell chromatin accessibility (scATAC-seq) atlases aid in training sequence-to-function (S2F) deep learning models.
• Optimal criteria for scATAC-seq training datasets and platform suitability for S2F models are unclear.

Purpose of the Study:

• To benchmark scATAC-seq platforms for S2F model training and transcription factor (TF) footprinting.
• To evaluate the impact of cell number and fragment counts on training data quality.
• To assess the performance of S2F models trained on different data sources.

Main Methods:

• Introduction of HyDrop v2, an improved custom droplet scATAC-seq method.
• Benchmarking of scATAC-seq platforms for S2F model training and TF footprinting.
• Comparative analysis of S2F models trained on custom and commercial scATAC-seq data.

Main Results:

• Lower fragment counts can be compensated by increasing cell numbers in training datasets.
• S2F models trained on custom or commercial scATAC-seq data exhibit comparable performance in enhancer prediction, sequence explainability, and TF footprinting.
• Data integration from different scATAC-seq platforms facilitates large-scale, cost-efficient atlas construction.

Conclusions:

• scATAC-seq platform choice impacts S2F model training and TF footprinting capabilities.
• Data integration strategies can overcome limitations of individual platforms for building comprehensive regulatory atlases.
• This work provides guidelines for constructing effective scATAC-seq datasets for deep learning-based regulatory modeling.