GenTS: Multifaceted Methods in Bioinformatics & Time Series

• GenTS is a multifaceted framework that encompasses network-based gene-set testing in bioinformatics, benchmark pipelines for generative time series, and spatio-temporal calibration in regression analysis.
• In bioinformatics, it extends classical gene-set analysis by using Gaussian graphical models and multi-splitting to detect network differences beyond mean shifts, revealing subtle pathway rewiring.
• In generative time series, GenTS provides a modular benchmarking suite for synthesis, forecasting, and imputation, offering comprehensive evaluation metrics to guide model selection.

Searching arXiv for "GenTS" to identify relevant papers and disambiguate the term. GenTS is used in the arXiv literature for more than one methodological construct. In bioinformatics, it denotes a network-based gene-set testing framework that extends classical gene-set analysis from univariate differential-expression summaries to multivariate hypotheses on gene-gene interaction structure within predefined pathways (Städler et al., 2013). In generative time-series research, it denotes a comprehensive benchmark library for generative time series models, designed around synthesis, forecasting, and imputation rather than forecasting-only or otherwise discriminative workflows (Wang et al., 18 May 2026). Related methodological summaries also use “GenTS/TDS” for a Top-Down Scale calibration idea in multiscale geographically and temporally weighted regression, where ordered-grid, coordinate-wise AICc search is used to stabilize covariate-specific spatial and temporal bandwidth selection (Geniaux et al., 24 Mar 2026).

1. Terminological scope and disambiguation

The name has domain-specific meanings, and the surrounding literature is not uniform. The table summarizes the principal usages represented here.

Usage	Domain	Defining formulation
GenTS	Bioinformatics	Network-based gene-set testing of $\mathbf{H}_{0,s}:\Omega^x_s=\Omega^y_s$
GenTS	Generative time series	Benchmark library with unified preprocessing, model collection, and panoramic evaluation
GenTS/TDS	Spatial econometrics	Top-Down Scale calibration via ordered grids and coordinate-wise AICc search
GTaS	Computational neuroscience	“generalized thinning and shift,” a distinct acronym

This disambiguation matters because the three GenTS-related usages operate at different levels of abstraction. The bioinformatics GenTS is a statistical test on Gaussian graphical models; the time-series GenTS is an extensible software benchmark stack; and the GenTS/TDS usage refers to a calibration scheme for multiscale local regression. A separate acronym, GTaS, denotes a generative spike-train model and should not be conflated with GenTS (Städler et al., 2013, Wang et al., 18 May 2026, Geniaux et al., 24 Mar 2026, Trousdale et al., 2013).

2. GenTS in bioinformatics: multivariate gene-set testing on networks

In the bioinformatics paper, GenTS is introduced as an extension of classical gene-set analysis (GSA). Classical GSA methods aggregate univariate two-sample statistics across genes inside a pathway. This can detect coordinated mean shifts, but it cannot detect changes in covariance structure, conditional dependence, or network topology. GenTS targets precisely that missing regime: for each predefined gene set $A_s\subset\{1,\ldots,d\}$, it asks whether the multivariate dependency pattern differs between two biological conditions.

The formulation is expressed through Gaussian graphical models with zero mean and condition-specific concentration matrices,

$$f_s^x \sim N(0, (\Omega_s^x)^{-1}), \qquad f_s^y \sim N(0, (\Omega_s^y)^{-1}),$$

with graph edges encoded by nonzero off-diagonal entries of the precision matrix. The null hypothesis for gene set $A_s$ is

$\mathbf{H}_{0,s}: \; \Omega^x_{s}=\Omega^y_{s}.$

Accordingly, the test is not a mean-shift test. It isolates network differences, and the paper later suggests combining it with classical GSA when both mean and network changes are scientifically relevant.

The method is adapted from the high-dimensional two-sample framework DiffNet. Its rationale is that direct two-sample likelihood testing on full covariance structures is unstable or infeasible when the number of genes in a set is large relative to sample size. GenTS therefore turns network comparison into a split-sample procedure in which graph estimation and hypothesis testing are separated. This suggests a deliberate bias toward validity under high dimensionality rather than toward a single-shot full-model likelihood ratio test (Städler et al., 2013).

