Phase-Matched Genetic Instruments for Directional Time-Course Enrichment

When biologists measure gene expression repeatedly over time, for example in blood drawn from heart-attack patients from admission to six months later, or before and after vaccination, they usually summarize thousands of changing genes as a short list of changing biological pathways. That list cannot say whether a pathway pushes the disease forward or only reacts to damage the disease has already done. This is the problem of reverse causality: the disease may change the marker, rather than the marker changing the disease. The standard way to break this ambiguity is Mendelian randomization (MR). MR treats inherited DNA variants as natural experiments, because they are fixed at conception and so cannot be altered by the disease. The difficulty is that the variants MR normally uses describe lifelong average expression, while time-course pathways often matter only during a brief phase, such as the first days of immune activation.

We propose to label every pathway found in a time-course enrichment analysis as a driver, a passenger, a feedback loop, or unresolved. The label comes from bidirectional MR in which the genetic instrument for a pathway is matched to the phase in which that pathway actually moves. Concretely, we use variants whose effect on the pathway's genes appears only in that phase, not variants chosen from resting-state tissue. The central bet is that a pathway which matters only transiently will look causally null, or will point the wrong way, when tested with static instruments, and will resolve correctly when tested with phase-matched ones. If that bet fails in the experiments below, the method reduces to ordinary pathway-level MR and we would say so.

Core Claim

Time-course enrichment already solves the question of when. Time-Course Gene Set Analysis (TcGSA) fits mixed-effects spline models to every gene in a pathway across repeated measurements. On the DALIA-1 HIV therapeutic vaccine trial (19 patients, 14 time points) it found 69 significant gene sets where gene-by-gene testing and GSEA adapted for time series found none. That paper motivates our work: it shows that pathway trajectories are the right unit of analysis. It also leaves translational users with 69 candidates and no way to rank them by causal role.

Three lines of work supply the causal side, and each misses one ingredient. Pairwise bidirectional MR can find feedback loops at the level of a single protein. SPAG11B bidirectional MR reports that genetically higher SPAG11B lowers rheumatoid arthritis risk (IVW odds ratio 0.49, 95% CI 0.40–0.61) while arthritis lowers SPAG11B (odds ratio 0.93, p = 0.004). Its instruments are static, and it tests one protein at a time. MR.RGM generalizes this to Bayesian networks of many traits with feedback, but was run only on cross-sectional GTEx expression and on individual variables, not pathways. The ADNI longitudinal pathway study comes closest to combining all three directions. It tests SNP-by-time interactions on Alzheimer's biomarker slopes and runs MAGMA pathway enrichment on the results, but it stops at association: it performs no MR and measures no expression dynamics.

Our contribution sits in the gap these three leave. We build phase-specific pathway instruments, run them through bidirectional MR, and attach the resulting direction label to time-course enrichment output.

Research Program

The pipeline has three stages.

Stage 1: phases. A TcGSA-style mixed model is fit to a longitudinal cohort. Each significant pathway gets a trajectory, and its dominant phase is defined as the time window of largest departure from baseline. In the acute myocardial infarction series analyzed by Liu et al., for instance, NK-cell cytotoxicity dominates early and metabolic programs dominate during remodeling.

Stage 2: phase-matched instruments. For each pathway and phase, we collect variants that change expression of the pathway's member genes in a cellular state resembling that phase. The source is context-specific expression-genetics resources. The clearest example in our material is the T-cell division-state study, which found genetic regulation visible only in particular activation and division states and linked it to immune-disease gene programs. Effects on member genes are combined into a single pathway-score effect, using gene loadings from the mixed model. Following MR.RGM, we keep the number of pathways jointly modeled small, because that paper shows feedback networks are unidentifiable without sparsity or strong instruments.

Stage 3: bidirectional estimation. The forward estimate for pathway score $X$ in phase $\phi$ on disease $Y$ is the usual inverse-variance-weighted ratio, computed with phase-specific variant effects:

$$\hat\theta^{(\phi)}_{X\to Y}=\frac{\sum_j \hat\beta^{(\phi)}_{Xj}\hat\beta_{Yj}\,\sigma_{Yj}^{-2}}{\sum_j (\hat\beta^{(\phi)}_{Xj})^2\,\sigma_{Yj}^{-2}}.$$

Here $\hat\beta^{(\phi)}_{Xj}$ is variant $j$'s effect on the pathway score in phase $\phi$, $\hat\beta_{Yj}$ is its effect on disease risk from a large disease genome-wide association study (GWAS), and $\sigma_{Yj}$ is the standard error of the latter. The static-instrument version is identical except that $\phi$ is replaced by resting-state effects. That substitution is exactly our ablation.

The reverse estimate $\hat\theta_{Y\to X}$ uses disease GWAS lead variants as instruments. Their effects on the phase-$\phi$ pathway score are looked up in the same context-specific resource. MR-Egger and weighted-median estimates flag variants that affect disease through routes other than the pathway, which MR calls horizontal pleiotropy.

A pathway is labeled a driver if only the forward estimate is nonzero after correction, a passenger if only the reverse is, feedback if both are, and unresolved otherwise. Unresolved is an honest and expected outcome, because reverse estimates rely on weak effects of disease variants on genes elsewhere in the genome.

