The study evaluated three unsupervised models for imputation of proteomics data, comparing them to heuristic-based approaches like median imputation. The models included collaborative filtering (CF), denoising autoencoder (DAE), and variational autoencoder (VAE) with a stochastic latent space. The results showed that deep learning methods, along with K-nearest neighbors (KNN) and median imputation, were effective in imputing large datasets. However, nine methods implemented in R failed to scale to high-dimensional data.
The development dataset consisted of 564 HeLa runs, and the analysis revealed the structure of the dataset using principal components. The models were evaluated on imputing precursors, aggregated peptides, and protein group data, showing that the self-supervised models performed similarly to other methods across different levels of data aggregation.
The study also assessed the impact of imputation on a large real-world dataset of ALD patients and healthy controls. The results indicated that the deep learning approaches outperformed traditional imputation methods, leading to more differentially abundant protein groups being identified.
Furthermore, the study analyzed the robustness of differently abundant protein group identification and found variability in the analysis due to imputations. Novel protein groups identified through the PIMMS approach were found to be associated with fibrosis, suggesting biological relevance.
Lastly, the additional protein groups included using the PIMMS approach were predictive of fibrosis in the ALD cohort, with deep learning models showing improved performance compared to traditional imputation methods. Overall, the study demonstrated the effectiveness of self-supervised models for imputing missing values in proteomics data and their potential impact on downstream analyses.
Source link
Source link: https://www.nature.com/articles/s41467-024-48711-5