Even when cholesterol and blood pressure are driven to target with the best available drugs, heart attacks and strokes keep occurring — this is residual cardiovascular risk, and the literature increasingly pins it on inflammation inside the artery wall rather than on lipids alone (Asada; Kexin Wang et al.; Ching-Hui Sia et al., who showed that heart-attack patients with no standard risk factors die more often, not less). One druggable node in this residual pathway is the angiotensin II type 1 receptor (AT1R), a cell-surface receptor that signals through two separable internal channels: a G-protein channel that maintains normal vascular function, and a β-arrestin-2 channel whose activation drives degradation of the vessel wall. Yi Fu et al. showed that the body's own protein COMP selectively blocks the destructive arrestin channel while leaving G-protein signaling intact — and that its absence causes a 94% incidence of fatal aortic aneurysm in mice. Separately, Justine S. Paradis et al. proved for other receptors that the deciding factor in such pathway choice is not the shape of a single receptor but the geometry of two receptors paired into a dimer: one "closed" dimer arrangement physically bars the arrestin protein from docking, while G-protein docking proceeds unimpeded.
That geometric switch is, mathematically, a rotational symmetry-breaking event: a two-fold rotationally symmetric dimer ($C_2$) becomes asymmetric ($C_1$) the moment one effector docks to one protomer. No current computational tool can model it. Strictly symmetry-respecting neural models cannot emit asymmetric outputs from symmetric inputs — this is Curie's Principle, formalized for neural networks by Sékou-Oumar Kaba and Siamak Ravanbakhsh — and atomistic simulation is far too slow for the millisecond-scale dimer reorientations involved (Mahdi Hijazi et al.). We propose to close this gap: treat quaternary bias at AT1R as a symmetry-breaking transition, build a generative structure model that can break symmetry in a principled way, and validate the designs with live-cell signaling assays and a mouse model of vascular injury.
Core Claim
The claim is threefold. First, mechanistically: functional selectivity at dimerized GPCRs is governed by a $C_2 \to C_1$ rotational symmetry-breaking transition in the dimer interface, and AT1R — the receptor at the center of vascular-injury signaling — carries a homolog of the conserved transmembrane-helix-5 tryptophan hotspot ($W5.34/33$) that Justine S. Paradis et al. identified as the quaternary bias switch in CXCR4, μOR, and V2R. Second, methodologically: a discrete-diffusion protein structure generator, extended from ESMDiff to two-chain dimers and equipped with relaxed rotational equivariance, can sample the symmetry-broken effector-bound states that strictly equivariant models cannot. Third, therapeutically: mutations or stabilizing ligands that lock AT1R into the closed, arrestin-occluded dimer state should suppress β-arrestin-2-driven vascular injury while preserving G-protein signaling — a sharper instrument than the unbiased receptor blockade that Xi Zhang et al. describe as eliminating protective signaling along with the harmful kind.
Research Program
The program has four linked parts, moving from formalism to wet-lab validation.
Part 1 — an order parameter for quaternary symmetry breaking. We define a continuous coordinate for dimer asymmetry, for example the deviation of the dimer's transmembrane-helix-5 interface geometry from its $C_2$-symmetric reference, and verify on existing CXCR4, μOR, and V2R structures and MD trajectories that open and closed dimer states separate cleanly in this coordinate. This borrows the logic of two-step symmetry breaking from electronic nematicity in quantum materials, where directional order emerges first and only later locks into spatial order (S. Raghu et al.; Jonatan Wårdh et al., who showed a fluctuating ordered state can leave behind vestigial directional order). The analogy is deliberate, not decorative: the dimer's directional (rotational) order is set before any effector engages, and effector docking then selects one orientation from a degenerate pair.
Part 2 — a symmetry-relaxed dimer generator. We extend the discrete structure language model of ESMDiff (codebase lujiarui/esmdiff), which compresses 3D protein backbones into discrete tokens and samples conformational ensembles by masked diffusion 20–100× faster than continuous diffusion models, to two-chain GPCR dimers. The key new ingredient is relaxed rotational equivariance. Strict equivariance demands $\phi(g \cdot x) = g \cdot \phi(x)$; following Kaba and Ravanbakhsh we instead learn
$$\phi(g \cdot x) = g \cdot \phi(x) + \delta(g, x)$$
where $\delta$ is a learnable symmetry-breaking term, realized practically as the small learnable perturbation matrices $\Delta$ that R2Det introduced for relaxed rotation-equivariant convolutions in vision. The model is conditioned on effector identity (G-protein versus β-arrestin-2) and on ligand state (angiotensin II alone, or with the COMP EGF2 domain bound), and is trained on dimeric GPCR structures, ATLAS-style MD ensembles, and the apo/holo and fold-switching benchmarks that ESMDiff was validated on. Sanity checks are the BPTI millisecond benchmark (rare-state recovery) and the intrinsically disordered protein ensembles, both of which ESMDiff already handles.