3. Statistical machinery, workflow, and empirical interpretation in the bioinformatics setting

For each gene set, GenTS randomly splits the samples into two halves. The first half is used to estimate condition-specific graphs $\hat G_s^x,\hat G_s^y$ and a pooled graph $\hat G_s^{xy}$. The second half is then used for likelihood-based comparison between the separate and pooled models. The central statistic is an AIC difference,

$$\Delta\mathrm{AIC}_s = \mathrm{AIC}^{\rm ind}_s-\mathrm{AIC}^{\rm joint}_s = 2\left(L^{\rm ind}_s-L^{\rm joint}_s\right)-2(\mathrm{df}^{\rm ind}_s-\mathrm{df}^{\rm joint}_s),$$

with

$$\mathrm{df}^{\rm ind}_s = 2d_s + |E(\hat G^x_s)| + |E(\hat G^y_s)|,\qquad \mathrm{df}^{\rm joint}_s = d_s + |E(\hat G^{xy}_s)|.$$

Under the null, $\Delta\mathrm{AIC}_s$ is asymptotically distributed as a shifted weighted sum of chi-squares, yielding asymptotic p-values rather than permutation p-values. Multiple testing is handled by FDR correction across gene sets, and repeated random splitting is stabilized by multi-splitting with aggregation by the median corrected p-value.

Validity depends on two assumptions imposed on the graph estimator used on the first split. The Screening assumption (ScA) requires

$A_s\subset\{1,\ldots,d\}$0

and the Sparsity assumption (SpA) requires the estimated graph to be sparse enough for stable maximum-likelihood fitting on the second split. A sparsity index is defined as

$A_s\subset\{1,\ldots,d\}$1

with its inverse interpreted as “samples per predictor” in a regression view of Gaussian graphical model estimation.

The default estimator is the Graphical Lasso,

$A_s\subset\{1,\ldots,d\}$2

with tuning by either cross-validation or BIC. The preferred implementation is GL-BIC-AT: Graphical Lasso, tuning by BIC, and adaptive thresholding when the graph is too dense. Adaptive thresholding removes edges with the smallest absolute partial correlations until the inverse sparsity $A_s\subset\{1,\ldots,d\}$3 is below a threshold $A_s\subset\{1,\ldots,d\}$4, with $A_s\subset\{1,\ldots,d\}$5 used in the paper.

Simulation studies compare Graphical Lasso, Meinshausen–Bühlmann neighborhood selection, and shrinkage covariance estimation with partial-correlation testing. The reported findings are that Graphical Lasso and Meinshausen–Bühlmann generally control type I error well in one sparse-random-precision setting, that shrinkage is often too dense or misspecified and has poor power especially in the autoregressive setting, and that adaptive thresholding is needed because Graphical Lasso without thresholding can yield inflated false positives in the second simulation model. Multi-splitting outperformed single-splitting, classical mean-based GSA could miss network-only differences, and combining classic GSA with NetGSA improved coverage of both signal types. Real-data applications on CCLE, lung cancer, and breast cancer datasets using BioCarta gene sets produced rankings that often differed substantially from classical GSA, with highlighted pathways including IL3_PATHWAY, SHH_PATHWAY, and IL7_PATHWAY. The paper interprets these results as evidence that pathway rewiring can be biologically informative even when average expression changes are modest, while also emphasizing limitations: Gaussian graphical model assumptions, computational cost, dependence on screening and sparsity conditions, sensitivity to non-normality, and the fact that the method tests network differences only (Städler et al., 2013).

4. GenTS in generative time-series research: benchmark architecture and task formalization

In the 2026 paper, GenTS is a comprehensive benchmark library for generative time series models. It is motivated by a mismatch between existing time-series toolkits and the needs of generative modeling. Libraries such as TSLib, GluonTS, and ProbTS are described as primarily designed around forecasting-style objectives and standard regression or classification losses, whereas generative models rely on adversarial training, variational objectives, denoising diffusion, normalizing flows, and continuous-time differential equations. GenTS is therefore positioned as a full benchmarking pipeline with three tightly connected parts: unified data preprocessing, a model collection spanning major generative families, and panoramic evaluation metrics.

