Molecular mimicry is the phenomenon in which a pathogen protein resembles a human protein closely enough that antibodies or T cells primed against the pathogen also attack the host, producing autoimmunity. Machine-learning detection of mimicry today mostly reads amino-acid sequences: a protein language model trained to tell human from viral proteins flags the viral proteins it mistakenly calls human as candidate mimics. This works partly — but antibodies do not bind sequence, they bind three-dimensional surface patches, and those patches move. Proteins flex continuously, and transient or cryptic surface states that appear only rarely can still be recognized by antibodies. A static structure snapshot, whether from a crystal or from AlphaFold, never shows these states. The central problem is that current tools entangle what a surface patch is (its shape, which is largely invariant as the protein tumbles and wiggles) with how it moves (which states it visits and how often). This proposal argues that mimicry detection requires splitting those two things apart.
Concretely, we propose to build a two-tower representation over four-dimensional molecular surfaces — trajectories of conformational change rendered as sequences of surface point clouds. A Geometry Tower attaches local coordinate frames to moving surface patches (in the spirit of frame-based 4D point cloud networks) so each patch is described by a shape embedding that is invariant to the protein's global rotation and to its own motion. A Motion Tower encodes how those patches deform over time and estimates how often each conformational state occurs. The mimicry question then becomes precise: does a pathogen surface patch's invariant geometry match a host patch's geometry in a transient state, weighted by how reachable that state is? No existing method can ask this question, because none separates the ingredients.
Core Claim
The claim is that conformational molecular mimicry — cross-reactivity driven by 3D surface similarity that sequence alignment cannot see, occurring in host conformational states that static structure prediction cannot show — is detectable only when static patch geometry and conformational motion are represented in separate, explicitly decoupled streams. We formalize the mimicry score between a pathogen protein $P$ and a host protein $H$ as
$$ S(P,H) \;=\; \sum_{s} \pi_s \,\max_{i}\; \langle\, g_P,\; g_{H,s}^{(i)} \rangle, $$
where $g_P$ is the Geometry Tower embedding of a pathogen surface patch, $g_{H,s}^{(i)}$ is the embedding of host patch $i$ in conformational state $s$, and $\pi_s$ is the state occupancy estimated by the Motion Tower — the fraction of the trajectory the host spends in state $s$. The decoupling is the point: $\pi_s$ lets a rare transient state contribute in proportion to its physical reachability, and the frame-invariant $g$ lets shape be compared without the confound of motion. A monolithic encoder over the same 4D data, or an average over per-frame embeddings, destroys exactly this structure — a claim we will test by ablation rather than assert.
Research Program
Three components get built. First, the decoupled surface encoder. For each protein we obtain a conformational trajectory from molecular-dynamics corpora and generative conformation samplers of the kind used by ProTDyn: A Foundation Protein Language Model for Thermodynamics and Dynamics Generation, which was trained on the mdCath and BioEmu corpora spanning more than 5,000 proteins at multiple timescales (1ns, 10ns, 100ns). Each frame is converted to a solvent-accessible surface point cloud; a rotation-equivariant frame network in the style of LeaF: Learning Frames for 4D Point Cloud Sequence Understanding predicts a local SO(3) frame per patch, yielding the Geometry Tower's invariant patch embeddings, while the Motion Tower models patch deformation fields and infers a discrete set of conformational states with occupancies $\pi_s$. LeaF's frames assume quasi-rigid local motion, which molecular surfaces violate — sidechains vibrate at nanosecond scale while domain hinges bend at hundred-nanosecond scale — so we treat frames as smoothly re-estimated over trajectory windows and use velocity quantization of the kind introduced in Geometric Flow Grounding: A Unified Manifold Decoupling Framework for Dynamics Discovery and Verification to prevent averaging across branched motions, which GFG showed produces off-manifold drift (baseline MSE ~1.29 versus 0.0003 for decoupled projection on nonlinear physical systems).
Second, an on-manifold filter. GFG's tangent projection residual — the quantity that quantifies how badly a predicted state violates the physical data manifold — is repurposed as a quality gate on inferred transient states: transient host patches with high manifold violation are discarded as hallucinations before they enter the mimicry score. This adapts GFG from Euclidean systems data to molecular surfaces, which is itself an open technical risk given GFG's smooth-manifold assumptions; abrupt conformational flips are handled by segmenting trajectories into piecewise-smooth regimes.
