Splitting Self-Signals from Neighbor-Signals in Time-Resolved Gene Regulatory Networks

Most molecular measurements of tissue signaling are blind to who is talking to whom. A ligand such as prostaglandin E2 (PGE2) is measured as a bulk concentration, yet its biological effect depends entirely on its origin: when CD4+ T cells make PGE2 for themselves, it fuels inflammatory colitis, while the same molecule delivered by neighboring non-lymphoid cells induces protective regulatory T cells (mPGES-1-Mediated Production of PGE2 and EP4 Receptor Sensing Regulate T Cell Colonic Inflammation). Gene regulatory network (GRN) inference — reconstructing the directed wiring from transcription factors to their target genes — and longitudinal expression profiling have each matured separately, but no existing pipeline combines them to answer the question that decides therapeutic outcomes for shared ligands: for this signaling molecule, in this cell type, at this moment, is the regulatory effect driven by the cell's own autocrine output or by a paracrine delivery from the microenvironment?

This proposal builds a deliberately simple model that answers that question. The core idea: in a time series of single-cell expression profiles across multiple cell types, a transcription factor's regulon activity (the collective behavior of its target genes, which tracks functional impact far better than the TF's own mRNA; Accurate Transcription Factor Activity Inference to Decipher Cell Identity from Single‐Cell Transcriptomic Data with MetaTF) in cell type $c$ at time $t+1$ should be predictable from ligand production in cell type $c$ at time $t$ (autocrine) and from ligand production in other cell types at time $t$ (paracrine). Fitting one sparse, lagged regression per receptor–TF module yields a signed, time-resolved decomposition of each ligand's regulatory influence into a self-loop component and a neighbor component. The output is an origin-resolved dynamic GRN: for every signaling molecule, a quantitative attribution of how much of its downstream regulatory effect is autocrine versus paracrine, and how that attribution shifts over time.

Core Claim

The claim is that the autocrine/paracrine split, which today can only be established with expensive conditional genetic experiments (as in the PGE2 study, where the authors needed mPGES-1 knockout T cells and EP4-floxed recipients to separate the loops), can be recovered computationally from ordinary multi-time-point single-cell RNA data — provided the model uses time lags and cell-type-resolved sender terms rather than static ligand–receptor scores. This is distinct from the closest existing work. SigXTalk models ligand–receptor–TF–target pathways with a hypergraph and tracks them across discrete time points, but its own authors state that it "relies on static gene expression snapshots" and it never mathematically decouples the self-derived and microenvironment-derived pools of the same ligand (Dissecting crosstalk induced by cell-cell communication using single-cell transcriptomic data). CLARIFY co-infers cell–cell interactions and GRNs from spatial data but is strictly a single-time-point method and assumes spatially adjacent cells have similar GRNs — exactly the assumption that fails when a T cell and an adjacent stromal cell execute opposing programs in response to the same PGE2 (CLARIFY: cell-cell interaction and gene regulatory network refinement from spatially resolved transcriptomics). MMRA disentangled intrinsic from stromal signatures in colorectal cancer, but on bulk tumors and without any temporal resolution (MicroRNA–mRNA interactions underlying colorectal cancer molecular subtypes). The gap — a dynamic GRN whose inter-cellular edges carry an explicit origin label — is real and, on the evidence in this literature, unoccupied.

Research Program

The method, which we call AutoPara, has three components, each chosen to stay on the simple side of the "simpler is better" tension that the geneRNIB benchmark exposed: complex foundation models (scPRINT) and multi-omic pipelines were consistently beaten by plain Pearson correlation and GRNBoost2 across thousands of perturbation conditions (geneRNIB: a living benchmark for gene regulatory network inference). We therefore build on sparse regression, not deep learning.

