Which Treg Circuits Translate? Cross-Species Conservation of Regulatory T-Cell Networks in Tumors versus Autoimmune Lesions

The same cell type sits at the center of two opposite therapeutic problems. Regulatory T cells (Tregs) — the immune system's dedicated suppressor cells — are too weak in tumors, where they help the tumor silence attacking immune cells, and too strong in autoimmune disease, where they fail to restrain the immune system from attacking the body's own joints, kidneys, and nerves. Oncology therefore wants to deplete or disarm Tregs inside tumors, while rheumatology wants to expand or strengthen them elsewhere. The uncomfortable fact, documented in both literatures, is that interventions built for one side of this divide bleed into the other: systemic Treg depletion to break tumor immunosuppression triggers colitis, hypophysitis, and nephritis (ROSE12, a novel anti-CTLA-4 FcγRs binding-enhanced antibody activated by extracellular adenosine triphosphate, shows tumor-selective regulatory T-cell depletion and antitumor efficacy without systemic immune activation), while tolerance-inducing drugs that expand Tregs to treat autoimmunity could, in principle, feed tumor immunosuppression — a possibility raised but never tested by LPX-TI641, a Tim3/4 Agonist, Induces Long-Term Immune Tolerance in Multiple Sclerosis Models.

This proposal argues that the missing piece is a map that separates two kinds of Treg regulatory circuit: circuits that are shared between the tumor microenvironment (TME) and autoimmune lesions — and therefore dangerous to touch — and circuits that are specific to one context and therefore safe, or at least safer, targets. And because every preclinical test of these interventions happens in mice, the map must be built to answer a second question at the same time: which of those circuits are conserved between mouse and human, so that a rodent result can be trusted at all. Single-cell RNA sequencing (scRNA-seq — measuring gene expression in individual cells rather than averaged over tissue) is the measurement tool; cross-species regulatory conservation is the analytical core; autoimmunity and the TME are the two disease regimes being contrasted.

Core Claim

The gene regulatory network of intratumoral Tregs can be decomposed into modules that differ in two independent ways: whether they are conserved across species within a disease context, and whether they are conserved across disease contexts within a species. A circuit that is strongly conserved between mouse and human tumor Tregs but weakly shared between tumor Tregs and autoimmune-lesion Tregs is a high-value, low-collateral target. A circuit conserved in both axes is where autoimmune toxicity will come from. No existing study measures either axis for tissue-infiltrating Tregs, let alone both.

The claim has a concrete anchor. The most detailed existing reconstruction of intratumoral Treg control, built from 51,195 single-cell CD4+ T-cell transcriptomes across human head and neck tumors, blood, and tonsils, identifies BATF as the central regulator of a TNFR-activated suppressive Treg state (4-1BB+, OX40+, CD39+, CD96+), and shows that BATF knockout paradoxically increases some immunosuppressive markers (ICOS, LAG3, neuropilin-1) (1049 Reconstruction of gene regulatory networks dissects transcriptional control of intratumoral regulatory T cells). That study was entirely human and explicitly left open whether the BATF circuit is conserved in animal models and whether it operates in autoimmune lesions. Meanwhile, the tools that could answer this — SATURN, which aligns cells across species using protein-language-model embeddings instead of strict one-to-one gene matching (Towards Universal Cell Embeddings: Integrating Single-cell RNA-seq Datasets across Species with SATURN), and CACIMAR, which scores conservation of regulatory networks and intercellular circuits directly (CACIMAR: Cross-species Analysis of Cell Identities, Markers, Regulations and Interactions Using Single-cell RNA Sequencing Data) — have only ever been run on healthy whole-body atlases, embryogenesis, or injured retina. The immunological atlas covering the most species, PBMCatlas with 77,957 peripheral immune cells across 12 vertebrates, is explicitly limited to circulating cells and omits tissue microenvironments entirely (Molecular mechanisms governing circulating immune cell heterogeneity across different species revealed by single‐cell sequencing). The intersection is genuinely empty.

Research Program

We will build a two-axis atlas of CD4+ Treg states and compute, for every regulatory module, a context-selectivity index. The reused data are the human HNSCC CD4+ dataset (51,195 cells) and the PBMCatlas peripheral cells, which serve as the healthy-circulating baseline defining the "core" Treg program. The new data are generated symmetrically: scRNA-seq of sorted CD4+ T cells from mouse TME models (MC38 and Hepa1-6, the same lines used to validate ROSE12) and from mouse autoimmune lesions in three established models — MOG-induced EAE spinal cord, collagen-induced arthritis joints, and pristane-induced lupus kidney (the models used in LPX-TI641, a Tim3/4 Agonist, Induces Long-Term Immune Tolerance in Multiple Sclerosis Models and Disulfiram alleviates pristane-induced lupus via inhibiting GSDMD-mediated pyroptosis). Donors are age-, sex-, and batch-matched, addressing the covariate confounding that PBMCatlas itself flags as a limitation.

