Cells in time-course experiments are rarely left alone. Researchers irradiate them, dose them with drugs, or switch on a growth factor such as TGF-β, and then measure gene expression over hours or days. Most dynamical methods for inferring gene regulatory networks (maps of which genes switch which other genes on or off) nonetheless model the cell as an autonomous system, in which each gene's rate of change depends only on the other genes. When a known external stimulus reaches two genes at different delays, such a model has only one way to explain the lagged co-movement: it adds a regulatory edge from the early gene to the late one. This proposal treats the measured stimulus as a first-class input. Each gene's rate of change is the sum of a sparse set of regulator effects plus a learned delayed response to the stimulus. Every term is a Gaussian process (a probabilistic model that places a distribution over smooth functions and reports how uncertain it is). Edges are then scored by the posterior probability that a regulator's contribution is nonzero.
The second half of the idea makes this tractable beyond a handful of genes. We estimate derivatives from smoothed trajectories, or from RNA velocity when splicing counts are available. That turns inference into one independent regression per gene. Because the model is additive, every Gaussian process component has a one-dimensional input. This matters because the known cost guarantees for sparse Gaussian processes degrade exponentially with input dimension and are very favourable in one dimension. The result is a model whose cost grows linearly with the number of samples and in parallel across genes, and which can be compared directly against autonomous network methods on whether it avoids stimulus-induced false edges.
Core Claim
When an experiment applies a known, time-varying stimulus, modelling that stimulus as an explicit delayed input to each gene lowers the false-positive edge rate relative to autonomous dynamical models. Comparing responses across several stimulus schedules (doses, timings) further makes edge direction identifiable in settings where a single time course cannot. Both effects can be obtained at thousands of genes with cost linear in time points, because the additive, per-gene structure keeps every Gaussian process component one-dimensional.
The closest existing work plays different roles. Gene Regulatory Network Inference with Latent Force Models supplies the modelling template. It drives mRNA equations with unobserved transcription-factor proteins given Gaussian process priors, with Hill kinetics and a translation delay, and fits them with full MCMC on 5-gene and 22-gene systems. Its forces are latent and endogenous, the graph is essentially fixed, and it does not scale. We keep the kinetic form, replace the latent force with a partially observed exogenous input, and learn the graph.
Approximate Latent Force Model Inference supplies a method we reuse. It makes latent force models fast with variational inference, gradient-matching pre-estimation, and a neural operator. It also evaluates on small, pre-specified transcriptional systems, including ionizing-radiation data, and it warns that parameters stop being identifiable once the response function becomes an unconstrained neural network. That warning constrains our design: our response functions are one-dimensional Gaussian processes with shrinkage, not multilayer perceptrons.
Modeling Temporal scRNA-seq Data with Latent Gaussian Process and Optimal Transport supplies the scaling device. It uses a truncated Hilbert-space basis expansion to scale Gaussian processes to hundreds of thousands of cells. Its dynamics live in a latent space, it has no exogenous input channel, and it produces no gene-by-gene adjacency matrix.
EgPDE-Net motivates the proposal from outside biology. It shows that giving exogenous drivers a dedicated pathway into a continuous-time model beats folding them into the state. Its reported gain is 9.85% lower RMSE than the best continuous baseline. It is deterministic, however, and has never been applied to transcription.
The Model
For gene $i$ with expression $x_i(t)$ and a known stimulus $u(t)$, we write
$$\frac{dx_i}{dt} = b_i - d_i\,x_i(t) + \sum_{j \in C_i} s_{ij}\, f_{ij}\big(x_j(t-\Delta_{j})\big) + \int_0^{\infty} h_i(\tau)\, u(t-\tau)\, d\tau .$$
The symbols are defined as follows. The constant $b_i$ is basal transcription and $d_i$ is the decay rate. The set $C_i$ contains candidate regulators of gene $i$, pre-screened to at most a few dozen transcription factors per gene using motif or ChIP-Atlas priors plus correlation. Each $f_{ij}$ is a one-dimensional Gaussian process describing how regulator $j$ affects gene $i$, allowing switch-like Hill shapes without imposing them. The delay $\Delta_j$ is a per-regulator translation lag, as in the latent force model. The weight $s_{ij}$ is a shrinkage scale under a horseshoe-type prior, so that most regulator effects are pushed to zero. Finally, $h_i$ is a causal impulse response with a Gaussian process prior, which lets the stimulus reach each gene with its own delay and shape. The score for edge $j \to i$ is the posterior probability that $s_{ij} f_{ij}$ exceeds a small effect threshold.
