Attribution That Tracks the Weights: Sanity-Checked Explanations of Tethered GPCR Activation

When a deep model predicts something a biologist cares about — say, whether a receptor protein is in its active or inactive shape — we usually also ask it why: which residues mattered? But the field of attribution, the study of these "why" maps, has an uncomfortable finding at its core. Several popular explanation methods produce nearly the same heatmap whether the network's weights are trained or completely random, meaning the "explanation" describes the input's edges and priors rather than anything the model learned (Adebayo et al.). The standard diagnostic is cascading parameter randomization: scramble the network layer by layer, from the output backward, and watch whether the explanation collapses into noise. If it does not, the explanation is not tracking the learned parameters (Adebayo et al.). Meanwhile, in structural biology, a family of receptors called GPCRs has yielded an unusually clean mechanistic puzzle: some receptors carry their own built-in agonist. In the relaxin receptor RXFP1, a loop called ECL2 physically inserts into the receptor's own signaling pocket once an external lid domain is displaced (Erlandson et al.); in GPR161, that same loop sits constitutively in the pocket, stabilized by membrane sterols (Hoppe et al.). Mutagenesis has pinned the mechanism to specific residues — mutating Phe564 and Leu566 in RXFP1's ECL2 completely abolishes signaling.

These two threads meet at a gap nobody has closed. Machine learning models of GPCR activation exist — an XGBoost model over hand-engineered geometric features reaches 97.37% activation-state classification accuracy on GPCRdb structures (Mollaei and Barati Farimani), and a protein language model fine-tuned on GPCRs explains itself with raw attention weights (GPCR-BERT) — but neither has ever been subjected to parameter randomization testing, and recent evidence shows automated interpretability metrics can fail to distinguish trained protein transformers from randomly initialized ones (Heap et al.). Worse, all of these models were built for canonical ligand-driven receptors; none addresses tethered agonism, where the trigger is an intramolecular loop insertion rather than an external molecule.

Core Claim

The claim is that a geometric deep network — a model that reads the receptor's 3D atomic coordinates directly, with no hand-picked distance features — trained on GPCRdb structural ensembles can match the accuracy of engineered-feature models while producing residue-level attributions that (i) degrade properly under cascading weight randomization and (ii) localize the tethered-agonist module in RXFP1 and GPR161, recovering the experimentally implicated residues without being told where to look. The attribution method must be perturbation- or masking-based rather than gradient- or attention-based, because modified backpropagation provably collapses toward rank-1 maps that forget both the class objective and the upper-layer weights (Sixt et al.), and learned extremal masks probe the model as a function rather than its gradients (Fong et al.), in the spirit of counterfactual explainers already shown to be sensitive to internal weights (Barkan et al.).

The sanity check is formalized as a degradation profile. Let $a_\theta(x) \in \mathbb{R}^{n_{\text{res}}}$ be the attribution over residues for input structure $x$ under weights $\theta$, and let $\theta_{\geq \ell}$ denote weights with all layers from the output down to depth $\ell$ re-randomized. Define

$$d(\ell) = \operatorname{sim}\big(a_\theta(x),\, a_{\theta_{\geq \ell}}(x)\big),$$

where $\operatorname{sim}$ is a rank correlation between residue-importance rankings. A parameter-sensitive explanation has $d(\ell)$ falling monotonically from 1 toward the trained-vs-fully-random reference level as $\ell$ deepens; a fake explanation stays flat. This ranking-based variant sidesteps known failures of pixel-similarity sanity metrics, which can report false results depending on layer initialization distributions and structural input correlations (Hedström et al.; Lin et al.; Tan).

Research Program

Three components. First, the model: an SE(3)-respecting graph network over residue nodes (coordinates, residue identity, secondary structure) and edge features (contacts, distances), trained on the GPCRdb ensemble of 555 experimental structures across 105 receptor types with active/intermediate/inactive labels, the same corpus used by the XGBoost baseline (Mollaei and Barati Farimani), plus a continuous activity-level head evaluated in the same 0–100% scale. Second, the explanation layer: learned residue masks in the extremal-perturbation family (Fong et al.), trained to minimally mask residues that flip the predicted state, with the counterfactual objective of the LXR line (Barkan et al.) as the sequence-space analogue. Third, the audit: cascading randomization profiles for our attributions, for gradient baselines, and — as an external subject — for the raw-attention explanations used in GPCR-BERT (GPCR-BERT), directly testing whether its attention maps survive the demolition that theory says they should not (Heap et al.; Sixt et al.).