Third, a reachability test for predicted mimics. For any predicted mimic pair we ask whether the pathogen surface can physically adopt the mimicking geometry, using decoupled latent steering in the style of COMPASS: Decoupled Latent Steering for Protein Conformational Transitions: steering updates the latent conditioning of a diffusion model at an intermediate Tweedie projection step rather than pushing atomic coordinates, which COMPASS showed keeps the steric clash score at 0.192 versus 0.515 for direct 3D coordinate guidance. This closes the loop: the decoupled representation predicts a mimic, and the decoupled steering verifies it is physically reachable without forcing clashes.
The proposal deliberately does not attempt to solve the bystander-activation versus true-mimicry ambiguity that pervades post-infectious autoimmunity, but it gives the debate a tool: cases such as anti-NF186 nodopathy after SARS-CoV-2 (Autoimmune nodopathy with anti-NF186 antibodies following SARS-CoV-2 infection: a case report) or anti-NMDAR encephalitis after herpes simplex encephalitis (Specific viral antibodies associate with anti-NMDAR encephalitis after herpes simplex encephalitis) can be scored for geometric mimicry in transient states; a high score is evidence for the mimicry route, a near-zero score pushes weight toward inflammatory bystander mechanisms.
Baselines
The naive reference point is chance-level ranking of pathogen–host pairs, which anchors the retrieval metrics at AUROC 0.5; any method must beat this trivially, so the real comparisons are the following. The strongest sequence competitor is a fine-tuned protein language model in the style of Protein language models expose viral mimicry and immune escape, which on SwissProt human/viral classification achieved an ROC AUC of 99.7% with ~97% accuracy — but with a 9.48% rate of viral proteins misclassified as human, precisely the asymmetric failure that defines candidate conformational mimics. We rerun this baseline ourselves on our pair-retrieval task with cosine similarity of ESM-2 embeddings, so the comparison is on identical data. The strongest static structure competitor is the transfer-learning epitope predictor of SEMA: Antigen B-cell conformational epitope prediction using deep transfer learning, which reported an ROC AUC of 0.76 on its retrospective post-2020 PDB test set; we pair SEMA-3D with rigid structural alignment (RMSD between static predicted structures) as the static-geometry mimicry score, given that Geometric epitope and paratope prediction requires a co-crystallized antibody–antigen complex as input and therefore cannot score unpaired pathogen–host pairs at all. The strongest dynamic-but-monolithic competitor is an encode-then-average pipeline: the same conformational trajectories, each frame passed through the same Geometry encoder, embeddings averaged over time — plus, as a tougher variant, the learned weighted per-conformation fusion of Learning Protein Representations with Conformational Dynamics, which outperformed static SaProt-650M by +11.1 accuracy points on the CoDNaS-Q multi-state subset, making it the best existing dynamic representation baseline rather than a strawman. Finally, the key ablation is our own architecture with the decoupling removed: one monolithic 4D encoder consuming surface trajectories directly, plus a variant with geometry matching but no occupancy weighting $\pi_s$. All methods receive identical trajectory inputs, identical splits with UniRef50-style homology clustering (following the SwissProt protocol of the viral-mimicry paper), and the same tuning budget and compute; a fair win means higher AUROC on cross-reactive pair retrieval stratified to low-sequence-identity pairs, and epitope localization above SEMA's 0.76 ROC AUC on the same retrospective split.
Why not just…?
Why not just use sequence language models, which already reach 99.7% ROC AUC? That headline number is for the easy task — telling human from viral proteins overall. The failure mode relevant here is the asymmetric 9.48% of viral proteins misclassified as human in Protein language models expose viral mimicry and immune escape, and those authors explicitly list the absence of structural information as their main limitation. More fundamentally, mimicry mediated by conformational epitopes — patches assembled from sequentially distant residues — is invisible to any sequence metric when linear homology is low. Our experiments settle this directly: the retrieval task is stratified by sequence identity, and the claim survives only if the decoupled method wins specifically in the low-identity stratum where sequence methods sit near chance.
