GPCR Ligand Bioactivity Prediction with Physics-Informed Dual-State Query Learning
Shuo Zhang, Huifeng Zhang, Rongqi Hong, Jian K. Liu
q-bio.QM
Quick verdict
DSQ's MWC-inspired gate lifts GPCR agonist prediction over general baselines, but single-run metrics and absent GPCR-specific baselines cap confidence.
Summary
The paper introduces DSQ (Dual-State Query), a multimodal deep learning architecture for predicting both binding affinity and functional efficacy (agonist versus antagonist) of small molecules against GPCRs. Two sets of learnable queries extract putative active- and inactive-state representations from frozen ESM-2 protein embeddings and frozen MoLFormer ligand embeddings, and a 'neural MWC' gating module combines the two state-specific affinity predictions through a sigmoid of their difference minus a learned per-receptor energy barrier, motivated by the Monod-Wyman-Changeux allosteric model. Training adds gate supervision, an orthogonality constraint between state views, and a margin-ranking loss enforcing differential affinities. On a 260k-pair human GPCR benchmark curated from GLASS and GPCRdb (222 training receptors, 203 test receptors), DSQ attains overall RMSE 0.999 and R2 0.517, beating retrained DeepDTAGen, DrugBAN, and GraphDTA, with the largest gains on the agonist subset (R2 0.269 versus negative or near-zero for baselines), though it loses to some baselines on antagonist metrics. Ablations attribute the gains to the MWC gate parameterization, learnable temperature, ranking loss, and orthogonality. Attention analyses recover known activation motifs (2x50, 7x53), and perturbing them shifts predicted activation probability. Key limitations, mostly acknowledged by the authors: all models fail on low-homology receptors (DSQ R2 -0.081 below 60% identity), no receptor-disjoint cold-target evaluation, no comparison against GPCR-specific baselines such as AiGPro, and no variance or significance testing is reported.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 3
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Embedding the MWC allosteric model as a differentiable gate yields better GPCR bioactivity prediction than general DTI baselines, with the clearest gains on agonists (overall R2 0.517 vs 0.446 best baseline; agonist R2 0.269 vs at most 0.099).
- At inference DSQ produces activation probability and bioactivity jointly, without an externally supplied agonist/antagonist class token.
- The MWC gate parameterization, learnable temperature, ranking loss, and orthogonality constraint are each necessary, per ablations (e.g., removing the ranking loss collapses agonist R2 to 0.015).
- Learned query attention recovers canonical activation switches (2x50 sodium pocket, 7x53 NPxxY motif), validated by perturbation that lowers predicted agonist activation probability from 0.5147 to 0.4294.
- Sequence-based models including DSQ fail on low-homology receptors (DSQ R2 -0.081 below 60% identity), an acknowledged generalization boundary.
Strengths
Honest negative results
The paper openly reports that DSQ loses to DeepDTAGen and GraphDTA on several antagonist metrics (antagonist RMSE 1.125 vs 1.078/1.120; R2 0.364 vs 0.417/0.370) and documents low-homology failure rather than hiding it (§4, Table 3).
Above-average evaluation breadth
Beyond Table 2 leaderboarding: homology-stratified evaluation (Table 3), a five-variant ablation (Table 4), differential-affinity consistency checks (74.91% agonists / 90.17% antagonists correctly ordered, §4.3), motif perturbation (§4.4), and quantitative latent-space metrics (Silhouette 0.0022→0.6086, Table 5).
Practical inference design
Interpretability anchored in known biology
Reproducibility infrastructure
Weaknesses
No statistical uncertainty anywhere
Missing the most relevant baseline
Train/test receptor overlap likely inflates headline results
Physics framing exceeds what the math enforces
Efficacy binarization underspecified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Preliminary Review: DSQ - Physics-Informed Dual-State Query Learning for GPCR Bioactivity
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the DSQ model for GPCR bioactivity prediction?
DSQ (Dual-State Query) is a deep learning model that predicts, for a protein-ligand pair, both how strongly a small molecule binds a GPCR (binding affinity) and whether it activates the receptor (agonist) or blocks it (antagonist). It uses frozen ESM-2 protein and MoLFormer ligand encoders, learnable 'query' vectors that probe active and inactive receptor states, and a gating layer inspired by the Monod-Wyman-Changeux allosteric model.
How well does DSQ perform compared to other models?
On a benchmark of about 260,000 human GPCR-ligand pairs (222 training, 203 test receptors), DSQ reaches overall RMSE 0.999 and R2 0.517, ahead of retrained DeepDTAGen, DrugBAN, and GraphDTA. Its clearest advantage is on agonists (R2 0.269 versus negative or near-zero for baselines); on antagonist metrics some baselines beat it, and all tested models fail on receptors with less than 60% sequence identity to the training set.
Is the code and data for DSQ available?
The authors state that code, trained weights, and the dataset are available at github.com/jiankliu/DSQ. Note that the underlying database GLASS2 is cited as still in submission, so reproducing the exact data split may depend on the authors' released files.
Does DSQ really simulate receptor thermodynamics?
Not fully. It embeds an MWC-inspired formula: activation probability is a sigmoid of (active-state affinity minus inactive-state affinity minus a learned per-receptor energy offset), divided by a learnable temperature. This is a soft architectural constraint - after training the 'temperature' settles at 0.2627 and the energy offsets are unitless calibration parameters, not physical quantities.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Text extraction truncated several sections: the tail of §2 (Problem Definition), part of §3.7 (dataset construction and split stratification), some Table 4 rows (e.g., full simple-gate/w-o-gate entries), and part of Related Work - judgments there rest on surrounding prose.
- Figures 1-5 and S1-S2 are unavailable as images; interpretability (§4.4) and t-SNE (§4.5) claims were assessed from text descriptions only.
- Agonist/antagonist label assignment criteria and the bioactivity unit (presumably a pKi/pKd scale) are not visible in the extracted text.
- The GitHub repository was not inspected; code, weights, and data availability are as stated by the authors.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.