.png){kind=icon}
.png){kind=icon}
Return
mbSparse: an autoencoder-based imputation method to address sparsity in microbiome data
.png){kind=icon}
OA
AI
DOI：10.1080/19490976.2025.2552347.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The involvement of gut microbiota in host physiological activities is crucial, yet the high sparsity of microbiome data, marked by numerous zeros in count matrices, presents huge analytical challenges. To overcome this, we developed mbSparse, an imputation algorithm that leverages deep learning rather than traditional predefined count distributions. Utilizing a feature autoencoder for learning sample representations and a conditional variational autoencoder (CVAE) for data reconstruction, mbSparse effectively integrates these processes to enhance imputation. Our results demonstrate that mbSparse achieves exceptional accuracy, with mean squared error reductions of up to 4.1 compared to existing microbiome methods, even amid outlier samples and varying sequencing depths. In colorectal cancer analysis, mbSparse increases the detection of validated disease-associated taxa from 7 to 27, while predictive accuracy improves, as evidenced by area under the precision-recall area under the curve values rising from 0.85 to 0.93. Additionally, mbSparse addresses non-biological zeros by restoring over 88% of removed counts and achieving a Pearson correlation of 0.9354 at a 10% removal rate, preserving essential taxonomic relationships. Finally, our exploration of mbSparse variants reveals that the CVAE is critical for enhancing accuracy, providing valuable insights for further optimizing microbiome data imputation techniques.
Keywords:
Microbiome
deep learning
sparsity
imputation