Fig. 1

Overview of mbDenoise and the noise model. a mbDenoise distinguishes biological zeros from technical zeros, and assumes that the true nonzero abundance data lie on a low-dimensional latent space embedded in the high-dimensional feature space, reflecting the observed redundancy in microbiome data. mbDenoise recovers the true abundance levels, that is, the latent signal matrix, by fitting a zero-inflated probabilistic PCA (ZIPPCA) model. The ZIPPCA framework takes into account uneven library size, overdispersion, and sparsity using a mixture model that consists of a negative binomial count distribution and a point mass at zero. b Input data (that is, observed count matrix) are assumed to be samples from this mixture model, and the posterior mean estimate of the latent signal matrix by variational approximation represents the denoised output. The denoised abundance can be used for multiple downstream analysis tasks