What a receptor-state model learns about a drug

A drug can bind a receptor without producing the response a researcher wants. For a predictive model, that distinction changes the task: matching an experimental activity number is useful, but explaining whether the molecule activates a receptor requires knowing what the number represents. A recent GPCR model puts an active–inactive distinction inside its architecture and obtains its clearest gains for records labelled as agonists.¹

Dual-State Query, or DSQ, comes from Shuo Zhang and colleagues at the University of Birmingham. This assessment examines the September arXiv manuscript; its repository reports conference acceptance, which we have not independently verified. The interesting question is how far its receptor-state construction improves prediction—and when the resulting output can reasonably be interpreted as pharmacology.¹

The answer is encouraging and specific. DSQ improves several benchmark measures, and comparisons with simpler versions support the value of its state-aware design. Yet its activation score remains a learned quantity tied to the training labels. Understanding those labels is essential before treating the score as a description of what receptors in a cell are doing.

Put the assay before the architecture

The study draws recorded human GPCR interactions from GLASS and GPCRdb. Its curation assigns EC50 measurements to the agonist category and Ki or IC50 measurements to the antagonist category. Values become a common negative-logarithmic target: for measurements expressed in nanomolar units, the transformation is nine minus the base-ten logarithm of the numerical value.¹

This common scale helps the regression, but it does not erase the difference between measurements. An EC50 summarizes a concentration–response relationship; a binding affinity and a functional response are different observables. The primary experimental literature makes that distinction tangible. In work on the human M1 muscarinic receptor, Canals and colleagues separately measured binding and several signaling responses. The same allosteric compound, BQCA, produced little direct response in one assay and full agonism in more efficiently coupled pathways.²

That example is relevant to how we read DSQ, rather than an objection to learning from historical records. A model trained on the stated endpoint-based categories is first demonstrating its ability to reproduce that categorization and the associated activity values. A positive class score does not independently establish maximal response, pathway preference or the fraction of receptors occupying a physical conformation. Those interpretations would require observations matching those quantities.

DSQ inherits this broad data strategy from AiGPro, the earlier GPCR profiling model by Brahma and colleagues. AiGPro likewise grouped IC50 and Ki records separately from EC50 records. Its architecture supplies an agonist or antagonist class token as part of the prediction context; the paper explicitly describes obtaining different predicted activities for the same ligand when the requested activity class changes.³

DSQ makes a distinct modeling choice. It uses class labels while learning, but does not require the user to supply that class token when making a prediction. Instead, it estimates a class-associated activation score alongside bioactivity. That is a useful change in the question the software can answer. It does not make previously published AiGPro performance numbers a fair head-to-head comparison: the inputs and evaluation protocols differ.¹³

Two views of one receptor

DSQ begins with representations from pretrained protein and chemical language models. The protein features come from ESM-2 and the chemical features from MolFormer; these encoders remain frozen during training. Two sets of learned queries extract different views of each receptor sequence. The architecture encourages these views to remain distinct, then combines each with the ligand representation to estimate two state-associated affinity values.¹

The views are numerical representations, not solved active and inactive structures. No crystal structure is required as an input. Calling one branch active gives it a role in the learning objective; it does not by itself demonstrate that the branch reconstructs an experimentally observed conformation. This distinction lets us appreciate the practical attraction—working without a structure—without crediting a structural determination that was never performed.

The two branches meet in a gate inspired by the Monod–Wyman–Changeux account of conformational selection. The model compares its state-associated affinities, subtracts a learned receptor-dependent offset, and passes the result through a sigmoid function. That gate then weights the two affinity estimates to produce the final bioactivity prediction. Training also rewards alignment with the binary class labels and the expected direction of the difference between the state estimates.¹

The construction gives the network a useful constraint: its classification and regression outputs cannot vary entirely independently. An agonist-associated prediction is encouraged to arise through a preference for the active branch. This is a reasoned biological prior within a statistical model, rather than a new measurement of thermodynamic constants.

The experimental roots deserve attention. Canals and colleagues examined the two-state idea using binding, signaling and receptor mutations in cellular systems. BQCA became more active at an activating M1 receptor mutant. Their work demonstrates how a conformational model can organize experimentally distinguishable behaviors.² DSQ borrows that organizing logic. It asks whether the logic improves learned predictions across many receptors, a different and worthwhile test.¹

Where the improvement appears

The training set contains 260,816 records across 222 receptors, with 200,258 antagonist-labelled and 60,558 agonist-labelled observations. A separate validation set supports model development. The test set contains 11,464 interaction pairs across 203 receptors; those test interactions were excluded before training.¹

The strongest comparison is the one the authors actually control. DeepDTAGen, DrugBAN and GraphDTA were retrained using the same interaction split and activity targets. Under those conditions, DSQ reduces overall root-mean-square error from the strongest baseline's 1.071 to 0.999 on the logarithmic activity scale. Its overall R-squared is 0.517, compared with DrugBAN's 0.446.¹

These results support improved prediction of the recorded target under this benchmark. The logarithmic error deserves careful reading: one unit corresponds to a tenfold difference in the underlying concentration value. An RMSE close to one is not a guarantee that an individual prediction lies within tenfold, and the pooled target combines different assay endpoints. A model can improve substantially enough to help prioritization while still leaving considerable uncertainty around a particular compound.