First, for each cell type $c$ and each time point, we compute regulon activity scores $a_{r,c}(t)$ for every transcription factor $r$ using the enrichment-of-targets approach validated in metaTF, rather than the TF's own mRNA level. Second, for each known ligand $L$ with receptor $R$, we compute a sender score $s_{c'}^L(t)$ — ligand gene expression weighted by receptor expression in the receiving population — for every cell type $c'$. Third, we fit, per receiver cell type and per receiver TF:

$$a_{r,c}(t+1) \;=\; \beta_a\, s_c^L(t) \;+\; \sum_{c' \neq c} \beta_p^{c'}\, s_{c'}^L(t) \;+\; \gamma\, a_{r,c}(t) \;+\; \epsilon$$

where $\beta_a$ is the autocrine coefficient (the cell's own ligand production at the previous time point predicting its regulon shift now), $\beta_p^{c'}$ are paracrine coefficients (each neighbor type's ligand production), and $\gamma$ captures cell-intrinsic persistence such as maturation programs that continue without any sender input. All coefficients are fitted with an elastic-net penalty and non-negativity on $\beta_a$, $\beta_p$ where biology demands it. The headline quantity is the origin fraction

$$\rho_{L,c} \;=\; \frac{\beta_a}{\beta_a + \sum_{c'} \beta_p^{c'}}$$

which is near 1 for a purely autocrine loop, near 0 for a purely paracrine input, and time-varying when the regime switches. The lag structure is what makes identification possible: an autocrine signal and a paracrine signal produce the same static correlation, but only the autocrine term has its sender fluctuation locked to the receiver's own earlier expression state within the same cell type.

An important interpretive use follows directly: the model separates intrinsic maturation from externally-triggered change, the confound that plagues longitudinal studies. Poon et al. deliberately purified cardiomyocytes to >95% before profiling maturation, sacrificing all heterotypic signaling to remove this ambiguity (Transcriptome-Guided Functional Analyses Reveal Novel Biological Properties and Regulatory Hierarchy of Human Embryonic Stem Cell-Derived Ventricular Cardiomyocytes Crucial for Maturation); AutoPara is designed to work on the unpurified tissue where such questions are usually unanswerable. It also speaks to the "shadow problem" reported for Kolmogorov-Arnold GRN models, where branching trajectories overlap in expression space and edge functions become ambiguous (Kolmogorov-Arnold Network for Gene Regulatory Network Inference): adding sender identity and lag as conditioning variables disambiguates trajectories that are identical in target-gene space but differ in upstream origin.

Baselines

The do-nothing reference is a single pooled GRN fitted on all cells and time points jointly, ignoring cell type and time — the default output of standard pipelines. Anything that cannot beat this on origin-relevant questions has no value, and it defines the accuracy floor for edge recovery. It will be run on exactly the same inputs as AutoPara with the same prior ligand–receptor list.

The strongest existing competitor is SigXTalk, the hypergraph method that connects ligands, receptors, intermediate TFs, and targets, scores pathway regulatory strength with Random Forests, and computes signal fidelity and specificity metrics (Dissecting crosstalk induced by cell-cell communication using single-cell transcriptomic data). It is the only published method spanning all three of our directions simultaneously. On the SERGIO simulator with known ground truth, SigXTalk reported an 18% increase in AUROC and a 15% increase in AUPRC over 12 representative GRN methods including GENELink and PIDC; we treat its recovered-pathway AUROC on the same simulations as the number to exceed. For fairness, we will run SigXTalk on identical data with its default priors, and additionally in a per-time-point mode so its pathway scores can be differenced into a pseudo-dynamic signal — giving it every chance to look temporal without our model's advantages.

The strongest simple method is GRNBoost2 (gradient-boosted trees decomposing inference into per-gene regression), which geneRNIB found consistently outperforms complex models across thousands of perturbation conditions (geneRNIB: a living benchmark for gene regulatory network inference). We will run it both per cell type (the two-step cluster-then-infer design) and, as a direct ablation of our key ingredient, on our lagged sender features — replacing the elastic net with boosted trees on the same autocrine/paracrine covariates. This ablation isolates whether our contribution is the lagged origin-resolved feature construction (which the ablation keeps) or the specific regression head (which it swaps). Beating SigXTalk's pathway AUROC on matched simulations, and beating plain GRNBoost2 on origin-label accuracy where ground truth exists, counts as success. No reported result was found on Lacuna for absolute geneRNIB metric values, so our GRNBoost2 comparison must establish its own numbers on our splits rather than compare to published geneRNIB tables.