Part 3 — prediction and design. The trained model is asked, prospectively and before any experiment: does AT1R's dimer landscape contain a closed, arrestin-occluded state reachable through the $W5.34$ hotspot? We then design a panel of 32 AT1R variants — matching the panel size of Hijazi et al. for comparability — including interface mutations predicted to lock the closed state, plus small peptidomimetics predicted to stabilize it, in the spirit of the COMP EGF2-derived rescue compound of Fu et al.. The design objective is a bias factor
$$B(m) = \log \frac{E_{\text{arr}}(m)\,/\,E_{G_q}(m)}{E_{\text{arr}}(\text{WT})\,/\,E_{G_q}(\text{WT})}$$
where $E$ is signaling efficacy measured by biosensor, $m$ is a designed variant, and WT is wild type; we search for variants that strongly decrease $B$ (arrestin-silencing) while keeping G-protein efficacy within roughly 80% of wild type.
Part 4 — tiered experimental validation. Designs are tested first in HEK293T cells with the bioluminescence resonance energy transfer (BRET) biosensor panel that Paradis et al. used for QUESTS: G-protein coupling and cAMP inhibition on one side, β-arrestin-2 recruitment and downstream ERK phosphorylation on the other. Variants that achieve the intended bias then enter the angiotensin-II-infused $ApoE^{-/-}$ and $COMP^{-/-}$ mouse aneurysm model of Fu et al., with endpoints of aneurysm incidence, elastin degradation scoring, and — critically — preservation of G-protein-dependent vascular function, head-to-head against the EGF2 peptidomimetic that study already validated. This closed computational-to-physical loop directly answers the surrogate-oracle failure mode that SMDD-Bench exposed, where the best agent solved only 40.2% of drug-design tasks and scored near 0% on 3D spatial reasoning against purely computational oracles.
Baselines
The naive reference point is wild-type AT1R under angiotensin II infusion with no intervention, anchored by the number Yi Fu et al. report: a 94% incidence of angiotensin-II-induced abdominal aortic aneurysm in COMP-deficient mice, rescued by AT1R, β-arrestin-2 knockout, or the EGF2 peptidomimetic. Any designed variant must materially beat this reference in the same model while preserving G-protein function. The EGF2 peptidomimetic itself, from the same study, is the strongest existing therapeutic comparator and will be run in parallel arms.
The strongest existing design method is the QUESTS workflow of Paradis et al., which designed CXCR4 dimer geometries with RosettaMembrane and validated them with BRET biosensors; we will run it unchanged on AT1R, designing the same number of variants under the same compute budget, testing them in the same blinded BRET panel. No reported result found on Lacuna for a QUESTS design success rate on new receptors, so the head-to-head must be generated by us. For monomer-level allosteric design, the strongest reported number anywhere in this space is AlloCraft's 81% experimental success rate across 32 designed dopamine-receptor variants; we run its MD-plus-graph-theory pipeline on AT1R as a monomer-only comparator, again with 32 variants. Beating these baselines means a higher proportion of tested variants achieving at least a ten-fold shift in the bias factor $B$ with retained G-protein efficacy, and, in mice, aneurysm incidence significantly below the 94% reference with intact G-protein-dependent responses.
Two ablations isolate the proposal's key ingredient. The first removes the relaxed term $\delta$ from our generator, restoring strict $C_2$ equivariance; by Curie's Principle as proved by Kaba and Ravanbakhsh, this model cannot generate asymmetric effector-bound dimers, and we expect it to collapse open and closed states onto a single symmetric mode. The second replaces principled relaxation with random noise injection into the symmetric model, which the same authors show degrades convergence and generalization. Finally, an unconstrained continuous diffusion model trained on the identical dimer dataset serves as a generative baseline; ESMDiff reports continuous diffusion competitors such as AlphaFlow run 20–100× slower, so we grant it equal wall-clock time rather than equal sampling count, which is the generous comparison.
Why not just…?
Why not just use the COMP EGF2 peptidomimetic that already works? This is the most serious objection, and we treat it as a real threat rather than arguing it away. Fu et al. already rescued aneurysms in mice with a peptidomimetic and showed COMP deficiency marks human aneurysm patients. But their mechanism acts on the receptor's extracellular N-terminus and leaves open whether quaternary geometry governs AT1R bias at all — the question Paradis et al. answered only for CXCR4, μOR, and V2R. The head-to-head mouse comparison in Part 4 settles the therapeutic question honestly: if the peptidomimetic matches the closed-dimer designs, the therapeutic increment is modest and the proposal's lasting contribution is the symmetry-aware design method and the mechanistic answer — a dimer-level control point that a peptidomimetic dosing strategy cannot reach, since it cannot choose which quaternary state it stabilizes without a model of the symmetry-broken landscape.