Why not just compare static AlphaFold structures with RMSD or TM-align? This is a real threat and we treat it as the primary static baseline rather than dismissing it. The evidence against it is that static snapshots omit functional states: Learning Protein Representations with Conformational Dynamics showed static PLMs lose 11.1 accuracy points to dynamics-aware representations exactly on multi-state proteins, and the geometric epitope predictors of Geometric epitope and paratope prediction degrade on unbound versus bound structures because induced-fit changes are unrepresented. But if static alignment turns out to recover most documented cross-reactive pairs, our hypothesis about transient-state mimicry is wrong for those pairs, and we will report that honestly — the stratified evaluation exists to find out.
Why not just average embeddings over a molecular-dynamics trajectory? Because averaging collapses exactly the states we care about. Learning Protein Representations with Conformational Dynamics explicitly evaluated encode-then-average and found it inferior to their learned weighted fusion, and Geometric Flow Grounding: A Unified Manifold Decoupling Framework for Dynamics Discovery and Verification showed that averaging conflicting velocities causes smoothing drift and hallucinated trajectories (MSE ~1.29 versus 0.0003 when decoupled). Our ablation makes this an empirical result on our task rather than an appeal to their domains: if the averaged variant matches the decoupled score on mimicry retrieval, the core claim fails.
Why not just steer the pathogen structure directly toward the host epitope in 3D to test mimicry? Because forcing coordinates off the learned physical manifold produces garbage structures. COMPASS: Decoupled Latent Steering for Protein Conformational Transitions measured exactly this: direct 3D guidance inflates the steric clash score to 0.515, while decoupled latent steering keeps it at 0.192, statistically indistinguishable from the unsteered 0.191, while still achieving a 55.6% hit rate on large displacements (>15 Å) where unsteered models achieve 0%. We adopt the decoupled steering as our reachability test for precisely this reason — it is the alternative, done the way the evidence supports.
Proposed Experiments and Decisive Checks
Experiment one is cross-reactive pair retrieval. We curate positive pathogen–host pairs from IEDB immunogenicity records and the mechanistic literature — including the biofilm peptide/P4HB pair in lupus nephritis (A Novel Autoantibody Induced by Bacterial Biofilm Conserved Components Aggravates Lupus Nephritis) and the bacterial GAD/human GAD65 homology axis — with homology-matched negatives, and report AUROC and top-$k$ precision for every baseline, stratified by sequence identity. The decisive slice is pairs below 20% identity: the proposal wins only if the decoupled score reaches AUROC ≥ 0.75 there while sequence and static baselines sit near chance. Experiment two is epitope localization on the SEMA retrospective split (post-January 2020 PDB complexes with no training homologs): a prediction head on our geometry tower must clear SEMA's reported 0.76 ROC AUC under the same split and budget. Experiment three is reachability: we first reproduce decoupled steering on the curated 98-pair apo–holo benchmark of COMPASS, where COMPASS reported a +42.9% gain in strict 1.0 Å target-reaching efficiency over unsteered models, then apply it to predicted mimic pairs and require a steric clash score within the unsteered range (~0.19). Experiment four is prospective: rank untested pathogen–human pairs, correlate scores with IEDB immunogenicity, and nominate the top pairs with low sequence identity and high transient-state occupancy for wet-lab antibody cross-binding assays. The honest failure modes are known in advance: if GFG-style tangent projection breaks on discontinuous conformational flips, the piecewise-smooth segmentation must be shown to hold; and if labeled cross-reactivity data proves too sparse for stable training, the epitope-localization task (experiment two) becomes the primary quantitative anchor, since its benchmark is public and its baseline number is on record.
References
- ProTDyn: A Foundation Protein Language Model for Thermodynamics and Dynamics Generation
- LeaF: Learning Frames for 4D Point Cloud Sequence Understanding
- Geometric Flow Grounding: A Unified Manifold Decoupling Framework for Dynamics Discovery and Verification
- COMPASS: Decoupled Latent Steering for Protein Conformational Transitions
- Autoimmune nodopathy with anti-NF186 antibodies following SARS-CoV-2 infection: a case report
- Specific viral antibodies associate with anti-NMDAR encephalitis after herpes simplex encephalitis
- Protein language models expose viral mimicry and immune escape
- SEMA: Antigen B-cell conformational epitope prediction using deep transfer learning
- Geometric epitope and paratope prediction
- Learning Protein Representations with Conformational Dynamics
- A Novel Autoantibody Induced by Bacterial Biofilm Conserved Components Aggravates Lupus Nephritis