Mimetic-Aware Affinity Learning for Membrane Proteins

Most affinity measurements for membrane proteins — the receptors, transporters, and channels behind a large share of approved drugs — are not made on the protein as it sits in a cell. They are made on a protein ripped out of its lipid bilayer and restabilized in a surrogate: a detergent micelle (a small, curved ball of soap-like molecules), a bicelle, or a nanodisc (a belt of scaffold protein holding a small patch of real bilayer). These surrogates, collectively "mimetics," are not inert containers. Dodecylphosphocholine (DPC) micelles collapse the yeast mitochondrial ADP/ATP carrier so completely that its thermal melting transition vanishes and its inhibitor no longer binds, while cardiolipin acyl chains are forced into non-native contacts that mimic real binding signals Mitochondrial ADP/ATP Carrier in Dodecylphosphocholine Binds Cardiolipins with Non-native Affinity. Detergent display similarly reshapes which epitopes antibodies can even see: in a 407-antibody screen across 215 GPCRs in DDM micelles, selective binders clustered on long disordered loops (mean antigen length 54 amino acids) while structured transmembrane epitopes went unrecognized Multiplexed selectivity screening of anti-GPCR antibodies.

Meanwhile, the databases that feed every modern affinity predictor — BindingDB alone holds roughly 2.9 million measurements across 1.3 million compounds BindingDB in 2024: a FAIR knowledgebase of protein-small molecule binding data — record pH and temperature per entry but almost never record which mimetic the protein was sitting in when the number was measured. This is a specific instance of biological data fragmentation: the assay context that most determines whether a membrane-protein affinity number means anything is trapped in the methods sections of papers, not in the machine-readable record. The consequence is that current predictors pool DPC-collapse artifacts, DDM distortions, and genuine bilayer measurements into a single regression target and treat them all as truth. Boltz-2, the strongest structure-plus-affinity model to date, already conditions on experimental method (X-ray versus NMR) but treats membrane targets as isolated solutes and never stratifies affinity records by solubilizing environment Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction.

Core Claim

The mimetic environment is a systematic, recoverable confounder in membrane-protein affinity data — not random noise, and not a problem that more parameters will wash out. If we (a) recover the mimetic context from the fragmented literature for the membrane-protein slice of BindingDB and ChEMBL, and (b) make it an explicit conditioning variable in an affinity model, then three things should become measurable and falsifiable: the model should beat mimetic-blind models on held-out membrane targets; the difference between predictions under a micelle condition and a bilayer condition, which I'll write as $\Delta_m = f(c, m_{\text{micelle}}) - f(c, m_{\text{bilayer}})$ for a complex $c$ and mimetic label $m$, should be large and consistent for proteins known to be detergent-sensitive; and cross-mimetic transfer should show directional bias (models trained on micellar records should mispredict bilayer measurements in a coherent direction, not with symmetric scatter).

Research Program

The program has three linked parts, one per role each direction plays: fragmentation supplies the recovery problem, mimetic fidelity supplies the biophysical semantics, and affinity learning supplies the test bed.

First, build a mimetic-annotated membrane affinity corpus. BindingDB already groups experiments into "Entry" objects keyed by study conditions like pH and temperature The Binding Database: data management and interface design; we extend this with a mimetic field (detergent identity and concentration, bicelle, nanodisc scaffold, cell-based assay, or unknown). Recovery is a literature-mining problem of the kind already proven tractable: Nobata et al. extracted a full metabolome from free text Mining metabolites: extracting the yeast metabolome from the literature, and CoV-AbDab demonstrated curated literature-to-database pipelines for antibody binding data CoV-AbDab: the Coronavirus Antibody Database. Schema design follows the lesson of SEA CDM, whose Study–Experiment–Assay model harmonized 1,278 studies and over two million samples precisely because it made experimental context a first-class entity rather than an afterthought SEA CDM: Study-Experiment-Assay Common Data Model and Databases for Cross-Domain Data Integration and Analysis. Critically, automated extraction will be validated against a manually curated gold set — BindingDB's own curators document that automated ingestion leaves factor-of-1000 affinity errors and misidentified targets, and state plainly that no current AI replaces human curation BindingDB in 2024: a FAIR knowledgebase of protein-small molecule binding data. We will report coverage honestly: the fraction of membrane-target entries for which mimetic context is recoverable at all is itself a first result about how fragmented this metadata is.