For the tethered-agonism regime, the model is applied to the RXFP1 active-state structure at 3.2 Å (Erlandson et al.) and the GPR161–$G_s$ complex at 2.7 Å (Hoppe et al.), with short MD ensembles seeded from these endpoints to populate the inactive and intermediate basins. We deliberately do not attempt to simulate the full millisecond transition across the activation barrier — that remains computationally prohibitive (Hijazi et al.) — instead, the differentiable surrogate interpolates between endpoints the way the XGBoost model did for β2AR, and its attributions tell us which residues carry the transition. This is the point of contact with the bioinformatics gap: allosteric pathway tools such as AlloCraft map conduits with mutual information and KL-divergence statistics over long trajectories (Hijazi et al.); a parameterized surrogate offers attributions without the sampling bottleneck, and the randomization audit is what makes those attributions trustworthy rather than decorative.

Proposed Experiments

Experiment 1 is predictive parity on the shared benchmark. Train on the GPCRdb structural split with receptor-type-disjoint folds so the model cannot memorize receptor families; report activation-state accuracy and activity-level mean absolute error against the XGBoost figures of 97.37% accuracy and 8.55% MAE reported by Mollaei and Barati Farimani. Matching within a point or two, from raw coordinates rather than 58 engineered features, is the entry condition; the attribution results are the actual contribution.

Experiment 2 is the sanity audit. Compute $d(\ell)$ profiles for learned masks, gradient explanations, and raw attention, on our model and on a GPCR-BERT-style protein language model fine-tuned on its 254 Class A GPCR sequences with the motif subsets it reports (98.05% NPxxY, 100% E/DRY, 86.29% CWxP masked-token accuracy; GPCR-BERT). Cross-check method agreement and disagreement with the tooling of missingness and unity-in-xai-court, and use conex to generate counterfactual residue sets as a second, independent explanation. The prediction, grounded in Sixt et al. and Heap et al., is that attention and modified-gradient maps stay flat in $d(\ell)$ while learned masks degrade.

Experiment 3 is the tethered-agonism localization test, the decisive one. For RXFP1, the attribution for the active-vs-inactive decision must enrich ECL2 and specifically rank Phe564 and Leu566 — the residues whose mutation completely abolished signaling in the functional assays of Erlandson et al. (no reported numerical value found on Lacuna for the signaling abolition; the finding is qualitative) — above the great majority of the ~600-residue receptor, and must implicate the LRR lid and hinge that autoinhibit ECL2 in the resting state. For GPR161, the attribution must localize ECL2 as the built-in agonist and the conserved sterol-binding transmembrane pocket identified by Hoppe et al.. The metric is rank percentile and motif-level enrichment of the implicated residues, with significance against a null of uniformly distributed importance.

Experiment 4 stress-tests the surrogate on dynamics: apply the model to the two milliseconds, roughly 150,000 frames, of β2AR MD trajectories from Mollaei and Barati Farimani, and check that the attribution reproduces their nonlinear finding — the hair-trigger sensitivity of activity to small Helix-3–Helix-6 distance changes, including the intermediate oscillations their linear pairwise metrics miss. Optionally, on the mutational-effect side, evaluate against ProteinGym's supervised benchmark, where the best supervised model, ProteinNPT, reports a Spearman rank correlation of 0.613 (Notin et al.), to check that the geometric model is not regressing on general protein fitness relative to strong existing predictors.

Baselines

The "do nothing" reference point is the untrained network of identical architecture: compute learned-mask attributions on the same inputs with random weights, and measure how often they still enrich the ground-truth residues. If trained-model attributions do not beat this reference on the Experiment 3 localization metric, the explanation is an architectural artifact, which is precisely the failure Heap et al. documented for automated interpretability metrics on protein transformers. Passing this reference is the minimum bar.

The strongest existing competitor is the XGBoost pipeline of Mollaei and Barati Farimani, which reports 97.37% activation-state accuracy and 8.55% activity MAE on the 555-structure GPCRdb corpus. We will rerun it, or its faithful reimplementation, on our exact receptor-disjoint splits, with the same compute budget and a tuning budget at least as generous as ours, including its 55 contact-distance and 3 angular features and its native gain-based feature importance as the competing explanation. Beating it means two things jointly: accuracy within about one point on the shared split, and attributions that localize the tethered-agonist residues where its gain importance — which can only speak in the vocabulary of its pre-defined pairwise distances — cannot, because the LRR-lid/ECL2 interaction is not one of its 58 features. If it wins on accuracy by a wide margin, we report that honestly and reframe the contribution as the audit and localization results, which remain unclaimed.