Scalability comes from two choices. First, following the pre-estimation step of Approximate Latent Force Model Inference, we fit smooth trajectories and their derivatives once. For single-cell data we can instead use splicing-based velocity as the derivative target, as in spliceJAC. The left-hand side then becomes an observed noisy derivative, and genes decouple into independent regressions. This also sidesteps the covariance blowup that Derivative Observations in Gaussian Process Models of Dynamic Systems incurs when derivative observations enter a joint covariance, because each gene's derivative is a scalar target in its own model. Second, each $f_{ij}$ and $h_i$ is represented with a Hilbert-space basis of $B$ functions. The per-gene cost is therefore linear in the $T$ observations and quadratic in the $|C_i| B$ basis coefficients. Because each component is one-dimensional, the result of Rates of Convergence for Sparse Variational Gaussian Process Regression implies that for a squared-exponential kernel the required number of basis or inducing functions grows only polylogarithmically, as $\log N$ with $D=1$. The full-dimensional alternative has the exponent-$D$ blowup that the same paper identifies as the main obstacle. A final forward-solve refinement, backpropagating through an ODE solver on a subset of genes, checks that the gradient-matching estimates produce trajectories that actually integrate to the data.
Research Program
The work has three parts.
The first is a controlled simulator. It uses Hill-kinetics networks with known topology, including hub-heavy degree distributions, forced by a known stimulus under several dosing schedules (step, pulse, ramp, repeated pulses). Crucially, it includes decoy configurations in which the stimulus drives two unconnected genes with different lags. We will also embed the linear, converging and bifurcating backbones used in the dyngen suite that RENGE evaluates on, adding a forcing term to them. This is our own extension, not a published benchmark.
The second part applies the model to real stimulus time courses. These are the ionizing-radiation transcriptional datasets used in Approximate Latent Force Model Inference, which span multiple doses; the MCF7 p53 data from the latent force model paper; the A549 TGF-β-induced EMT time course analysed by spliceJAC; and the hiPSC CRISPR time series from RENGE. In the RENGE data, the knockout at transduction is treated as a gene-specific step input.
The third part studies the approximation. We ask how basis size and derivative-estimation error trade off against edge recall on rare transition states. This is the concern raised in Approximation-Aware Bayesian Optimization, that budgets allocated for global fit smear out localized behaviour. If the globally allocated basis measurably hurts recovery of edges active only in transient states, such as hybrid epithelial–mesenchymal states, we will test a time-localized basis concentrated where $u(t)$ changes.
Baselines
The naive reference point is correlation between regulator expression and target derivative, with the stimulus ignored. It tells us how much any structured model buys at all.
The first strong competitor is a tree-ensemble derivative regressor of the dynGENIE3 type, run with the stimulus added as an extra candidate regulator. We will include lagged copies of $u$ at the same lag grid our impulse response spans, so it has access to the same information. The review Machine Learning Methods for Gene Regulatory Network Inference describes GENIE3-style tree ensembles as the standard supervised-regression approach. This baseline is the most dangerous one because it shares our per-gene decoupling and is cheap. Both methods get identical smoothed derivatives, identical candidate-regulator sets, and matched hyperparameter search budgets.
The second is RENGE, the strongest published method for separating direct from indirect effects in time-series perturbation data. On hiPSC it reports an average correlation of about 0.3 for predicting day-5 expression changes (RENGE). Beating it means exceeding 0.3 on that task with the same held-out knockouts, and matching or exceeding its agreement with ChIP-Atlas on the 19 genes it scores. RENGE's direct-versus-indirect advantage on dyngen is reported only qualitatively; we found no reported edge-recovery AUPRC on Lacuna for dyngen.
The third is spliceJAC, run per cell state on the A549 EMT data. It represents the cheap linearized alternative, and we found no reported quantitative edge-recovery result on Lacuna for it on A549.
The fourth covers autonomous Gaussian process dynamics. We run our own model with the stimulus term deleted. It serves as the key ablation and isolates the contribution of the exogenous channel. We also compare against a variational latent force model in which the stimulus's effect must be absorbed by a latent force. For that model the only reported number on Lacuna is a runtime figure, a ~20,000x prediction speedup of the neural operator over the standard variational latent force model; no edge-recovery number was found.
A second ablation replaces the Gaussian process impulse response with an instantaneous linear stimulus term, testing whether learning the delay is what matters.
