Proximity-Measured FOXP3 Complexes as a Physical Bridge Across the Tolerance Translation Gap

Therapies that induce immune tolerance — teaching the immune system to stand down against a graft or its own tissue, instead of globally suppressing it with drugs — routinely work in a culture dish and then fail in a living animal or patient. One reason is that the field measures tolerance at the wrong level of description. The unit that actually enforces tolerance is not the number of regulatory T cells (the immune system's built-in brake cells) and not the mRNA for Foxp3, the transcription factor that defines them, but the physical protein assemblies inside those cells that hold the Foxp3 program in place under inflammatory stress. Three clinical observations in the compiled literature say counts and transcripts are the wrong currency: in type 1 diabetes, a regulatory CD8⁺ T cell subset is numerically indistinguishable from healthy donors yet fails to suppress (Analysis of frequency changes of CD8+ regulatory T cell subsets in patients with type 1 diabetes); in Graves' disease, regulatory B cell counts normalize after remission while their suppressive function stays broken (Bingbing Zha et al.); and in lupus, interferon-alpha converts expanded regulatory T cells into inflammatory effectors through a timing-dependent STAT4 switch (TIGIT reverses IFN-α-promoted Th1-like Tregs via in-sequence effects dependent on STAT4).

The proposal here is to measure tolerance-relevant protein complexes where they live — inside intact regulatory T cells, before any cell is broken open — and to use that physical "complex state" as the connective tissue between three literatures that have never touched. Dynamical models already rank upstream regulators of FOXP3 from time-series transcriptomics, but never check whether those regulators physically touch Foxp3 protein (Causal dynamical modelling predicts novel regulatory genes of FOXP3 in human regulatory T cells). Co-immunoprecipitation — pulling a "bait" protein out of a cell lysate with an antibody to see which "prey" proteins come with it — is the gold standard for physical interactions, but lysis destroys the cell's compartmentalization and washes away weak, transient complexes. Two studies show the way out: in situ proximity ligation assays (PLA), which light up as fluorescent dots when two proteins sit within 40 nanometers inside an intact cell, were essential to confirm the Tax–Scribble interaction in T cells recovered from humanized mice (Ian T. Foster et al.), and a multi-assay pipeline of PLA plus co-IP plus fragment mapping successfully guided a peptide that disrupts the SOX2/CDP complex from the dish into xenografts (Targeting the SOX2/CDP protein complex with a peptide suppresses the malignant progression of esophageal squamous cell carcinoma). Finally, machine-learning work on the in vitro-to-in vivo gap — transferring drug response from cell lines to patients (DeepSADR) or predicting whole-cell network rewiring (Cell-Level Virtual Screening) — operates entirely at the transcript level and never asks whether the physical complexes it implicitly depends on actually assemble in a living animal. Nobody has connected transcript-level regulation, physical complex assembly, and cross-environment survival in one pipeline, on the cell type where it matters most therapeutically.

Core Claim

The claim is that the in vitro-to-in vivo translation gap in immune tolerance can be localized to specific, nameable protein complexes, and that a proximity-based complex-state measurement will predict durable tolerance better than cell counts or transcript levels. Concretely: for candidate FOXP3-stabilizing complexes identified by dynamical transcriptomic ranking and cross-linked co-IP–mass spectrometry, the PLA-measured assembly state in intact primary human regulatory T cells — measured under the inflammatory timing conditions that drive lineage collapse — will predict Foxp3 protein retention and suppressive function after in vivo transfer, while cell frequency and FOXP3 mRNA will not. To make "translation gap" quantitative rather than rhetorical, I define a per-complex translation-gap index: if $s_{\text{vitro}}(c)$ and $s_{\text{vivo}}(c)$ are the mean PLA puncta per cell for complex $c$ in matched in vitro and in vivo conditions, then

$$G(c) = 1 - \rho\big(s_{\text{vitro}}(c),\, s_{\text{vivo}}(c)\big),$$

where $\rho$ is the rank correlation across matched donor-and-condition pairs. $G(c) \approx 0$ means the complex state transfers across environments; $G(c) \approx 1$ means the dish lies. Interventions that lower $G$ for a Foxp3-stabilizing complex should transfer to in vivo tolerance; interventions that only look good in vitro should not.

Research Program

The program has four linked stages, each reusing a technique the literature has already validated in isolation. First, candidate generation: re-fit the delayed linear time-invariant dynamical models of Causal dynamical modelling predicts novel regulatory genes of FOXP3 in human regulatory T cells on the existing 19-time-point donor series (GEO: GSE11292), but extend the experiment with the inflammatory timing arms from the TIGIT study — TCR activation followed by interferon-alpha, versus interferon-alpha first — so the ranking surfaces regulators whose influence on FOXP3 is conditional on the in-sequence stress that actually causes lineage collapse in lupus. The original pipeline ranked 176 transcripts (161 genes) and validated 3 of 5 candidates by siRNA knockdown at the mRNA level; the open question it explicitly left is whether any of those regulators (NRBF2, NCOA7, MAP1LC3B) physically bind Foxp3 or its cofactors. Only NRBF2 knockdown moved Foxp3 protein, which makes NRBF2–Foxp3 the flagship candidate complex, with the others as controls.