The library is implemented on top of PyTorch and PyTorch Lightning. A BaseDataModule handles dataset splitting, dataloader construction, and task-specific preprocessing. It can parse .csv, .txt, and .tsf, convert datasets into tensor windows, create training, validation, and test loaders, simulate missingness for imputation, assign look-back and prediction windows for forecasting, and optionally simulate irregular time gaps for continuous-time models such as neural differential equations. The benchmark includes more than fifteen datasets from six domains, covering traffic, energy, finance, medical, weather, physics, and related settings, alongside two synthetic datasets, Spiral2D and SineND.

The model collection spans five families—GANs, VAEs, flows, diffusion models, and differential equation-based models—and the paper states that the library currently includes more than 25 SOTA models and variants. Four simple baselines are included: VanillaVAE, VanillaGAN, VanillaDDPM, and VanillaMAF. The design is modular: both datasets and models are expressed through base classes, and the BaseModel standardizes training_step, validation_step, and sample, including support for conditional sampling and batched generation.

Task formalization follows a generative-distribution perspective. Time series generation is defined as modeling a proxy distribution $A_s\subset\{1,\ldots,d\}$6 for the true but unknown data distribution $A_s\subset\{1,\ldots,d\}$7, or a conditional proxy $A_s\subset\{1,\ldots,d\}$8. Synthesis is split into unconditional synthesis and class-label-guided synthesis. Forecasting is treated as conditional generation of future samples given observed history, with the goal of modeling a conditional future distribution rather than just producing a point estimate. Imputation is framed through a masked observed time series $A_s\subset\{1,\ldots,d\}$9, followed by conditional generation from $$f_s^x \sim N(0, (\Omega_s^x)^{-1}), \qquad f_s^y \sim N(0, (\Omega_s^y)^{-1}),$$0.

Evaluation is explicitly multi-perspectival. Model-free metrics include Wasserstein Distance and variants, CRPS, MSE, and related distance-based measures. Model-based metrics include Predictive Score (PS), Discriminative Score (DS), and Context-FID (C-FID). The appendix defines

$$f_s^x \sim N(0, (\Omega_s^x)^{-1}), \qquad f_s^y \sim N(0, (\Omega_s^y)^{-1}),$$1

where $$f_s^x \sim N(0, (\Omega_s^x)^{-1}), \qquad f_s^y \sim N(0, (\Omega_s^y)^{-1}),$$2 is the accuracy of a post-hoc classifier trained to distinguish real from generated sequences. Predictive Score trains a post-hoc GRU regressor on synthetic data and evaluates it on real data, and Context-FID adapts FID to time series by computing Fréchet distance in a TS2Vec embedding space. GenTS also includes t-SNE-based visualization and one-call plotting utilities for forecasting and imputation, including prediction intervals (Wang et al., 18 May 2026).

5. Benchmark findings, model-selection guidance, and extensibility in the generative time-series library

The benchmark experiments are organized around synthesis, forecasting, and imputation, and the reported outcomes are strongly task dependent. For unconditional synthesis, diffusion models are generally strongest overall, especially on model-based metrics such as PS and DS, while VAEs are also competitive on Wasserstein Distance and C-FID. Among individual models, ImagenTime performs very strongly overall and is especially good on PS and DS; COSCIGAN is notable among GANs, especially on C-FID and Wasserstein Distance; and TimeVAE gives balanced but moderate results. For class-label-guided synthesis, TimeVQVAE is described as especially effective in aligning with class-structured data.