The file offers a reported AUPRC of 0.89 for GRNFormer (Machine Learning Methods for Gene Regulatory Network Inference). The benchmark behind it is unspecified, so we will not treat it as a comparable target.
Why not just…?
Why not just add the stimulus as another input feature to GENIE3 or a transformer network model? This is the real threat, and the proposal is built so that it can lose to it. Lagged stimulus features let a tree ensemble explain a delayed response in principle. The open question is whether it does so reliably with a few dozen time points, and whether it assigns the lagged co-movement to the stimulus rather than to a spurious regulator edge. ST-ResNet and Fusing Exogenous Context into Spatio-Temporal Neural Networks report that naive concatenation underperforms dedicated exogenous pathways, but that evidence comes from traffic forecasting, not network recovery. The decoy experiment settles it. If augmented dynGENIE3 matches our false-positive rate on decoys and our AUPRC, we would conclude that the contribution is the modelling of the stimulus, not the Gaussian process. The paper would then be reframed around that finding and around the calibrated edge probabilities, which tree ensembles do not provide.
Why not just use a neural ODE with an exogenous input, as EgPDE-Net does? Such a model can fit trajectories. It does not yield a sparse adjacency matrix or edge probabilities without additional machinery. The latent Gaussian process work reports that scNODE overfits and loses coherence on hard extrapolation on the 236,285-cell iPSC data (LGP-OT). Approximate Latent Force Model Inference notes that parameters become unidentifiable once the response function is an unconstrained network. We will still run an exogenous-input neural ODE on the held-out dosing-schedule task. If it wins clearly on prediction, we would report that our advantage is limited to structure recovery.
Why not just fit each stimulus condition separately and compare the networks, as differential network inference does? Double Optimal Transport infers networks that differ between conditions from unpaired samples. It asks which edges change, whereas we want a single mechanism shared across schedules, with the schedule difference acting as a natural intervention. Fitting conditions separately discards exactly the contrast that should identify edge direction. The multi-schedule experiment below measures that loss directly.
Proposed Experiments
The first decisive check is the decoy test in simulation. Across hundreds of simulated networks, we measure the fraction of stimulus-induced decoy pairs that each method reports as edges, together with overall AUPRC and early precision against the true graph. The hypothesis fails if our model's decoy false-positive rate is not clearly below that of both the autonomous ablation and lag-augmented dynGENIE3.
The second is the identifiability test. Holding the network fixed, we vary the number of distinct dosing schedules from one to four and track the accuracy of edge direction. The prediction is that direction accuracy rises with schedule diversity for our model. It should stay flat for autonomous models, which ignore the schedule, and for per-condition fitting.
The third test is generalization to unseen doses on the radiation data. We train on some doses and predict held-out dose trajectories. Error and predictive log-likelihood are compared with the variational latent force model and the exogenous neural ODE.
The fourth is scaling. We fit the A549 and hiPSC data at 1,000, 5,000 and 10,000 genes, reporting wall-clock time against basis size. We also check against the ChIP-Atlas and STRING references used by RENGE and HyperG-VAE that edge quality does not collapse as genes are added.
As a sanity check, on the seven BEELINE-style datasets of HyperG-VAE, which have no stimulus, the model reduces to its autonomous form. It should be roughly competitive there, not dominant; a large win would suggest leakage from the priors rather than a real effect.
References
- Gene Regulatory Network Inference with Latent Force Models
- Approximate Latent Force Model Inference
- Modeling Temporal scRNA-seq Data with Latent Gaussian Process and Optimal Transport
- EgPDE-Net: Building Continuous Neural Networks for Time Series Prediction with Exogenous Variables
- <scp>spliceJAC</scp> : transition genes and state‐specific gene regulation from single‐cell transcriptome data
- Derivative Observations in Gaussian Process Models of Dynamic Systems
- Rates of Convergence for Sparse Variational Gaussian Process Regression
- RENGE infers gene regulatory networks using time-series single-cell RNA-seq data with CRISPR perturbations
- Approximation-Aware Bayesian Optimization
- Machine Learning Methods for Gene Regulatory Network Inference
- Predicting citywide crowd flows using deep spatio-temporal residual networks
- Long-Term Mobile Traffic Forecasting Using Deep Spatio-Temporal Neural Networks
- Double optimal transport for differential gene regulatory network inference with unpaired samples
- Inferring gene regulatory networks by hypergraph generative model