Predicting the Pre-Suppression Window from Host Pathway Dynamics

When someone starts antiviral therapy, they do not become non-infectious immediately. For weeks to months, virus in the body decays along an individual-specific curve, and during that decay the person can still transmit. This is the pre-suppression transmission window, and it is not a marginal effect: pooling serodiscordant-couple cohorts, Loutfy and colleagues found a transmission rate of 0 per 100 person-years (95% CI 0–0.05) once full viral suppression was confirmed, but 0.14 per 100 person-years (95% CI 0.04–0.31) when events shortly after treatment start were included — and when those early-initiation transmissions were excluded, the rate collapsed to 0 per 100 person-years (95% CI 0–0.01). Nearly all residual transmission on therapy is packed into the initiation window. The START trial analysis by Lampe and colleagues shows the same thing from the behavioral side: immediate treatment did not reduce condomless sex (among heterosexuals it rose from 5.6% to 9.3% at month 24, $P = 0.007$), yet transmission-risk sex fell from 11.0% to 0.2% in MSM and 7.7% to 0.6% in heterosexuals ($P < 0.001$) — because suppression itself, not behavior, is what closes the window.

Epidemiological models treat the duration of this window as an empirical nuisance: a population-averaged delay. Yet the duration varies widely between individuals, germline genetics does not explain it (a Swiss HIV Cohort Study analysis of 797 long-suppressed patients found no genome-wide significant association with reservoir decay, all $P > 5 \times 10^{-8}$, Host genomics of the HIV-1 reservoir size and its decay rate during suppressive antiretroviral treatment), and cross-sectional transcriptomics of untreated patients shows that functional programs — interferon-stimulated genes, oxidative phosphorylation, apoptosis regulation — separate viral-control phenotypes cleanly (Transcriptomic meta-analysis identifies gene expression characteristics in various samples of HIV-infected patients with nonprogressive disease; A Computational Study Identifies HIV Progression-Related Genes Using mRMR and Shortest Path Tracing). The open question is whether the dynamics of these functional pathways, measured at the moment therapy starts, determine how fast the window closes — and whether that information changes population-level transmission estimates.

Core Claim

The pre-suppression transmission window is partly a host-pathway phenomenon. By sampling whole-blood transcriptomes densely during the first weeks of antiviral therapy and analyzing them with longitudinal gene-set methods — which track predefined groups of co-functioning genes over time rather than testing single genes — we can (i) identify functional pathways whose early trajectories predict an individual's time to viral suppression, (ii) convert those trajectories into calibrated, individual-level survival predictions, and (iii) propagate the resulting individualized lag distributions into epidemic simulators, replacing the fixed population-average delay that current models assume.

The proposal uses functional enrichment as the core mechanism (the unit of analysis is the pathway trajectory, not the single gene), treatment initiation window risk as the application regime, and epidemic simulation as the evaluation setting where the prediction actually matters.

Research Program

Three components, in order of dependence.

First, a pathway-kinetic survival model. Standard enrichment tools such as Fisher's exact test on differentially expressed genes, as unified in funkea, are cross-sectional: they ask whether a pathway is overrepresented at one snapshot. TcGSA fixed this for time series by modeling gene expression with linear mixed-effects models — one random effect per patient, capturing how each person's gene set drifts over repeated visits — and in the DALIA-1 HIV trial it found 69 significantly shifting gene sets where univariate gene testing and standard time-series GSEA found zero after multiple-testing correction. We reuse TcGSA's mixed-model machinery to extract, for each patient and each gene set, a baseline level and a trajectory slope, and then feed those patient-specific summaries into a survival model for time to suppression. Because patient subgroups in initiation cohorts plausibly have crossing hazard curves (fast clearers vs. delayed suppressors), we use a mixture survival form rather than a single Cox model, following Deep Cox Mixtures for Survival Regression, which on the FLCHAIN benchmark achieved an Expected Calibration Error of 0.0200 versus 0.0992 for Cox proportional hazards. Concretely, for patient $i$ with pathway-kinetic features $z_i$ (baseline level and slope per gene set), the instantaneous suppression probability per unit time is modeled as a mixture over $M$ latent subpopulations:

$$h_i(t) \;=\; \sum_{m=1}^{M} \pi_m(z_i)\, h_{0m}(t)\, \exp\!\big(\beta_m^{\top} z_i\big),$$

where $h_i(t)$ is the hazard of reaching suppression at time $t$ given no suppression yet, $\pi_m(z_i)$ are mixture weights learned from the pathway features, and $h_{0m}$ are subgroup-specific baseline hazards.