Second, train a mimetic-conditioned affinity model. Architecturally this is a Boltz-2-style model with one added input: an embedding of the mimetic label, entering the conditioning pathway alongside the existing experimental-method conditioning Boltz-2. The companion idea, drawn from Gap 4 of the literature, is that the negative examples used in contrastive affinity training should include detergent-distorted protein states, not just mis-docked ligands. DecoyDB built 5.3 million decoy poses by perturbing ligand geometry DecoyDB: A Dataset for Graph Contrastive Learning in Protein-Ligand Binding Affinity Prediction; we instead construct negatives by perturbing the protein environment, contrasting detergent-determined deposited structures of the same target against bilayer-embedded conformational ensembles. For the bilayer reference we use GPCRmd's 1.32 million conformational frames of receptors in explicit lipid bilayers All-Atom GPCR-Ligand Simulation via Residual Isometric Latent Flow, and we can scale ensemble generation using GPCRLMD, which reproduces per-target flexibility fluctuations with Pearson correlation near 0.84 while sampling an all-atom ensemble in 2.0 seconds — a 900x speedup over BioEmu All-Atom GPCR-Ligand Simulation via Residual Isometric Latent Flow. This addresses a bottleneck that OneProt's authors hit directly: they could only find adequate MD trajectories for 5,005 unique sequences, forcing tiny training batches for that modality When Protein Dynamics Matter: Integrating Molecular Dynamics into Protein Foundation Models.

Third, validate against biophysical ground truth and cross-mimetic transfer. The yAAC3 case gives a sharp counterfactual check: conditioned on DPC, the model should predict the loss of the carboxyatractyloside interaction that is unambiguously present in DDM (melting transition at 47.8–50.2 °C apo, near 80 °C with inhibitor) and absent in DPC (no transition at all) Mitochondrial ADP/ATP Carrier in Dodecylphosphocholine Binds Cardiolipins with Non-native Affinity. The antibody screen supplies a second check: predicted epitope accessibility under micellar conditioning should reproduce the measured bias toward long disordered antigens (54 versus 37 amino acids) Multiplexed selectivity screening of anti-GPCR antibodies. And where structural orientation matters — ligands that enter GPCRs from the bilayer itself, as characterized in The pathway of ligand entry from the membrane bilayer to a lipid G protein-coupled receptor — the micelle-conditioned model should systematically miss such routes, exactly as docking models blind to the membrane do.

Proposed Experiments

The primary test bed is a mimetic-stratified split of the membrane-protein slice of BindingDB/ChEMBL: hold out entire target families and entire mimetic classes, evaluating with RMSE and Pearson correlation on pKd. The critical splits are cross-mimetic (train on micelle-annotated records, test on nanodisc or cell-based records, and vice versa), because a pooled model cannot fail these tests in an informative way — it has no notion of the split axis. No reported result found on Lacuna for membrane-protein-specific affinity prediction on any of the standard benchmarks, so we will also run PDBbind-2016-core-style evaluation on a membrane subset we construct ourselves, with DecoyDB's reported RMSE of 1.188 and ELGN's 1.285 on the full core set DecoyDB, Predicting Protein-Ligand Binding Affinity with Equivariant Line Graph Network reported as context only, since the subsets are not comparable. The biophysical validation set is small but decisive: yAAC3/carboxyatractyloside, the transporter conformational cases where micelles fail to support functional helical dynamics Substrate-induced conformational dynamics of the dopamine transporter, and orientation-ambiguous micelle-derived models such as the prolactin receptor transmembrane domain A combined computational and structural model of the full-length human prolactin receptor.

