💻 bioinformatics

Bayesian model-based method for clustering gene expression time series with multiple replicates

This paper introduces MMDPGP, a Bayesian model-based method that utilizes Dirichlet process Gaussian processes to cluster gene expression time series with multiple replicates by modeling each replicate separately, thereby achieving superior clustering accuracy and biological insight compared to state-of-the-art approaches.

Original authors: Nushi, E., Douillard, F. P., Selby, K., Lindström, M., Honkela, A.

Published 2026-11-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nushi, E., Douillard, F. P., Selby, K., Lindström, M., Honkela, A.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Life inside a single cell is a constant, bustling conversation. Genes, the instruction manuals for building and running an organism, do not simply sit idle; they turn on and off in complex rhythms, responding to the environment and the organism's needs. To understand how a bacterium grows or how a disease progresses, scientists often look at these rhythms over time, creating a timeline of gene activity known as a transcriptomics time series. However, biology is rarely perfectly uniform. Even when scientists grow identical bacteria in the same lab under the same conditions, slight natural variations occur between different samples, much like how two people might react slightly differently to the same weather. These variations, called replicates, are not just noise to be ignored; they are a source of valuable information about the underlying biological reality. The challenge for researchers has long been how to group genes that behave similarly over time without being misled by these natural differences between samples.

A team of researchers at the University of Helsinki has developed a new way to solve this puzzle. They created a method called MMDPGP, which stands for multiple models Dirichlet process Gaussian process. To understand what this does, imagine trying to sort a large collection of musical recordings. Some recordings are of the same song, but each was played by a different band with its own unique style and tempo. Traditional sorting methods might try to find the average sound of all the bands playing that song and group them together. This approach often fails because the average smooths out the unique characteristics that define the song, or it gets confused when the bands play slightly differently. The new method, however, listens to each band individually while still recognizing that they are all playing the same song. It builds a separate model for each recording, allowing it to appreciate the natural differences between the bands while still correctly identifying the common melody they are all performing.

The researchers tested this approach on a specific type of bacteria called Clostridium botulinum, which is known for producing a potent toxin. They grew three separate cultures of this bacteria and measured their gene activity at thirteen different moments over a twenty-four-hour period. This created a rich dataset where every gene had three slightly different versions of its activity timeline. When they applied their new method to this data, it grouped the genes into twenty-six distinct clusters. In contrast, older methods produced very different results: one method created over a hundred tiny, fragmented groups, while another collapsed the three biological copies into a single average, resulting in only thirteen broad groups. The new method's ability to handle the three separate copies of the data allowed it to find a middle ground that preserved the biological structure of the data without getting lost in the noise.

The true test of whether these groups make sense came from looking at what the genes in each group actually do. The researchers used a standard biological catalog to check if the genes in each cluster shared common functions, such as energy production or cell movement. The groups created by the new method showed a much stronger connection to these real biological functions than the groups created by the older methods. Specifically, the new method identified five distinct groups of genes that were clearly linked to specific biological tasks, whereas the other methods found fewer such connections. This suggests that by respecting the individual variations between the bacterial cultures, the new method was able to uncover the true functional relationships between genes that were hidden when those variations were ignored or averaged out.

To be certain that their method was not just lucky with this specific bacteria, the researchers also created artificial data sets where they knew the correct answers in advance. They simulated gene activity with different levels of noise and variation, mimicking the messy reality of a real lab. In every single test, the new method sorted the genes more accurately than the existing approaches, correctly keeping genes that belonged together and separating those that did not. While one of the older methods performed slightly better at predicting missing data points, this was because it created so many tiny groups that it essentially memorized the data rather than finding the broader patterns. The new method struck the right balance, capturing the natural variability of the experiments to reveal the underlying biological story.

This work represents a shift in how scientists can analyze complex biological data. By treating each experimental copy as a unique voice in a chorus rather than a distortion to be corrected, the researchers have shown that natural variation is a tool for discovery, not an obstacle. Their method provides a clearer picture of how genes work together over time, offering a more reliable way to understand the intricate machinery of life. For scientists studying everything from bacterial growth to human disease, this approach offers a more precise lens through which to view the dynamic world of gene expression.