Second, a prospective initiation-window cohort — the data the field lacks. No dataset in the literature we surveyed measures longitudinal host transcriptomes from the moment therapy starts with paired viral-load decay. The closest precedents show feasibility on both sides: the influenza challenge study profiled whole blood every 8 hours for 7 days and detected infection in 94% of infected cases with an interferon-dominated signature appearing 29 hours post-exposure (A Host Transcriptional Signature for Presymptomatic Detection of Infection in Humans Exposed to Influenza H1N1 or H3N2), and TcGSA's DALIA-1 analysis handled 14 repeated samples per patient. We propose ~180 persons initiating combination antiretroviral therapy, with paired whole-blood RNA sequencing and viral load weekly for 12 weeks; the primary outcome is time to viral load below 200 copies/mL. A smaller influenza arm (~40 index cases treated early with a fast-clearing antiviral, sampled at treatment start and days 2, 4, 6) tests cross-pathogen generality, motivated by the baloxavir household trial showing a 29% reduction in household transmission within 5 days of index-case treatment (From symptom relief to household control: baloxavir redefines antiviral strategies for influenza).

Third, propagation into transmission models. SimpactCyan simulates HIV epidemics individual-by-individual and projected that universal test-and-treat in Eswatini would cut incidence by 64% by 2019 versus 12% under status quo — but its treated individuals move to suppression on a fixed or population-averaged schedule. Likewise CAST, which selects which infected network nodes to treat under budget and reduced average new infections by 56.1% versus greedy selection at a 20% budget on the ICPSR HIV network, assumes suppression is effectively a step function. We will re-parameterize both with the individualized hazard distributions from our survival model and measure how much population-average lags distort incidence projections and node-targeting decisions.

Proposed Experiments

The pipeline will be validated before any new data are collected, because the DALIA-1 dataset (19 patients, 14 time points) contains a usable time-to-event outcome: time to viral rebound after analytical treatment interruption, a mechanistic cousin of time to suppression after initiation. Experiment 1 reproduces TcGSA's 69-gene-set result on DALIA-1, then tests whether the pathway-kinetic survival model predicts rebound timing better than chance and than clinical covariates alone. Experiment 2 applies the full pipeline to the new initiation cohort, evaluating prediction of time-to-suppression by concordance index, Integrated Brier Score, and Expected Calibration Error. Experiment 3 runs the simulation study: SimpactCyan calibrated to the Eswatini setting as in the original paper, comparing incidence projections under (a) fixed population-average suppression lag, (b) empirical lag distribution, and (c) pathway-conditioned individualized hazards; and CAST on the ICPSR HIV network under the same three lag assumptions, measuring new infections averted at the 20% budget. Experiment 4, the influenza arm, asks whether the interferon signature measured at antiviral initiation predicts shedding-clearance speed and, in households, secondary symptomatic infection (the baloxavir trial's endpoint was 5.8% versus 7.6% under placebo, not statistically significant — a study with a molecular stratifier may explain that heterogeneity).

The biological payoff is the interferon paradox. Controllers show downregulated interferon-stimulated genes and suppressed oxidative phosphorylation (Transcriptomic meta-analysis...), yet acute infection shows interferon induction as the host's front-line defense (A Host Transcriptional Signature...). Whether a high-interferon, high-metabolic state at treatment start speeds drug-mediated clearance or prolongs the window is exactly what the pathway-slope-to-hazard linkage will test.

Baselines

The naive reference point is the fixed population-average suppression lag — what SimpactCyan and CAST effectively assume today. Every individual gets the same delay, and any prediction claim is vacuous. Beating it means showing, in Experiment 3, that replacing the fixed lag with individualized hazards changes projected incidence or infections-averted by a margin that matters for policy, not merely a rounding difference.