Integration runs through SATURN's "macrogene" embeddings, which matter here for a specific technical reason: the TNFR family at the heart of the intratumoral Treg phenotype (4-1BB, OX40, CD40) is a duplicated gene cluster where strict one-to-one orthology is exactly the mapping that fails (Towards Universal Cell Embeddings: Integrating Single-cell RNA-seq Datasets across Species with SATURN documents over 50% gene loss under 1:1 mapping in distant comparisons, and functional drift of nominally orthologous genes such as MYOC). Regulatory networks are then inferred per context with the GENIE3-based approach CACIMAR uses, and conservation is scored per module. Formally, for a module $M$, let $C_{\text{sp}}(M, d)$ be the network conservation score between mouse and human both in disease context $d$ (tumor or autoimmune lesion), and $C_{\text{ctx}}(M, s)$ the same score computed between tumor and autoimmune-lesion Tregs within species $s$. The context-selectivity index is

$$S(M) \;=\; C_{\text{sp}}(M,\text{tumor}) \;-\; \tfrac{1}{2}\big(C_{\text{ctx}}(M,\text{mouse}) + C_{\text{ctx}}(M,\text{human})\big).$$

High $S$ marks tumor-specific, cross-species-robust circuits; low or negative $S$ marks circuits shared with autoimmune inflammation. The precedent for this logic is the Alzheimer's epigenomic study, which found 90% of promoters and 84% of enhancers conserved between the CK-p25 mouse and human brain, but showed that only the immune-activated, PU.1-driven enhancers — not the repressed neuronal promoters — carried a 4.4-fold enrichment of human GWAS risk variants (Conserved epigenomic signals in mice and humans reveal immune basis of Alzheimer’s disease). Conservation alone was not the finding; conservation plus disease-specificity was. We port that logic from bulk epigenomics to single-cell regulatory networks in two disease regimes.

A dynamic extension uses PROFET, which reconstructs continuous cell-state trajectories from static snapshots and reduced trajectory prediction error by more than 50% relative to linear optimal-transport baselines (PROFET Predicts Continuous Gene Expression Dynamics from scRNA-seq Data to Elucidate Heterogeneity of Cancer Treatment Responses), to ask whether tumor-selective Treg depletion permanently redirects the differentiation trajectory of remaining Tregs.

Proposed Experiments

The first experiment is purely computational and delivers the map: transfer labels between mouse autoimmune-lesion Tregs and human HNSCC tumor Tregs using SATURN, then score regulatory-network conservation with CACIMAR's approach. The reference magnitude is SATURN's 85.8% median label-transfer accuracy with a 119% gain over the orthology-dependent SAMap on the frog–zebrafish embryogenesis benchmark (Towards Universal Cell Embeddings...); a sharply lower accuracy for cross-context transfer than cross-species-within-context transfer would itself be a result — it would mean disease context, not evolutionary distance, dominates Treg state divergence.

The second experiment tests the map's central prediction in vivo. Murine Batf knockout in tumor-infiltrating Tregs should reproduce the human paradox (increased ICOS, LAG3, neuropilin-1) if the BATF module scores as tumor-conserved across species; if it does not, the atlas has caught a rodent–human divergence exactly where preclinical testing of BATF-targeted TME interventions would have been misleading. This is the failure mode documented across the translation literature, e.g. murine BC094916 suppressing myeloma while its human ortholog IFI16 does not (Bridging the Cross Species Clinical Translation Gap).

The third experiment is the safety arm. Apply the ATP-gated anti-CTLA-4 antibody ROSE12 ($K_D \approx 4$ nM at 100 µM extracellular ATP, no detectable binding at 0 µM; a 30- to 300-fold wider therapeutic window than non-switching versions) (ROSE12...) in the MC38 model, then scRNA-sequence residual intratumoral Tregs and Tregs from autoantigen-bearing distant tissue (colon). The atlas's prediction is testable: modules ranked high on $S$ should be depleted or rewired only inside the tumor, while shared-context modules should shift in the colon too. ROSE12's developers performed no such transcriptomic profiling, so this fills a stated gap rather than duplicating a result. Success on all three arms means the $S$ index ranks targets before anyone spends a clinical-trial budget on them.

Baselines

The do-nothing reference point is the field's current practice: assume rodent efficacy predicts human outcome and pick targets from single-species networks. The translation literature quantifies why this fails — Bridging the Cross Species Clinical Translation Gap catalogues candidates with complete rodent efficacy that failed in humans, and even ROSE12's developers had to escalate through humanized mice and cynomolgus monkeys (well tolerated at 30 mg/kg/week) because rodent Fc-receptor biology misstates human immune-related adverse events (ROSE12...). Any performance we report for the $S$ index should be read against the implicit accuracy of that practice, which is unmeasured and, on the available evidence, poor.