The key ablation isolates the attribution method from the architecture: on the same trained network, compute raw-attention explanations in the GPCR-BERT style, vanilla gradients, and learned masks, and run all three through the cascading randomization audit and the Experiment 3 localization test. This separates the question "does the geometric architecture learn the right things" from the question "does the explanation report what the architecture learned" — the proposal's central claim lives in the second question.

Why not just…?

Why not just analyze the MD trajectories with classical statistics? This is the AlloCraft approach: mutual information and KL-divergence graphs over long simulations, validated at an 81% success rate across 32 designed dopamine receptor variants (Hijazi et al.). It is a real competitor for pathway discovery, and where millisecond trajectories exist it needs no neural network. But it needs the trajectories: the same authors note that MD is currently limited for millisecond-scale conformational transitions, and no such trajectory exists for the RXFP1 autoinhibition release. A differentiable surrogate interpolates between cryo-EM endpoints that statistics cannot bridge. On receptors where both apply, Experiment 4's comparison of our residue rankings with their conduit residues on the β2AR corpus is the head-to-head test; if the MI graphs dominate wherever trajectories exist, the surrogate's value is confined to the tethered, unsampled regime — which is exactly the regime this proposal targets.

Why not just read attention weights off a pre-trained protein language model? GPCR-BERT did this and found attention residues that correlate with mutagenesis data (GPCR-BERT). The problem is that correlation with mutagenesis does not establish the attention reflects the model's learned parameters: Adebayo et al. showed visually plausible maps can be weight-invariant, Sixt et al. showed why modified backward passes converge to input-driven rank-1 maps, and Heap et al. showed automated interpretability metrics fail to separate trained from random protein transformers. Experiment 2 settles this directly rather than by citation: we subject GPCR-BERT-style attention to cascading randomization and report the profile. If its attention degrades properly, the field learns something important and our masking approach loses part of its motivation; if it stays flat, the "long-range residue dependencies" it reports are architectural priors.

Why not just use tree-based feature importance, as the 97.37% XGBoost model already does? For canonical Class A receptors this is a serious threat, and we run it as our strongest baseline precisely because it might win on accuracy. Where it structurally cannot follow us is tethered agonism: its explanation vocabulary is fixed at 58 pre-engineered distances and angles chosen because the polar network and NPxxY motifs were already known (Mollaei and Barati Farimani). The RXFP1 trigger is a multi-domain steric event — an LRR lid unblocking a loop with a $C\Phi P\Phi$ motif that then inserts into the seven-helix bundle (Erlandson et al.) — and no one knows in advance which pairwise distances to engineer for it. That said, a fair test requires giving the baseline a chance: we will also run an XGBoost variant with an expanded, unbiased feature set (all loop–bundle contacts) so the comparison is about the learning regime, not about our stinginess with features.

Why not just use evolutionary couplings to find the conserved loop motif? Erlandson et al. themselves used coupling analysis to help identify the $C\Phi P\Phi$ motif, so this works for discovering which residues are conserved. But conservation is not state-dependence: evolutionary couplings cannot say that ECL2 inserts only when the LRR lid is displaced, or that GPR161's ECL2 is constitutively inserted and sterol-stabilized (Hoppe et al.). Large-scale evidence also caps expectations for alignment-only methods on context-dependent effects: in ProteinGym, indel and context-heavy settings favor supervised models, with ProteinNPT reaching a 0.613 Spearman correlation against weaker alignment-only baselines (Notin et al.). Couplings are a complementary prior we will feed the model, not a substitute for it.

Decisive Checks

The proposal stands or falls on three pre-registered outcomes: the learned-mask attribution degrades under cascading randomization while attention and gradient baselines do not (Experiment 2); Phe564, Leu566, the LRR hinge, and the GPR161 sterol pocket rank in the top decile of residue attributions against the random-weight and uniform-importance nulls (Experiment 3); and accuracy stays within roughly a point of the 97.37% / 8.55% MAE XGBoost benchmark on shared splits (Experiment 1). If the first holds but the third fails, the audit methodology still stands as a contribution — a demonstrated, rank-based parameter-sensitivity test for 3D structural models, a domain where, to our knowledge, no randomization sanity check has ever been run, despite the field's reliance on attention-based stories about allostery.