Why not just…?

Why not just cluster cells first and run GENIE3 per cluster? This two-step design is the natural first suggestion, and the literature gives direct evidence against it for mixed or transitional populations: VMPLN showed that joint inference of cell identity and GRNs consistently outperformed sequential pipelines including GENIE3 in pAUPRC and early precision ratios when populations are mixed, and CLARIFY showed that ignoring microenvironmental context inflates GRN false positives (Single-cell gene regulatory network analysis for mixed cell populations; CLARIFY: cell-cell interaction and gene regulatory network refinement from spatially resolved transcriptomics). More fundamentally, per-cluster GRNs contain only intracellular edges and cannot, even in principle, say whether a ligand's effect is self-derived or neighbor-derived — the origin question is not a accuracy gap but an unmodeled variable. We will run this pipeline as a baseline anyway, and if it matches AutoPara on any origin-labeled metric we will report that plainly; we expect it to match on TF–target edge recovery while failing on origin attribution by construction.

Why not just use GRNBoost2 or Pearson correlation, since geneRNIB says simpler wins? This is a genuine threat, and we take it seriously by making AutoPara simple: it is an elastic net, the same family as the sparse regressions behind the Inferelator's AMuSR approach at million-cell scale (High-performance single-cell gene regulatory network inference at scale: the Inferelator 3.0). The honest possibility is that lagged correlation alone, without our regression structure, recovers most of the signal. The decisive experiment is the ablation above: gradient-boosted trees on the same lagged sender features. If lagged GRNBoost2 on our features matches AutoPara, we will publish the feature construction as the contribution and drop the model — the lagged, cell-type-resolved sender design, not the fitting machinery, is the idea. What correlation genuinely cannot do, on the file's evidence, is assign edge signs and origin labels: scKAN authors note that correlation and tree ensembles cannot distinguish activation from repression, and SigXTalk showed pairwise correlation misses multi-step ligand–receptor–TF crosstalk (Kolmogorov-Arnold Network for Gene Regulatory Network Inference; Dissecting crosstalk induced by cell-cell communication using single-cell transcriptomic data). Also, correlation in mixed profiles actively misassigns stromal signals to tumor cells, the exact origin error this proposal targets (MicroRNA–mRNA interactions underlying colorectal cancer molecular subtypes).

Why not just score static ligand–receptor pairs from databases, as NicheNet- or KEGG-style tools do? Because static presence is precisely the quantity that fails on the motivating biology. Banerjee et al. proved that identical EP4-mediated sensing of identical PGE2 produces opposite downstream programs depending on source, so a static score is structurally incapable of predicting the outcome (mPGES-1-Mediated Production of PGE2 and EP4 Receptor Sensing Regulate T Cell Colonic Inflammation). And geneRNIB warns that static prior databases systematically discard context-specific links absent from reference sets (geneRNIB: a living benchmark for gene regulatory network inference). Temporally, the alveolar macrophage series showed that lethal and non-lethal pathogens are indistinguishable in early snapshots — the discriminating signal is IRF7's kinetic timing, visible only with lags (Large Scale Comparison of Innate Responses to Viral and Bacterial Pathogens in Mouse and Macaque). We will nevertheless run static ligand–receptor scoring as a baseline, since beating it is a low bar that must be cleared publicly.

Why not just profile purified single lineages over time and avoid the ambiguity altogether? That design eliminates paracrine inputs rather than measuring them. Poon et al.'s purified cardiomyocyte time series can only reveal intrinsic maturation; it cannot say which regulatory shifts would have been altered by neighbors (Transcriptome-Guided Functional Analyses Reveal Novel Biological Properties and Regulatory Hierarchy of Human Embryonic Stem Cell-Derived Ventricular Cardiomyocytes Crucial for Maturation). Conversely, isolated cancer cultures reveal autonomous autocrine loops like gastric cancer's self-produced acetylcholine while missing the co-existing neural paracrine input seen in vivo (Acetylcholine acts through M3 muscarinic receptor to activate the EGFR signaling and promotes gastric cancer cell proliferation). We repurpose this limitation as a validation instrument: in single-lineage time series, AutoPara should attribute essentially all ligand regulation to the autocrine term, and deviation from that expectation on isolated-culture data (the alveolar macrophage series) would flag a modeling failure.