Second, physical identification under lysis-minimizing conditions: chemical cross-linking applied to intact cells before lysis (freezing weak and transient assemblies in place), followed by co-IP coupled to shotgun mass spectrometry, to enumerate what actually co-assembles with Foxp3 under each timing arm. Third, quantitative in situ measurement: PLA probes for the validated bait–prey pairs, imaged in intact cells, yielding puncta-per-cell distributions per condition. This pairing is exactly the escalation that Ian T. Foster et al. and the SOX2/CDP study found necessary, because cross-linked co-IP tells you the complex exists and PLA tells you where, in which cells, and at what abundance — and the abundance distribution, not the yes/no answer, is what I claim carries predictive signal. Fourth, interface perturbation: once a complex's structure or interface is modeled, use the mutational-effect predictors already benchmarked on SKEMPI v2 — H3-DDG as the strongest supervised option, with BA-DDG and DiffAffinity as cross-checks — to rank point mutants and short constrained peptides predicted to stabilize or disrupt the interface, then test the top stabilizing probes with the same PLA-plus-functional-assay pipeline. The SOX2/CDP study showed this in silico-to-peptide route is feasible for intracellular transcription-factor complexes; what it did not have, because its hosts were immunodeficient, is an immune system to answer back.

Baselines

The naive reference point is regulatory cell frequency, the readout on which most tolerance trials are still stratified. The literature already documents its failure qualitatively — identical CD8⁺ regulatory counts in type 1 diabetes patients versus donors with loss of suppressive function (Analysis of frequency changes...), normalized regulatory B cell counts in remitted Graves' disease with defective function (Bingbing Zha et al.) — but no reported result found on Lacuna for a quantitative head-to-head of cell counts versus any functional or molecular predictor of tolerance. That absence is itself part of the gap this proposal fills: the first deliverable is precisely that head-to-head, so the naive baseline will be run on the same donor cells, same time points, and same inflammatory conditions as every other predictor, with no tuning, because there is nothing to tune.

The strongest existing methodological competitor is the dynamical-model ranking pipeline of Sawlekar et al., which is the only prior work that goes from data to specific FOXP3 regulator candidates with wet-lab validation. Its reported result is a 3-of-5 siRNA hit rate on FOXP3 mRNA (MAP1LC3B, NCOA7, NRBF2), with only NRBF2 affecting Foxp3 protein. For a fair comparison I will run that pipeline unchanged on the same GSE11292 data, then give it and my extended pipeline the same siRNA budget on the same donor cells and score both on protein-level and functional outcomes, not mRNA — its documented weak point. My method beats it if the PLA complex-state features predict which knockdowns lose Foxp3 protein and suppressive function with substantially higher accuracy than mRNA-level ranking; the ablation that isolates my key ingredient is to keep everything else fixed and swap in cross-linked co-IP band intensity for PLA puncta, which separates "physical interaction exists" from "physical assembly state in intact cells carries the signal."

The strongest functional assay system is the in-sequence interferon-alpha / TIGIT-agonist setup of Shihan Yu et al., with its readouts of IL-10 restoration and suppression of autologous B-cell differentiation into antibody-secreting plasmablasts; I adopt it as the functional ground truth rather than as a competitor, because it supplies the stressor and the endpoint but no physical-complex layer of its own. For the interface-perturbation stage, the computational baseline is H3-DDG at its reported Pearson correlation of 0.75 on SKEMPI v2 (and its reported 5.4% performance drop when using AlphaFold3-predicted rather than crystal structures, which is the regime I will actually be in); any peptide design I test must beat interface-blind peptide selection from the same library at holding complex state under stress.

Why not just…?

Why not just do standard co-IP and Western blot? Because lysis manufactures both false positives and false negatives exactly in the regime this proposal cares about: it mixes proteins from compartments that never meet in a living cell, and washes away the weak, phosphorylation-dependent assemblies that the TIGIT/STAT4 and NF-κB work implicates in regulatory T cell stability (Shihan Yu et al.; Hypothalamic kisspeptin alleviates myasthenia gravis...). The precedent is direct: Ian T. Foster et al. needed in situ PLA in cells recovered from humanized mice to see Tax–Scribble sequestration, and the SOX2/CDP study needed PLA plus fragment mapping to localize its interface. My design keeps co-IP but moves it after cross-linking and pairs it with in situ quantification, and the ablation above — co-IP band intensity versus PLA puncta on identical samples — is the experiment that settles whether the extra layer earns its cost.

