DeorphaNN

Peptide-activated G protein-coupled receptors control metabolism, endocrine signalling, and neuromodulation, but a large fraction remain orphans: the receptor is known, its endogenous peptide ligand is not. Deorphanization is normally brute force — synthesize hundreds of candidate peptides and test each one in a cell-based activation assay. DeorphaNN, from Larissa Ferguson and colleagues at the MRC Laboratory of Molecular Biology, KU Leuven, and the University of Cambridge, turns that screen into a ranking problem so that the wet-lab work starts at the top of a shortlist.

The starting point is an observation about structure prediction: AlphaFold-Multimer confidence scores partially separate agonist from non-agonist GPCR–peptide complexes, and the separation improves when the receptor is modelled against an active-state template derived from AF-Multistate. The authors then went a step further and looked inside the network. AlphaFold's final hidden-layer pair representations — the residue-by-residue tensors, not the per-residue single representations — carry substantially more agonist signal, and different subregions of that tensor contribute complementary information.

DeorphaNN converts each predicted complex into a graph and learns over it. Receptor and peptide residues are nodes; intramolecular edges follow the primary sequence; intermolecular edges come from Arpeggio, which identifies biophysically defined interatomic interactions rather than a naive distance cutoff. Edge features carry the corresponding slices of AlphaFold's pair representation, so the graph encodes both which residues touch and what the structure predictor learned about that contact.

#Key Features

• Learns from AlphaFold's internals, not just its scores: Pair representations extracted from the final hidden layer outperform single representations and per-residue pLDDT, and let the model surface agonists that a confidence-score ranking would miss.
• Active-state structural input: Complexes are predicted against AF-Multistate-derived active-state templates, so the modelled pose reflects the conformation associated with receptor activation.
• Chemistry-aware graph topology: Ablations show that Arpeggio-defined interaction edges plus pair-representation edge features beat distance-based (≤6 Å) edges and beat either component alone.
• Frozen-checkpoint inference: An ensemble of ten released checkpoints is applied to new receptor–peptide complexes through deorphann_batch.py; retraining from scratch is offered only as an optional notebook path.
• Cross-species generalization: A model trained on C. elegans data ranks agonists above chance for Platynereis dumerilii receptors and reaches mean average precision 0.72 on a curated human dataset.

#Technical Details

The graph network uses GATv2 attention convolutions with 128-dimensional node inputs and 128-dimensional edge features, batch normalization, global mean pooling, and a linear binary classification head. Training data came from a system-wide C. elegans screen covering 161 putative GPCRs in an aequorin-based, CHO Gα16-mediated calcium mobilization assay; running every receptor–peptide combination through AlphaFold-Multimer produced 20,035 complexes, of which 457 were experimentally confirmed agonist pairs, and filtering yielded a final training set of 65 unique receptors from 55 genes. Evaluation uses mean average precision under stratified group cross-validation. On an augmented human benchmark — literature-curated agonist pairs plus synthetic non-agonists chosen by ESM-2 embedding distance — DeorphaNN reaches mAP 0.72 against a random baseline of 0.30, while SpatialPPIv2, a general protein–protein interaction model, reaches 0.33, statistically indistinguishable from random.

#Applications

The intended workflow is prospective deorphanization: predict complexes for a receptor against every candidate peptide in the organism's peptidome, rank them with DeorphaNN, and assay the top handful. The paper does exactly this. For the orphan receptor NPR-44, the peptide NLP-69-1 ranked 5th of 364 candidates and activated the receptor with an EC50 of 231.7 nM; for NPR-33, the top-ranked peptide NLP-70-2 gave an EC50 of 76.3 nM. In retrospective checks, known-but-held-out agonists for NPR-34 and SEB-2 both ranked first out of 364. Beyond neuropeptide biology, the same ranking is useful to pharmacologists prioritizing peptide leads for human receptors that still lack a confirmed endogenous ligand.

#Impact

DeorphaNN is a concrete demonstration that AlphaFold's intermediate representations are a usable feature source for interaction problems the model was never trained on, and it is one of the few structure-based screening papers to close the loop with new wet-lab-validated ligands rather than retrospective benchmarks alone. Practical caveats are worth stating plainly. The model is a ranker, not an end-to-end predictor: every candidate pair must first be run through AlphaFold-Multimer with an active-state template and then through Arpeggio, which dominates the compute cost. Training data are dominated by high-potency interactions (EC50 below 500 nM), and the paper notes that a genuine but weaker agonist, NLP-70-1, ranked only 156th of 364. The work remains a preprint. The code is MIT-licensed on GitHub, and the ten pretrained checkpoints, AlphaFold-Multimer complexes, contact tables, and active-state templates are all published in a single Hugging Face dataset repository.