Baselines

The naive reference point is TcGSA output as published: a ranked list of dynamic pathways with no direction. It amounts to calling every pathway unresolved. A method beats it only if its driver and passenger calls are correct more often than chance on cases with known truth. Simply shortening the list does not count.

The second reference is temporal precedence, meaning a cross-lagged panel model. It says a pathway is upstream if its change at one time predicts disease change at the next. It is the method most time-course papers implicitly use. We fit a random-intercept cross-lagged panel model to the same trajectories and sample sizes.

The strongest existing method is pathway-level bidirectional MR with static expression instruments. We run it two ways: pairwise IVW, MR-Egger, and weighted median, and jointly through MR.RGM using its summary-statistic input mode. Both versions get the same pathways, outcome GWAS, variant-selection thresholds, and correlated-variant pruning as our method. MR.RGM reports better true- and false-positive rates than OneSampleMR on synthetic networks. No reported result found on Lacuna for MR.RGM, or for any MR method, on time-varying or pathway-level exposures. The comparison is therefore on our shared benchmarks. Beating it means higher four-class direction accuracy in simulation at matched false-driver rate, and more correct calls on real positive and negative controls.

The ablation that isolates our key ingredient swaps phase-matched instruments for static ones while holding everything else fixed. A second ablation replaces pathway aggregation with single-gene exposures, to check that aggregation adds power rather than just averaging.

We also run the semi-supervised reverse-causality test of Testing of Reverse Causality Using Semi-Supervised Machine Learning on pathway scores. It is the only instrument-free competitor in the file.

Why not just…?

Why not just read direction off the time lags? Earlier change is not causation when traits are stable within people. In simulations built on stable-trait autoregressive structure, which is the typical structure when genetic baselines are fixed, the semi-supervised reverse-causality paper reports that random-intercept cross-lagged models falsely detected reverse causality 88% of the time. Causality for Functional Longitudinal Data further shows that when treatments and confounders feed back on each other over time, simply conditioning on the past biases estimates. We keep the lag model as a baseline and expect it to fail on the simulated passenger pathways.

Why not just run ordinary pathway MR with standard eQTLs? eQTLs are variants that shift a gene's expression level. This alternative is the real threat. If most disease-relevant pathways are also genetically regulated at rest, static instruments will give the same answers, with more variants and more power. The phase-versus-static ablation settles it. In simulations where the pathway affects disease only during a transient phase, we predict static instruments give attenuated or null forward effects. In simulations where regulation is constant over time, we predict the two agree. If they also agree on real data, the honest conclusion is that phase matching is unnecessary. The project would then deliver direction-labeled TcGSA built on MR.RGM, a smaller but still useful result.

Why not just use the instrument-free semi-supervised test on pathway scores? Collapsing genes into pathway scores does solve its dimensionality problem, which the file flags as an open question. The sample size problem remains. The test's authors report power above 0.95 only at N ≥ 500 and sharp loss below 250, while DALIA-1 has 19 people. We will measure its power on simulated pathway scores across a range of sample sizes. If it matches phase-matched MR at cohort-scale N, it becomes a complementary check rather than a competitor.

Why not just learn directions with a deep perturbation model? PerturbCausal found that no deep virtual-cell model significantly beat sparse ElasticNet on causal network recovery by F1, and that predicted effect sizes were inflated about 23-fold. Mechanisms Matter found that performance fell to near zero once the causal interaction structure changed across contexts. These results constrain our design toward linear, sparse estimators.

Proposed Experiments

The first experiment is semi-synthetic and is where ground truth is exact. We adapt the Ornstein–Uhlenbeck simulator from Mechanisms Matter, in which gene expression drifts toward a set point while being pushed around by noise. We add a disease variable and genotype-like instruments whose effects can be constant or switched on only in a phase. Pathways are planted as drivers, passengers, feedback loops, or unrelated, while varying cohort size, number of time points, instrument strength, and pleiotropy. Metrics are four-class macro-F1 and the false-driver rate, meaning passengers called drivers. The decisive check is the phase-versus-static contrast in the transient-driver regime.

The second experiment is perturbation-anchored. It exploits the fact that CRISPR guides in the Replogle K562 and RPE1 and Norman 2019 screens behave like randomized instruments. We treat guides targeting one pathway's genes as instruments and a second gene program's score as the outcome. Directions are then known by intervention, which avoids the curated-network ground truth that PerturbCausal found barely above chance.

The third experiment is real-data. For phases, we use the DALIA-1 and influenza-versus-pneumococcal vaccine time courses studied by TcGSA, and the admission-to-six-month myocardial infarction cohorts GSE59867 and GSE62646. Instruments come from the division-state T-cell resource. For disease outcomes, we use FinnGen rheumatoid arthritis (12,555 cases) and, for Alzheimer's, the Bellenguez GWAS used in the plasma proteome study. Positive and negative controls come from prior MR, run through the same estimator at single-gene resolution:

  • SPAG11B should come out as feedback.
  • In the sTREM2 proteo-genomics study, sTREM2 is protective (β = −0.236).
  • CD33 should come out causal.
  • YKL-40, which showed no genetic association, should come out passenger or unresolved.

Controls validate the MR machinery; the pathway labels are the new output. We count a real-data success only if two conditions hold. The labels must replicate across the two vaccine cohorts, or the two infarction cohorts, for pathways shared between them. And static instruments must fail on at least some pathways where phase-matched ones succeed.