A scalable Bayesian functional factor model for high-dimensional longitudinal molecular data

summary

The gist

Large-scale longitudinal molecular profiling is now firmly established in biomedical research.

In short

The paper proposes a Bayesian functional factor model to analyze large-scale, high-dimensional longitudinal molecular data. It combines latent factor modeling with functional principal component analysis to simultaneously capture shared temporal patterns across variables and individual differences. This scalable framework successfully recovers coordinated biological signatures in COVID-19 data, revealing distinct immune activation profiles linked to disease severity.

Key concepts

Bayesian Functional Factor Model
This is a statistical framework that uses Bayesian methods to decompose complex, high-dimensional data into simpler underlying components. It treats the observed molecular measurements as being driven by a few unobserved 'latent factors' that change over time, allowing researchers to understand shared biological processes and individual variations simultaneously.
Functional Principal Component Analysis (FPCA)
FPCA is a technique used within the model to identify subject-specific patterns. Instead of just looking at single variables, FPCA analyzes how groups of variables behave together over time for each person, creating low-dimensional scores that summarize an individual's unique trajectory through the biological data.
Annealed Variational Inference
This is the computational method used to solve the complex statistical problem. It uses a technique called variational inference, which approximates the true complex posterior distribution with a simpler one. Coupling this with simulated annealing helps ensure the algorithm explores all possible solutions efficiently, even in very large datasets.
Latent Factor Processes
These are unobserved, underlying biological processes that drive the variation seen across many different molecular measurements over time. The model assumes these factors are mutually uncorrelated, meaning they represent independent biological mechanisms—like inflammation or interferon responses—that influence the data independently.

Terminology used across episodes

This episode discusses

The paper

A scalable Bayesian functional factor model for high-dimensional longitudinal molecular data · Read on arXiv

Salima Jaoua, Daniel Temko, Áelene Ruffieux

University of Zurich · MRC Biostatistics Unit, University of Cambridge

Large-scale longitudinal molecular profiling is now firmly established in biomedical research, prompted by the need to uncover coordinated biomarker trajectories reflecting the dynamics of underlying biological mechanisms and characterise patient heterogeneity in disease progression. While a range of statistical tools exist for either longitudinal modelling or high-dimensional analysis, there is no unified framework tailored to address these questions jointly. Motivated by a longitudinal COVID-19 study conducted in Cambridge hospitals, we propose a Bayesian functional factor model to address this gap. The framework combines latent factor modelling with functional principal component analysis to represent shared temporal programmes across subsets of variables while capturing individual variation through low-dimensional functional scores. We specify sparsity-inducing priors that yield interpretable factor structure and allow the effective number of factors to be inferred via overspecification. An annealed variational algorithm ensures efficient joint posterior inference at scale. The approach achieves accurate recovery of temporal structure in simulations with up to 20 000 variables. Applied to a panel of 57 circulating biomarkers and to whole-blood expression of 13 177 genes from the COVID-19 cohort, the model recovers coordinated inflammatory, interferon and immune-cell signatures, with subject-level scores that follow disease severity. The methodology is implemented in the R package bayesSYNC.

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "A scalable Bayesian functional factor model for high-dimensional longitudinal molecular data".

Jane: Here's the gist of the paper. Large-scale longitudinal molecular profiling is now firmly established in biomedical research.

Tom: The proposed Bayesian functional factor model addresses the gap by combining latent factor modelling with functional principal component analysis to represent shared temporal programmes across subsets of variables while capturing individual variation through low-dimensional functional scores.

Jane: This framework achieves accurate recovery of temporal structure in simulations with up to 20 000 variables and reveals clinically meaningful heterogeneity in recovery dynamics in COVID-19 data.

Tom: Model Specification The model decomposes the joint variation of the p functional variables into a low-dimensional latent representation where each variable is modeled as yi = µ + Bhi, i = 1,..., N (1).

Paper summary: Tom: So, this model aims to represent those shared temporal programs across subsets of variables while capturing individual variation with these low dimensional functional scores. It sounds like a way to find the structure underneath all that complexity.

Jane: Exactly, it's trying to find those underlying patterns in the data. The paper specifies sparsity inducing priors which are supposed to give us an interpretable factor structure and let us figure out how many factors there are by overspecification.

Lu: They use a mean field variational inference algorithm, which is a way to do posterior inference efficiently at scale when the posterior distribution is multimodal and high dimensional. They even couple it with a simulated annealing scheme with a temperature parameter T to explore that multimodal objective surface better.

Meng: From an engineering standpoint, that sounds complicated for production, but they claim this approach recovers temporal structure in simulations even with up to twenty thousand variables. That’s a big number for handling high dimensionality.

Lalam: They applied it to a panel of fifty seven circulating biomarkers and the whole blood expression of thirteen thousand seven hundred seventeen genes from the COVID-nineteen cohort <ref:2603.20844#pg1>. They recovered coordinated inflammatory, interferon and immune cell signatures, with subject level scores that follow disease severity.

Tom: So, to put it plainly, this Bayesian functional factor model recovers coordinated inflammatory programs in COVID-nineteen data and gives us subject specific scores that track how sick someone is <ref:2603.20844#pg1>.

Conclusion: Jane: Thinking about the title "A scalable Bayesian functional factor model for high-dimensional longitudinal molecular data", what does that really mean for someone listening who isn't deep into statistics?

Tom: It means they took a complex statistical problem and built a framework that can handle massive amounts of data without losing sight of the actual biological signals. It’s about finding the shared patterns in how things change over time across different measurements.

Lu: The core idea is combining latent factor modeling, which reduces the high dimensional data to a small number of latent components representing structured dependence, with functional principal component analysis, which breaks down subject specific temporal variability into orthogonal modes of variation.

Meng: It’s ambitious because it tries to solve the problem where factor models usually treat observations as multivariate but don't explicitly handle the functional aspect of time dynamics.

Lalam: The application showed that this model can reveal two distinct biological signatures in COVID-nineteen: one factor is characterized by strong positive loadings for classical inflammatory markers like CRP, and another is dominated by metabolites enriched for amino acid related pathways <ref:2603.20844#pg1>.

Tom: So, what does this mean practically? It means we can use these low dimensional scalar summaries to characterize the heterogeneity in disease severity and recovery dynamics in COVID-nineteen data <ref:2603.20844#pg1>.

Jane: It suggests that instead of looking at every single gene or biomarker individually, we can identify a few key coordinated programs that drive patient outcomes.

Lu: The authors note a limitation there, though: they mention the complexity of posterior inference with high dimensionality. The method relies on the annealed variational inference scheme to manage that scale effectively.

Meng: And I wonder how robust those factor structures are when we move this from simulation to real world traffic where things are messy and noisy.

Lalam: They say the methodology was directly motivated by a CITIID-NIHR Bioresource COVID-nineteen initiative at Cambridge University Hospitals during the first wave of the pandemic, involving dense longitudinal molecular profiling of individuals with PCR confirmed SARS-CoV two infection and uninfected controls <ref:2603.20844#pg1>.

Tom: So, they’ve shown that this framework works on real patient data to find these coordinated signatures related to immune activation in COVID-nineteen <ref:2603.20844#pg1>.

Jane: That's a big step because it moves beyond just finding correlations to actually modeling the dynamic processes happening within the patients over time.

More episodes

← Home