For probabilistic forecasting, CSDI and TMDM often rank among the best. The paper highlights CSDI as a strong default model and TMDM as especially effective for non-stationary data such as Stocks. It also observes that simpler baselines such as VanillaMAF and VanillaVAE can be surprisingly competitive on some datasets, which suggests that model complexity is not always necessary. For imputation, the reported pattern is clearer: diffusion-based models dominate, with CSDI consistently best or near-best, followed closely by ImagenTime and DiffusionTS, while differential equation-based models and naive baselines generally underperform.

The paper translates these results into practical guidance. It recommends DiffusionTS or ImagenTime as robust cross-task choices; COSCIGAN or ImagenTime for synthesis; FourierFlow for univariate synthesis; TimeVQVAE for class-label-guided synthesis; CSDI as a starting point for forecasting; TMDM for non-stationary series; and CSDI or other diffusion-based models for imputation. Differential equation-based models are described as promising in principle but under-explored and underperforming in the current benchmark.

Efficiency measurements in the appendix provide an additional systems-level perspective. GANs and VAEs are often expensive to train but fast at inference; diffusion models are relatively costly at inference because of iterative denoising; and differential equation-based models incur large computational overhead from repeated ODE or SDE solving. This is part of the benchmark’s purpose: not only accuracy comparison but also exposure of practical deployment tradeoffs.

The benchmark is explicitly extensible. The paper emphasizes modular architecture, custom dataset support through user-defined datamodules, custom model support through subclassing the base model interface, built-in tutorials and API documentation, open-source code, and official documentation. It also identifies three future directions: generative time series foundation models, more robust evaluation metrics because PS and DS can be brittle due to dependence on auxiliary evaluator models, and online and adaptive generation for drifting real-world series with manageable latency (Wang et al., 18 May 2026).

6. GenTS/TDS in multiscale spatio-temporal regression, and the boundary with adjacent acronyms

A third usage appears in the methodological summary of multiscale geographically and temporally weighted regression. There, the phrase “GenTS/TDS” refers to a Top-Down Scale calibration idea extended to the spatio-temporal setting. The model assigns each covariate its own spatial and temporal bandwidths in a multiscale GTWR specification,

$$f_s^x \sim N(0, (\Omega_s^x)^{-1}), \qquad f_s^y \sim N(0, (\Omega_s^y)^{-1}),$$3

with estimation by locally weighted least squares and a separable spatio-temporal kernel. The kernel can be multiplicative or additive, and the temporal component can be standard linear or cyclic/seasonal through circular temporal distance. The essential methodological point is not a new response model but a calibration scheme: bandwidth selection is converted into a structured coordinate-wise search on ordered grids, initialized at the largest bandwidths and updated by local candidate moves chosen by AICc. This is presented as a way to avoid a highly nonconvex unconstrained multidimensional optimization over many covariates and two bandwidth dimensions per covariate.

The extension preserves what the paper calls the stabilizing logic of Top-Down Scale calibration. The search begins with broad structure, large bandwidths are numerically stable, local candidate sets are small, and adaptive backfitting prioritizes covariates by a current scale-normalized contribution score. Prediction is handled through kernel sharpening rather than bandwidth re-optimization, using a power transform with default $$f_s^x \sim N(0, (\Omega_s^x)^{-1}), \qquad f_s^y \sim N(0, (\Omega_s^y)^{-1}),$$4. Monte Carlo experiments and empirical applications on Beet Yellows severity and Vaucluse house prices are used to argue that spatial multiscale structure is broadly essential and that spatio-temporal modeling adds value when temporal heterogeneity is real and well observed (Geniaux et al., 24 Mar 2026).

This GenTS/TDS usage should also be distinguished from GTaS, the “generalized thinning and shift” model for correlated spike trains. GTaS is a generative point-process construction based on a mother Poisson process, subset marking probabilities, and subset-dependent shift distributions. It produces marginally Poisson daughter processes with analytically tractable cumulants of all orders and is used to represent synchronous, sequential, cascade-like, and other time-structured higher-order correlations in neural populations. The similarity is therefore nominal rather than methodological: GTaS is a spike-train generator in computational neuroscience, not a gene-set test, benchmark library, or bandwidth-calibration scheme (Trousdale et al., 2013).