Stratification reveals the useful scientific pattern. For agonist-labelled test records, DSQ's R-squared is 0.269 and DrugBAN's is 0.099. On antagonist-labelled records, DeepDTAGen has the stronger result: 0.417 versus DSQ's 0.364. The paper appropriately identifies improved agonist modeling as the clearest benefit, rather than claiming universal superiority.¹

This is especially informative because the test composition differs from training. Agonist-labelled records are the minority during learning but number 6,369 of the 11,464 test pairs. The overall score therefore mixes categories in a different proportion from the training data. Reporting both categories is a methodological strength: it prevents the aggregate improvement from obscuring which pharmacological task benefits.¹

Does the physical constraint earn its place?

Comparisons with altered versions of DSQ go beyond a contest against unrelated architectures. Replacing the state gate with a conventional multilayer perceptron increases overall RMSE from 0.999 to 1.051. Accuracy against the agonist–antagonist labels falls from 0.840 to 0.747. Removing the ranking objective sharply reduces agonist regression performance, with R-squared falling to 0.015.¹

These interventions support the authors' case that the components contribute usefully to this fitted model. The simpler gate retains a similar computational role, making it a more informative comparison than merely removing the entire branch. It narrows the explanation for the gain toward the chosen parameterization and constraints, although it does not prove that the internal quantities equal physical state affinities.

A learned temperature parameter controls how sharply the gate changes between its two branches. Fixing it at its initial value weakens performance; learning it improves both classification and agonist regression.¹ Here, temperature is a scaling parameter in the model. Its fitted value describes the numerical calibration of the gate, not a measured temperature or a separately validated physical energy scale.

The interpretability analysis also contains a valuable distinction between highlighting and testing. In a human thromboxane receptor example, the authors zeroed input embeddings at selected sodium-pocket and NPxxY motif positions. For 46 agonist-labelled interactions, the mean predicted activation score changed from 0.5147 to 0.4294. Across 1,204 antagonist-labelled interactions, it remained low, moving from 0.0496 to 0.0393.¹

This is a perturbation of the model's input, not a mutation performed in cells. It shows that the prediction depends on features at the chosen positions. That is stronger than presenting an attention map alone, but the causal statement belongs to the software: removing those features changes its answer. Establishing a corresponding biological effect would require an experimental intervention. The integrated supplementary maps extend the attention analysis across receptor classes; they do not add such an experiment.¹

The receptor on the other side of the split

The most consequential test asks how closely a test receptor resembles receptors used in training. For the subset whose maximum sequence identity to training receptors is at least 80%, DSQ's overall R-squared is 0.584. Below 60% identity, it is −0.081. DeepDTAGen reaches 0.039 in that lower-identity subset, so DSQ does not retain its aggregate advantage there.¹

A negative R-squared means that the predictions perform worse, by this squared-error measure, than assigning the subset's mean activity value. It does not mean every prediction is wrong. It does mean the benchmark offers little support for accurate absolute activity estimation in that distant-receptor setting. The homology analysis is therefore practically valuable: it identifies where the model's demonstrated success stops.

Holding out interaction pairs also differs from holding out every observation for a receptor. The authors explicitly state that the present work does not include a receptor-disjoint cold-target test. Their sequence-identity analysis helps expose transfer difficulty, but cannot substitute for that separate experimental design.¹ For someone interested in a poorly characterized receptor, this is the first boundary to examine before deciding that the model is suitable.

A sensible use of DSQ would therefore keep the assay definition and receptor familiarity alongside each prediction. The paper supports considering its state-aware design when prioritizing activity measurements, particularly in agonist-focused settings resembling the evaluated data. It does not supply an exposure model, a dosing rule or evidence of clinical benefit. Its inputs and outputs concern molecular representations and historical activity records.¹

The achievement is to make a pharmacological idea do measurable work inside a predictive architecture. The same study shows why interpreting that architecture requires pharmacological discipline: a class label, a latent state and an observed cellular response remain different kinds of evidence. DSQ is most convincing when those distinctions clarify its useful advance rather than disappear behind the word “activation.”