Gaussian processes for time series with lead–lag effects with applications to biology data

Wancen Mu, Jiawen Chen, Eric S. Davis, Kathleen Reed, Douglas H. Phanstiel, Michael I. Love, Didong Li

TL;DR

The authors introduce GPlag, a Gaussian Process-based framework designed to model "lead-lag" relationships—where one time series follows another after a delay—in datasets with irregular and sparse sampling. By developing a novel class of kernels with interpretable parameters for time lag and dissimilarity, GPlag outperforms traditional methods like Dynamic Time Warping (DTW) in identifying functional biological interactions, such as those between enhancers and promoters in the human genome.

What Problem Does This Paper Address?

In many fields, from finance to biology, we observe pairs of time series where changes in a "leader" series trigger changes in a "follower" series after a specific time delay. This is known as the lead-lag effect.

However, analyzing these relationships is notoriously difficult due to three main hurdles:

  1. Irregular Sampling: Biological experiments are often expensive, leading to observations at non-uniform intervals (e.g., 0, 0.5, 1, 2, 4, and 6 hours).
  2. Data Scarcity: There are often very few time points available, making it hard to estimate transient effects.
  3. External Noise: Factors outside the pair can influence the signals, requiring a robust similarity metric to distinguish true regulatory relationships from coincidental correlations.

Existing tools like Time-Lagged Cross-Correlation (TLCC) or Dynamic Time Warping (DTW) typically require regularly spaced data or struggle to quantify the "strength" of the connection when the lag is unknown.

Methods

The authors propose GPlag (Gaussian Process for lead-lag). Instead of treating the two time series as separate entities, they transform the problem into a "dual state." They model the pair as a single-output Gaussian Process (GP) over an expanded domain that includes both time and an index for the specific series.

Key Innovations:

  • Novel Kernels: They introduce a family of "Time Lag" kernels (LRBF, LExp, and LMat) based on the Gneiting family. These kernels include two critical, interpretable parameters:
  • $s$ (Time Lag): Measures the specific delay between the series.
  • $a$ (Dissimilarity): Quantifies how much the series differ after accounting for the lag. A smaller $a$ indicates a stronger lead-lag relationship.
  • Identifiability Proofs: A major contribution is the theoretical proof that these parameters can be uniquely identified from the data, ensuring that the model's outputs have physical/biological meaning.
  • Flexible Inference: The model supports both Maximum Likelihood Estimation (MLE) for speed and Bayesian inference (via NUTS) for quantifying uncertainty.

Datasets

The paper evaluates GPlag using:

  1. Synthetic Data: Time series generated from known GP kernels to test if the model can accurately recover the true lag ($s$) and dissimilarity ($a$).
  2. Multi-omics Biological Data: A dataset of over 20,000 candidate enhancer-promoter pairs in human macrophages. The data includes measurements of enhancer activity (H3K27ac) and gene expression (RNA-seq) at seven irregular time points following cell activation.

Key Insights and Ideas

  • The "Dual State" Advantage: By modeling the cross-covariance between series directly within the kernel, GPlag "shares information" across the pair. This allows it to make better predictions at unobserved time points than models that treat series independently.
  • Dissimilarity as a Ranking Tool: The parameter $a$ provides a natural way to rank pairs. In the biological application, the authors used $a$ to identify the top 20% of "functional" enhancer-gene pairs, which are more likely to represent true regulatory mechanisms.
  • Continuous vs. Discrete Lag: Unlike TLCC, which can only estimate lags that align with observed time points, GPlag treats the lag $s$ as a continuous variable, allowing for much finer resolution in irregular data.

Key Results

  • Superior Prediction: In simulations, GPlag achieved significantly lower Mean Squared Error (MSE) and higher log-likelihood compared to standard GPs and splines.
  • Perfect Clustering: On a synthetic clustering task (grouping pairs by their distortion and lag), GPlag achieved an Adjusted Rand Index (ARI) of 1.00, whereas baseline methods like TLCC and DTW scored between 0.07 and 0.40.
  • Biological Validation: When applied to the macrophage data, the pairs ranked highly by GPlag showed:
  • Higher Hi-C Counts: Significantly more physical contact in 3D space ($p=0.019$), a "gold standard" for enhancer-promoter interaction.
  • eQTL Enrichment: A strong enrichment for known genetic variants (eQTLs) that regulate gene expression ($p = 0.18 \times 10^{-3}$), validating that the identified pairs are functionally relevant.

Limitations and Future Work

  • Scalability: Like all vanilla GPs, GPlag has a computational complexity of $O(n^3)$. While fine for the small number of time points in this study, the authors suggest using scalable GP methods (like NNGP or variational inference) for larger datasets.
  • Stationarity: The current kernels assume the relationship is "stationary" (the lag and similarity don't change over time). Future work could involve developing non-stationary kernels to capture more heterogeneous dynamics.
  • Higher Dimensions: The identifiability proofs currently hold for 1D time series; extending these proofs to higher-dimensional spatial data remains an open challenge.

Summary updated

Abstract

Investigating the relationship, particularly the lead-lag effect, between time series is a common question across various disciplines, especially when uncovering biological processes. However, analyzing time series presents several challenges. Firstly, due to technical reasons, the time points at which observations are made are not at uniform intervals. Secondly, some lead-lag effects are transient, necessitating time-lag estimation based on a limited number of time points. Thirdly, external factors also impact these time series, requiring a similarity metric to assess the lead-lag relationship. To counter these issues, we introduce a model grounded in the Gaussian process, affording the flexibility to estimate lead-lag effects for irregular time series. In addition, our method outputs dissimilarity scores, thereby broadening its applications to include tasks such as ranking or clustering multiple pairwise time series when considering their strength of lead-lag effects with external factors. Crucially, we offer a series of theoretical proofs to substantiate the validity of our proposed kernels and the identifiability of kernel parameters. Our model demonstrates advances in various simulations and real-world applications, particularly in the study of dynamic chromatin interactions, compared to other leading methods.
BibTeX
@inproceedings{mu2025gaussian,
  author = {Wancen Mu and Jiawen Chen and Eric S. Davis and Kathleen Reed and Douglas H. Phanstiel and Michael I. Love and Didong Li},
  title = {{Gaussian processes for time series with lead–lag effects with applications to biology data}},
  booktitle = {Biometrics},
  year = {2025},
  doi = {10.1093/biomtc/ujae156},
}
Figure