The strongest methodological competitor is the best existing cross-species pipeline run without our context contrast: SATURN for alignment plus CACIMAR-style network conservation scoring, applied to the same atlas with the same compute and the same hyperparameter search budget we use, but scoring conservation only between species and never between disease contexts. SATURN is the strongest aligner on record (85.8% median accuracy, 119% gain over SAMap; ~335,000 mammalian cells integrated) and CACIMAR matched 11 of 15 cell types and found 139 conserved markers across 38,305 mouse and 40,236 zebrafish retinal cells — these are the numbers to beat or match on our data. This baseline is also the key ablation: it isolates the proposal's actual ingredient, which is not a new integration algorithm but the two-axis $S$ contrast. Beating it means the context-selectivity ranking, not raw conservation, predicts the two in vivo outcomes — BATF-knockout phenotype agreement and tumor-versus-colon spillover under ROSE12. A second ablation replaces SATURN macrogenes with standard 1:1 ortholog tables in the same pipeline; the TNFR-cluster genes are where we expect the drop, and if there is none, that is an honest finding against our own design choice.

A third baseline is within-species differential expression: human HNSCC tumor Tregs versus human autoimmune-lesion Tregs, no mouse data at all. It competes directly on target ranking and we run it on our atlas so the comparison is fair. It will win on cost; the question the in vivo experiments settle is whether it can substitute, which we address below.

Why not just…?

Why not just use standard 1:1 ortholog tables? Because the gene family at the center of this phenotype is duplicated: 4-1BB, OX40, and CD40 sit in paralogous TNFR clusters where one-to-one mapping is precisely the operation that misassigns or discards genes, and SATURN's authors document over 50% gene loss under 1:1 mapping for distant comparisons plus functional drift in nominally clean orthologs like MYOC (Towards Universal Cell Embeddings...); PBMCatlas names low ortholog ratios as its own limiting distortion (Molecular mechanisms...). For mouse–human alone this critique is genuinely weaker — the two genomes are close — which is why the 1:1 ablation above is run rather than argued away. If the ablation shows no loss on our data, we will report that the expensive machinery was unnecessary here.

Why not just do bulk RNA-seq of tumors and lesions? Because the decisive population is rare and the signal would be averaged away. The BATF-governed TNFR-activated Treg state is a small CD4+ subpopulation within a mixed tumor; the same study that found it also showed bulk-level averaging cannot resolve it (1049 Reconstruction of gene regulatory networks...). The same lesson holds elsewhere: PROFET found a pre-existing drug-resistant subpopulation with six surface markers invisible to population averaging (PROFET...), and the lupus pyroptosis signal concentrates in glomerular infiltrating macrophages that whole-kidney bulk transcriptomics would dilute (Disulfiram alleviates pristane-induced lupus...).

Why not just compare human tumor Tregs to human autoimmune-lesion Tregs and skip the mice? This is the most serious alternative, and part of our design already concedes it: the within-human context contrast is one of the two axes of $S$. What the human-only comparison cannot deliver is the translational verdict. Human autoimmune-lesion Treg data at single-cell resolution is scarce — the tolerance-inducing compound that motivates this direction was validated ex vivo in only six MS patients (LPX-TI641...) — while every intervention we would rank gets tested in mice first. A circuit that looks tumor-specific in humans but has drifted in rodents will pass a human-only screen and then mislead at the preclinical stage, the BC094916/IFI16 pattern exactly (Bridging the Cross Species Clinical Translation Gap). The BATF-knockout experiment is designed to detect precisely this: if the human-only ranking and the two-axis ranking disagree, the in vivo phenotype tells us which one to trust.

Why not just use linear trajectory tools like standard optimal transport for the dynamics? Because the transitions we care about — whether residual Tregs after depletion redifferentiate — involve nonlinear turning points, and PROFET cut trajectory prediction error by more than 50% against WOT and TrajectoryNet precisely where linear interpolation assumptions break (PROFET...). The dynamic arm is secondary to the core map, so we run it with PROFET and report WOT alongside rather than relying on either alone.

Decisive Checks

Three results would decide the proposal. First, SATURN label transfer between mouse and human tumor Tregs should substantially exceed transfer between human tumor Tregs and human autoimmune-lesion Tregs, establishing disease context as the dominant axis of divergence. Second, murine Batf knockout should reproduce the human paradoxical marker increase if and only if the BATF module ranks as cross-species conserved — the first direct test of whether the flagship intratumoral Treg regulator translates to the animal models in which it will inevitably be preclinically tested. Third, after ROSE12 treatment, modules ranked high on $S$ should be rewired only in the tumor, while shared-context modules shift in colon tissue; if rewiring tracks conservation rank, the index is doing real predictive work. No reported result found on Lacuna for cross-species regulatory network conservation in tissue-infiltrating Tregs of any kind — which is the gap this proposal occupies.