Why not just profile gene expression and run gene set enrichment? Three results in the file say this falls short here. Only one of the three transcript-validated FOXP3 regulators moved Foxp3 protein, so mRNA ranking systematically misses post-transcriptional control (Sawlekar et al.). Curated enrichment libraries are fragile to context: across 727 cell lines and 138 drugs, Bateman et al. found MSigDB collection choice changes conclusions drastically (context-derived HGSK gave the top hit in only 40–60% of cases, and the ostensibly well-matched C6 oncogenic signatures performed worst, with Wilcoxon p-values as extreme as $1.24 \times 10^{-33}$ between collections). And static expression carries less therapeutic signal than network structure: CellVS-Net beat expression baselines on drug–disease retrieval (Hits@1 of 0.1250 versus 0.0357 for molecular fingerprints) — yet even that model, the file notes, has no physical protein-interaction ground truth. My proposal is the layer underneath all of these: it measures whether the implicated protein machinery actually assembles.

Why not just run the in vitro suppression assay and then a mouse efficacy study? This is the most serious alternative, and I want to be honest about that. The objection is that the whole apparatus is unnecessary: test candidate stabilizers in the Yu-style co-culture, take the winners into mice, and read out disease scores. Two things make that insufficient. The timing result shows static cultures mischaracterize the biology: interferon-alpha collapses regulatory T cells only when it arrives after TCR activation, a sequence a standard assay never controls (Shihan Yu et al.). And the translation literature shows in vitro efficacy is a poor predictor of in vivo outcome for exactly this kind of intervention — microRNAs that kill isolated cancer cells fail in animals due to systemic counter-mechanisms (Jian Wang et al.), and naive cross-domain aggregation collapses performance (removing DeepSADR's adaptive readout dropped clinical AUC from 0.856 to 0.662; DeepSADR). The efficacy-only route would tell me a peptide works or fails but not which complex failed to survive the environment, which is the mechanistic quantity the field lacks. My decisive experiment subsumes the alternative: if in vitro suppression assay scores predict in vivo outcomes as well as complex-state scores do, the proposal's central claim is wrong and I will say so.

Why not just predict everything in silico with AlphaFold-style models? Because unsupervised structural prediction of mutational effects is currently weak — BA-Cycle reached only a 0.3201 Spearman correlation on SKEMPI v2 (Boltzmann-Aligned Inverse Folding Model...) — and even the best supervised models say nothing about whether a complex assembles in a specific cell under inflammatory stress, which depends on expression, competition from endogenous binders, and post-translational state. In silico scoring enters my program where it is strong: ranking interface perturbations once a complex is physically validated, cross-checked against the reported H3-DDG performance ceiling.

Proposed Experiments and Decisive Checks

Experiment one ranks candidate regulators of FOXP3 under in-sequence inflammatory stress using the extended time-series design, scored by enrichment of physically validated complexes among the top-ranked genes relative to the Sawlekar baseline on GSE11292. Experiment two establishes the physical layer: cross-linked co-IP–mass spectrometry and PLA puncta quantification for the flagship NRBF2–Foxp3 complex and controls in primary human regulatory T cells across the three timing arms, with the Yu co-culture (IL-10 recovery, plasmablast suppression) as the functional endpoint; the decisive statistic is the area under a receiver-operating curve predicting functional collapse from complex state versus from cell frequency and FOXP3 mRNA on held-out donors. Experiment three tests ΔΔG-guided stabilizing and destabilizing probes against interface-blind controls from the same library, first in vitro (PLA plus function), using H3-DDG-ranked candidates and its reported SKEMPI v2 Pearson correlation of 0.75 as the reference for what the predictor itself can do. Experiment four closes the loop in vivo: adoptive transfer of labeled human regulatory T cells into humanized mice on the scale of the Foster study (32 humanized mice per comparison), recovery of cells from tissue, in situ PLA in the recovered cells, computation of the translation-gap index $G(c)$ for matched in vitro/in vivo pairs, and an inflammatory challenge to test whether low-$G$ complexes retain Foxp3 protein and function in vivo. The humanized setting is deliberate, since murine immune signaling diverges from human in cytokine receptor expression and Fc-receptor affinities (Chi Ngai Chan et al.), and unlike the SOX2/CDP xenografts it preserves a responsive immune compartment. The proposal succeeds if complex state beats counts and transcripts as a predictor of in vivo stability, and if at least one designed interface stabilizer demonstrably lowers $G$ for a Foxp3-anchoring complex while preserving suppressive function. It fails, informatively, if the physical assemblies turn out to be as environment-labile and uncoupled from function as the transcripts are — in which case the measured $G$ values and their per-complex variation would still constitute the first quantitative map of where, at the level of physical protein machinery, tolerance translation actually breaks.