The strongest competitor for the prediction task is a Cox proportional hazards model on clinical covariates only — baseline viral load, CD4 count, age, adherence — which is the accepted clinical standard for time-to-suppression analyses. It is strong precisely because these covariates are cheap and known to correlate with decay. It will be run on the identical cohort, with the same cross-validation folds and the same tuning budget (one hyperparameter search for the mixture model, matched by an equal-budget search over clinical covariate transformations for Cox), and evaluated on the same concordance, Brier, and calibration metrics. We know calibration is where single-distribution Cox models fail when subgroups cross: on FLCHAIN, Deep Cox Mixtures reports Expected Calibration Error of 0.0992 for Cox proportional hazards versus 0.0200 for the mixture model. Beating the clinical Cox baseline means a concordance improvement outside the fold-to-fold noise and materially better calibration; if pathway features add nothing over clinical covariates, the honest conclusion is that the window is clinically determined.

For the enrichment task, the strongest baseline is TcGSA itself, which on DALIA-1 found 69 significant gene sets where standard univariate testing and time-series GSEA found zero (Time-Course Gene Set Analysis for Longitudinal Gene Expression Data). Running standard overrepresentation analysis on pre- versus post-treatment snapshots is the weaker conventional baseline; the DALIA-1 result predicts it will find nothing, and if it does find something in the initiation cohort, that itself is informative.

The key ablation isolates the longitudinal gene-set ingredient: the same survival model fed with (a) baseline-only expression summarized by static overrepresentation scores, versus (b) individual differentially expressed genes without gene-set grouping, versus (c) full pathway trajectories with slopes. If (c) does not beat (a) and (b), the claim reduces to "baseline expression helps," which is much weaker.

Why not just…?

Why not just use baseline viral load and adherence in a Cox model? This is the most serious alternative, and we treat it as the primary baseline rather than arguing around it. The indirect evidence that clinical covariates under-explain suppression lag is that viral load–CD4 relationships are weak and that viremic non-progressors maintain high CD4 despite viremia through functional programs such as apoptosis suppression (A Computational Study Identifies HIV Progression-Related Genes...), and that START-participant suppression times were heterogeneous under uniform immediate treatment (Lampe et al.). But indirect evidence is not enough; the nested comparison in Experiment 2 settles it directly, and we commit to reporting the clinical-Cox result whatever it shows.

Why not just run Fisher's exact tests on pre- and post-treatment samples? Because it has already been shown to fail on longitudinal clinical data of exactly this shape: in DALIA-1, standard univariate gene testing and standard time-series GSEA detected zero significant gene sets after multiple-testing correction, while the mixed-model approach found 69 (TcGSA). Collapsing a 12-week trajectory into two snapshots discards precisely the kinetics the proposal needs. Static curation is also incomplete: DeepSyn recovered functional categories at least 18% better than static baselines and found uncurated pathways, so we will run a secondary analysis with its knowledge-graph expansion to check curated gene sets are not missing the relevant biology.

Why not accept the Swiss Cohort null result and drop host biology entirely? The SHCS study is a genuine constraint on this proposal, and we address it head-on: it found no germline genetic association with reservoir decay (Host genomics of the HIV-1 reservoir size and its decay rate...). But it measured static DNA variants in patients already suppressed for five or more years, predicting the decay of the latent reservoir at a median of $-0.06 \log_{10}$ copies per year — a different outcome, in a different regime, from dynamic RNA measurements of active immune pathways during the weeks of active viral clearance. If transcriptomes also come up null in our cohort, the SHCS result plus ours would jointly suggest the window is governed by drug and adherence factors alone, which would be a useful negative finding.

Decisive Checks

The proposal stands or falls on three numbers. First, in the initiation cohort, the pathway-kinetic survival model must beat the clinical Cox baseline on concordance and calibration, and the trajectory ablation must beat the static-expression ablation. Second, DALIA-1 rebound prediction must work at pilot scale before the full cohort is funded. Third, in simulation, individualized lags must shift projected incidence or CAST's infections-averted at the 20% budget by enough that the fixed-lag assumption is demonstrably a modeling error rather than a convenience — the benchmark being CAST's 56.1% reduction over greedy under its current assumptions and SimpactCyan's 64% incidence reduction under a fixed-lag universal test-and-treat policy. If all three hold, the field gains a molecular predictor of exactly the quantity that Loutfy et al. identified as the sole residual source of transmission on therapy: the initiation window itself.