Baselines

The do-nothing reference is pooled affinity prediction exactly as practiced today: all membrane-target affinity records, micellar and native and unlabeled alike, mixed into one training set with no mimetic metadata. Concretely this is Boltz-2 run as published, since it already ingests BindingDB and ChEMBL affinities with unit standardization and conditions on experimental method but not mimetic Boltz-2. It is the strongest competitor because it is the only published system that couples 3D complex prediction, affinity regression over these exact databases, and conditioning machinery we can slot a mimetic embedding into; on general benchmarks it reports Pearson R = 0.66 on a four-target FEP+ subset and an enrichment factor of 18.4 at the 0.5% threshold on MF-PCBA, both cited from its paper. We will retrain it under our splits with identical data, compute, and tuning budget; beating it means lower RMSE (and higher Pearson) on the cross-mimetic held-out splits, which is where the pooled model's blindness should hurt most. The 3D-graph baseline is the DecoyDB/GIGN/ELGN family, with DecoyDB's reported RMSE of 1.188 on the PDBbind 2016 core set as its reference figure DecoyDB; it is the strongest static-structure competitor but, like all of them Predicting Protein-Ligand Binding Affinity with Equivariant Line Graph Network, it cannot represent an environment at all, so the comparison isolates what conditioning adds. The key ablation is our own architecture with the mimetic label replaced by a shuffled or "unknown" token: same parameters, same data, same budget. If that ablation matches the properly conditioned model on cross-mimetic splits, the proposal's central claim is dead, and we will say so. Finally, a naive reference point: a docking-free 2D model of the MGraphDTA type with detergent identity appended as an ordinary categorical feature, which tests whether the signal is trivially extractable without any structural machinery.

Why not just…?

Why not just let a big network average out detergent noise? The evidence says detergent effects are not noise but structural collapse: in DPC, yAAC3 loses its thermal transition and inhibitor binding entirely while producing NMR contacts that mimic real binding — a systematically wrong label, not a scatter around a true value Mitochondrial ADP/ATP Carrier in Dodecylphosphocholine Binds Cardiolipins with Non-native Affinity. The antibody screen shows the same systematic skew at dataset scale Multiplexed selectivity screening of anti-GPCR antibodies. A pooled model has no mechanism to separate a collapsed-pocket measurement from a native one. Our cross-mimetic splits are precisely the experiment that would show whether scaling alone absorbs the effect: if the pooled Boltz-2 baseline matches the conditioned model there, the noise argument wins and we will report it.

Why not just use fast 2D or docking-free models and sidestep 3D artifacts entirely? They avoid the artifact question but also the physics. On "both-new" splits (unseen protein and unseen ligand), MGraphDTA reaches only Pearson 0.22 on DAVIS and 0.45 on KIBA, while the structure-using FDA pipeline lifts this to 0.29 and 0.51 A Folding-Docking-Affinity framework for protein-ligand binding affinity prediction. And 2D models collapse under structural decoys, whereas pocket-geometric 3D models stay robust On Machine Learning Approaches for Protein-Ligand Binding Affinity Prediction. Still, this alternative is a real threat to the scale of our contribution, which is why the naive categorical-feature control above exists: if the cheap version captures most of the gain, the structural conditioning is not earning its cost, and we will find that out early.

Why not just recollect everything in nanodiscs and retrain? Ideal science, impossible history: the two decades of micellar measurements already in BindingDB cannot be remeasured, they anchor every benchmark, and native-context assays for many targets do not exist. The fragmentation lens is exactly the argument that the confounder must be recovered from records, not re-collected — though we note honestly that as nanodisc and cell-based data accumulates, its value as a clean test set rises, and our cross-mimetic splits are designed to exploit it either way.

Why not just use raw AlphaFold structures and skip membrane re-orientation? Membranome 3.0 showed raw AlphaFold 2 models of membrane proteins misplace domains relative to the membrane plane and retain low-confidence tails that obscure pockets, requiring their PPM 3.0 and D-linker re-orientation pipeline to reach under-2 Å agreement with experiment — and that work covered only single-pass proteins, excluding the multi-pass GPCRs and carriers where detergent distortion bites hardest Membranome 3.0: Database of single‐pass membrane proteins with AlphaFold models. We adopt their re-orientation as a preprocessing step for our conditioning structures rather than repeating its construction.

Decisive Checks

The proposal stands or falls on three numbers. First, the shuffled-mimetic ablation versus the conditioned model on cross-mimetic splits: the conditioned model must win by a margin beyond seed variance, or the confounder story is wrong. Second, the yAAC3 counterfactual: the model, given DPC as condition, must predict loss of carboxyatractyloside binding, matching the measured disappearance of the thermal transition Mitochondrial ADP/ATP Carrier in Dodecylphosphocholine Binds Cardiolipins with Non-native Affinity. Third, the coverage statistic of literature-derived mimetic labels: if fewer than, say, a third of membrane-target affinity records yield recoverable context, the practical payoff shrinks and the finding becomes a measurement of the fragmentation itself — still worth publishing, but a different paper. Beyond the headline prediction gains, $\Delta_m$ values across the target landscape are the lasting deliverable: the first proteome-scale estimate of how much of our membrane-protein pharmacology was measured in the wrong medium.