Technical Summary: Multiple Models Dirichlet Process Gaussian Process (MMDPGP) for Transcriptomics Time Series Clustering

Problem Statement
Clustering gene expression time series is essential for understanding the roles of genes in biological processes and their interaction networks. While traditional methods (e.g., k-means, hierarchical clustering) rely on distance metrics that often ignore temporal dependencies, model-based approaches offer a probabilistic framework. However, existing model-based clustering methods for transcriptomics time series frequently fail to adequately account for the variability inherent in experimental replicates. Some methods ignore replicates entirely by averaging them, while others model all replicates within a cluster using a single Gaussian Process (GP), potentially obscuring natural biological variation. This paper addresses the need for a clustering technique that explicitly models multiple replicates to better capture underlying biological dynamics.

Methodology
The authors propose Multiple Models Dirichlet Process Gaussian Process (MMDPGP), a Bayesian non-parametric clustering method. The core innovation lies in the generative model's treatment of experimental replicates:

Dirichlet Process (DP) Prior: The method employs a DP to generate an infinite mixture model, allowing the number of clusters to be inferred automatically from the data rather than being fixed a priori.
Replicate-Specific Gaussian Processes: Unlike previous approaches, MMDPGP assumes that within each cluster, every experimental replicate is generated by its own distinct GP.
- Generative Model: For a gene $i$ in cluster $\kappa$ and replicate $r$ , the latent function $f^{(r)}_i(t)$ is drawn from a GP with a cluster-specific mean $m^{(r)}_\kappa(t)$ and kernel parameters (length scale, variance, noise).
- Independence: The model treats expression values from different replicates as conditionally independent given the cluster assignment, resulting in a block-diagonal covariance matrix. This reduces computational complexity compared to methods that model all replicates jointly.
- Handling Replicates: Biological replicates are modeled as distinct time series with their own latent functions. Technical replicates are modeled as noisy observations of the same latent trajectory.
Inference: The authors implement Neal's Gibbs sampling Algorithm 8 to sample cluster assignments. The posterior probability of assigning a gene to a cluster is calculated based on the likelihood of the gene's time series (across all replicates) being generated by the cluster's specific GPs.
Comparison Baselines: The method is compared against two state-of-the-art model-based approaches:
- MOHGP (Hensman et al.): Uses a single GP per cluster to model all replicates jointly.
- DPGP (McDowell et al.): Uses a single GP per cluster but collapses multiple replicates into a single mean time series, ignoring replicate variability.

Key Contributions

Novel Architecture: Introduction of MMDPGP, which utilizes a Dirichlet process mixture of GPs where each replicate within a cluster has its own GP model.
Computational Efficiency: By modeling replicates independently within the marginalized model, MMDPGP achieves lower computational complexity ( $O(N_\kappa R T^3)$ ) compared to MOHGP ( $O(N_\kappa (RT)^3)$ ), making it more scalable for dense time series with many replicates.
Biological Utility: The method leverages natural variability between replicates as a source of information rather than treating it as noise to be averaged out.

Results
The authors evaluated MMDPGP using both synthetic data and a real-world RNA-Seq time series dataset from Clostridium botulinum (three biological replicates, 13 time points).

Synthetic Data Performance:
- MMDPGP demonstrated significantly higher cluster purity across various noise levels and phase shifts compared to MOHGP and DPGP.
- It was less likely to incorrectly group time series generated from different underlying clusters.
Real Data Analysis (C. botulinum):
- Cluster Structure: MMDPGP identified 26 clusters, whereas MOHGP produced 116 (mostly small) and DPGP produced 13.
- Biological Enrichment: Gene enrichment analysis using COG (Clusters of Orthologous Groups) annotations revealed that MMDPGP clusters contained a higher number of significantly enriched functional categories (5 at $p < 0.05$ ) compared to MOHGP (2) and DPGP (4). At stricter significance levels ( $p \le 0.01$ ), MMDPGP detected all enrichments found by the other methods plus additional ones.
- Predictive Likelihood: A leave-one-out predictive likelihood analysis showed MOHGP had slightly higher scores. The authors attribute this to MOHGP's tendency to create very small clusters, leading to potential overfitting, whereas MMDPGP produced more biologically meaningful groupings.

Significance and Claims
The paper claims that MMDPGP offers a superior approach for clustering transcriptomics time series with multiple replicates. By modeling each replicate with a separate GP, the method utilizes the natural variability between experimental replicates to learn more about the underlying biology. The authors argue that this approach uncovers gene clusters that behave similarly within replicates but may differ across them due to natural variation, a nuance lost in averaging (DPGP) or joint modeling (MOHGP). The method is presented as a tool that provides finer, more detailed modeling of replicates, leading to clusters with higher biological coherence and functional enrichment. The authors do not claim to solve all clustering challenges but position MMDPGP as a robust alternative for handling multi-replicate data where biological variation is a key feature rather than a nuisance.

More like this