Proposed Experiments

Ground-truth evaluation comes first. We will use the scMultiSim simulator, which generates known truth for both inter-cellular interactions and intracellular GRNs and was CLARIFY's evaluation substrate (CLARIFY: cell-cell interaction and gene regulatory network refinement from spatially resolved transcriptomics), extended across simulated time steps so that sender–receiver edges carry known origin labels (same-type vs. cross-type). Metrics: AUROC and AUPRC for sender→receiver→TF pathway recovery against SigXTalk (whose matched-simulator reference point is an 18% AUROC / 15% AUPRC improvement over 12 conventional GRN methods, on SERGIO) and against CLARIFY-style static co-inference; and origin-label accuracy — the fraction of recovered inter-cellular edges correctly labeled autocrine or paracrine — for which no existing method produces a prediction at all, making pooled GRN and per-type GRNBoost2 the references. Second, on the BEELINE/BoolODE trajectory suite, we check that adding origin and lag features does not degrade intracellular TF–target recovery relative to scKAN, which outperformed nearest competitors by 5.40%–28.37% in AUROC and 1.97%–40.45% in AUPRC (Kolmogorov-Arnold Network for Gene Regulatory Network Inference).

Real-data validation then proceeds on three axes. On the rat burn wound atlas — 31,503 cells across 9 longitudinal time points with a documented biphasic regulatory shift from Fos/Jun/Stat motifs before day 7 to Tead4/Klf remodeling after (A scATAC-seq atlas of stasis zone in rat skin burn injury wound process) — AutoPara must reproduce that switch with the correct timing, and its origin attributions should assign early immune-to-stem paracrine inputs where the biology reports cross-talk, not self-loops. On the alveolar macrophage series — six time points (0, 1, 2, 6, 12, 24h) of isolated macrophages under lethal versus non-lethal pathogens (Large Scale Comparison of Innate Responses to Viral and Bacterial Pathogens in Mouse and Macaque) — the method must show near-total autocrine attribution, and must detect the stalled early IRF7 kinetics as a lagged autocrine coefficient difference between lethal and non-lethal challenge; missing this would falsify the lag mechanism. Third, on the mixed-population PBMC datasets used to validate VMPLN (Kang, 7,217 cells across 10 types; Zheng, 7,424 cells across 6 types) and the Liao COVID-19 macrophage set (29,980 cells, 8 patients, moderate versus severe), we test whether origin fractions for inflammatory ligands separate disease severity groups better than static ligand–receptor scores or pooled regulon activity alone. No reported result was found on Lacuna for absolute VMPLN metric values on these datasets, so we will establish matched numbers ourselves under identical preprocessing.

The decisive biological check is a computational prediction matched to a known genetic result: on gut inflammation data, AutoPara must assign PGE2 regulation of STAT3-pathway regulons in CD4+ T cells predominantly to the autocrine term and FoxP3-regulon induction to a paracrine term from non-lymphoid senders, reproducing the conditional-genetic findings of the mPGES-1 study, where colitic mice showed a nearly 5-fold increase in PGE2 and the autocrine/paracrine arms produced opposite T cell fates (mPGES-1-Mediated Production of PGE2 and EP4 Receptor Sensing Regulate T Cell Colonic Inflammation). That study leaves the exact protective paracrine sender unidentified — an open problem the model's per-sender $\beta_p^{c'}$ coefficients are built to address — and the same logic applies to the unresolved origin of IL-6 in oral squamous cell carcinoma (Interleukin-6 mediated inflammasome activation promotes oral squamous cell carcinoma progression via JAK2/STAT3/Sox4/NLRP3 signaling pathway). If AutoPara's origin fractions reproduce known genetics where genetics exist and generate sender-level predictions where they do not, the proposal will have done its job: converting cellular origin ambiguity from an experimental bottleneck into a computable, time-resolved quantity.