Why not just run RosettaMembrane design (QUESTS) on AT1R directly? Because RosettaMembrane optimizes a static interface geometry and does not model effector docking as a symmetry-breaking transition, it cannot rank which symmetry-broken state is thermodynamically accessible or how a flexible extracellular modulator like COMP shifts the population between states. Our generative ensemble explicitly predicts state populations conditioned on ligand and effector. But we concede QUESTS could plausibly succeed on the narrow task of finding interface mutations, which is exactly why it is our primary baseline rather than a footnote: the head-to-head, same receptor, same panel size, same blinded assays, is the decisive test of whether symmetry-relaxed generation adds anything over static design.
Why not just run longer molecular dynamics? Hijazi et al. state explicitly that standard atomistic MD cannot capture millisecond-scale quaternary reorientations without enhanced sampling or restraints — and even AlloCraft, built on MD, was confined to monomers and deferred oligomers and arrestin recruitment to future work. The 20–100× sampling speedup that ESMDiff reports over continuous diffusion is what makes ensemble-level dimer modeling tractable at all. Moreover, SMDD-Bench shows that staying inside computational oracles without physical testing produces near-zero success on 3D spatial tasks — which is why every design tier in this proposal terminates in a wet-lab readout rather than a simulated score.
Proposed Experiments and Decisive Checks
The experimental sequence is fixed in advance. First, the Curie test: the strictly equivariant ablation must fail to produce asymmetric effector-bound dimers while the relaxed model does — this validates the methodological core on existing CXCR4 data before any new receptor is touched. Second, the prospective hotspot test: the model must predict the AT1R $W5.34$-centered quaternary switch from sequence and structure alone, before mutagenesis. Third, the BRET panel: 32 designed variants versus 32 QUESTS-designed and 32 AlloCraft-designed comparators, assayed blind, success defined as a ten-fold or greater reduction in arrestin-biased signaling at preserved G-protein efficacy, benchmarked against AlloCraft's reported 81% success rate across 32 variants (Hijazi et al.) and against the qualitative open/closed-dimer BRET signatures of Paradis et al., for which no reported result found on Lacuna for a quantitative design success rate. Fourth, the in vivo test: winning variants in the angiotensin-II-infused $ApoE^{-/-}$/$COMP^{-/-}$ model against the 94% aneurysm incidence reference (Fu et al.) and against the EGF2 peptidomimetic, with elastin histology and G-protein-dependent function scored in parallel.
The broader payoff extends past AT1R. If dimer symmetry breaking proves to be a general control point for biased signaling, the same generator applies to the GLP-1 receptor, whose agonists confer anti-inflammatory, plaque-stabilizing vascular protection through multiple downstream pathways (Tianyu Wang et al.), and to chemokine receptors on plaque-infiltrating T cells, offering a locally acting alternative to the systemic checkpoint blockade that Fan Lin et al. achieved in oncology patients (median plaque reduction of $-2.72$ mm² with Fc-competent anti-PD-1) but which carries unacceptable autoimmune risk in ordinary cardiovascular patients. Symmetry breaking, in short, is not just a physical curiosity here — it is a designable dial for pathway selectivity in exactly the patient population where current drugs have run out of levers.
References
- Prevention of cardiovascular event targeted vascular inflammation, and pathology of thrombosis
- Remnant cholesterol and atherosclerotic cardiovascular disease: Metabolism, mechanism, evidence, and treatment
- Comparison of Mortality Outcomes in Acute Myocardial Infarction Patients With or Without Standard Modifiable Cardiovascular Risk Factors
- Cartilage oligomeric matrix protein is an endogenous β-arrestin-2-selective allosteric modulator of AT1 receptor counteracting vascular injury
- Computationally designed GPCR quaternary structures bias signaling pathway activation
- Symmetry Breaking and Equivariant Neural Networks
- Computational design of allosteric pathways reprograms ligand-selective GPCR signaling
- Structure Language Models for Protein Conformation Generation
- Advances in the allostery of angiotensin II type 1 receptor
- Microscopic theory of the nematic phase in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mtext>Sr</mml:mtext></mml:mrow><mml:mn>3</mml:mn></mml:msub><mml:msub><mml:mrow><mml:mtext>Ru</mml:mtext></mml:mrow><mml:mn>2</mml:mn></mml:msub><mml:msub><mml:mtext>O</mml:mtext><mml:mn>7</mml:mn></mml:msub></mml:mrow></mml:math>
- Nematic single-component superconductivity and loop-current order from pair-density wave instability
- R2Det: Exploring Relaxed Rotation Equivariance in 2D Object Detection
- SMDD-Bench: Can LLMs Solve Real-World Small Molecule Drug Design Tasks?
- Glucagon-like peptide-1 receptor agonists: new strategies and therapeutic targets to treat atherosclerotic cardiovascular disease
- Targeting pro-inflammatory T cells as a novel therapeutic approach to potentially resolve atherosclerosis in humans