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 pair-wise 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.
Introduction
In fields ranging from finance to molecular biology, researchers frequently encounter pairs of time series where one signal (the "leader") precedes another (the "follower") by a specific delay. Identifying and quantifying this lead-lag effect is crucial for understanding underlying causal or regulatory mechanisms. For instance, in genomics, the activation of an enhancer often precedes the expression of its target gene.
However, traditional time series analysis methods often falter when faced with real-world data constraints. Biological experiments, in particular, are often limited by irregular sampling at non-uniform intervals, data scarcity for transient effects, and external noise that can mask true regulatory relationships.
As shown in the table above, existing tools like Time-Lagged Cross-Correlation (TLCC) and Dynamic Time Warping (DTW) are often restricted to regularly spaced data or lack the ability to simultaneously estimate lag and quantify relationship strength in sparse datasets.
The GPlag Framework
To address these challenges, we introduce GPlag (Gaussian Process for lead-lag). GPlag reframes the problem by modeling a pair of time series as a single-output Gaussian Process over a joint domain of time and series index. This "dual state" approach allows the model to share information across the pair, improving predictive power even at unobserved time points.
Novel Kernels and Identifiability
The core of GPlag lies in a new class of "Time Lag" kernels (e.g., LRBF, LExp, LMat) derived from the Gneiting family. These kernels incorporate two interpretable parameters:
- $s$ (Time Lag): The continuous delay between the two series.
- $a$ (Dissimilarity): A metric quantifying the difference between the series after accounting for the lag. A smaller $a$ indicates a stronger lead-lag relationship.
We provide theoretical proofs for the identifiability of these parameters, ensuring that the estimated lag and dissimilarity are unique and biologically meaningful. The framework supports both Maximum Likelihood Estimation (MLE) for efficiency and Bayesian inference via the No-U-Turn Sampler (NUTS) for robust uncertainty quantification.
Applications in Functional Genomics
The utility of GPlag is most evident in its application to multi-omics biological data. We analyzed over 20,000 candidate enhancer-promoter pairs in human macrophages, where activity was measured at seven irregular time points (0, 0.5, 1, 1.5, 2, 4, and 6 hours) following cell activation.
Key Results
- Superior Performance: In synthetic benchmarks, GPlag achieved perfect clustering (Adjusted Rand Index = 1.00), significantly outperforming TLCC and DTW, which scored between 0.07 and 0.40.
- Biological Validation: By using the dissimilarity parameter $a$ to rank pairs, we identified the top 20% of "functional" interactions. These pairs showed significantly higher physical contact in 3D space (validated by Hi-C data, $p=0.019$) and were strongly enriched for expression quantitative trait loci (eQTLs, $p = 0.18 \times 10^{-3}$).
- Continuous Lag Estimation: Unlike discrete methods, GPlag provides a continuous estimate of the lag $s$, offering finer resolution for understanding the timing of regulatory events in irregular data.
Conclusion
GPlag provides a flexible and theoretically grounded framework for analyzing lead-lag relationships in sparse, irregularly sampled time series. Its ability to rank and cluster pairs based on learned parameters makes it a powerful tool for discovering functional interactions in complex biological systems. While currently assuming stationarity, future work will focus on scaling these methods to larger datasets and extending the kernels to capture non-stationary dynamics.
Summary updated