We propose a method to improve the efficiency and accuracy of amortized Bayesian inference by leveraging universal symmetries in the joint probabilistic model of parameters and data. In a nutshell, we invert Bayes’ theorem and estimate the marginal likelihood based on approximate representations of the joint model. Upon perfect approximation, the marginal likelihood is constant across all parameter values by definition. However, errors in approximate inference lead to undesirable variance in the marginal likelihood estimates across different parameter values. We penalize violations of this symmetry with a self-consistency loss which significantly improves the quality of approximate inference in low data regimes and can be used to augment the training of popular neural density estimators. We apply our method to a number of synthetic problems and realistic scientific models, discovering notable advantages in the context of both neural posterior and likelihood approximation.

MOTIVATION: A key process in anti-viral adaptive immunity is that the human leukocyte antigen (HLA) system presents epitopes as major histocompatibility complex I (MHC I) protein-peptide complexes on cell surfaces and in this way alerts CD8+ cytotoxic T-lymphocytes (CTLs). This pathway exerts strong selection pressure on viruses, favoring viral mutants that escape recognition by the HLA/CTL system. Naturally, such immune escape mutations often emerge in highly variable viruses, e.g. HIV or HBV, as HLA-associated mutations (HAMs), specific to the hosts MHC I proteins. The reliable identification of HAMs is not only important for understanding viral genomes and their evolution, but it also impacts the development of broadly effective anti-viral treatments and vaccines against variable viruses. By their very nature, HAMs are amenable to detection by statistical methods in paired sequence/HLA data. However, HLA alleles are very polymorphic in the human host population which makes the available data relatively sparse and noisy. Under these circumstances, one way to optimize HAM detection is to integrate all relevant information in a coherent model. Bayesian inference offers a principled approach to achieve this. RESULTS: We present a new Bayesian regression model for the detection of HAMs that integrates a sparsity-inducing prior, epitope predictions and phylogenetic bias assessment, and that yields easily interpretable quantitative information on HAM candidates. The model predicts experimentally confirmed HAMs as having high posterior probabilities, and it performs well in comparison to state-of-the-art models for several datasets from individuals infected with HBV, HDV and HIV. AVAILABILITY AND IMPLEMENTATION: The source code of this software is available at https://github.com/HAMdetector/Escape.jl under a permissive MIT license. The data underlying this article were provided by permission. Data will be shared on request to the corresponding author with permission of the respective co-authors. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.