ProteinRPN

A 422-residue enzyme may carry its catalytic function in ten residues that sit together in three-dimensional space while being scattered along the sequence. Most structure-based function predictors never look for those ten: they embed the whole residue graph, predict Gene Ontology terms from the pooled representation, and only afterwards run a saliency map to ask which residues mattered. That retrospective answer is typically a diffuse spray of positions under no requirement to cluster, and because it plays no part in the prediction, it cannot improve it.

ProteinRPN reverses the order. Borrowing the region proposal network from Faster R-CNN, it first nominates candidate functional regions inside the residue graph, prunes them to a small coherent set, and only then predicts function from what survives. A protein becomes a graph whose nodes are residues and whose edges join Cα atoms within 10 Å. A k-layer graph convolutional network gives every node a receptive field covering its k-hop subgraph — k = 2 proved sufficient — so each node stands in for a region, or anchor. A graph attention layer classifies each anchor by whether it contains functional residues, tuned for recall; a node-drop pooling layer supplies the missing precision by scoring residues on three signals at once: penumbral cone attention, a hyperbolic alternative to dot-product attention that encodes the nesting of secondary structure inside domains inside complexes; DSSP secondary structure, since catalytic residues sit preferentially in helices and sheets; and a proximity score rewarding residues tightly clustered with a few neighbours rather than loosely near many. A functional attention layer re-weights edges toward the survivors, and a Graph Multiset Transformer pools the graph into one embedding that an MLP reads out as a probability vector over each sub-ontology's GO terms.

The model comes from Shania Mitra, Lei Huang and Manolis Kellis at MIT CSAIL and City University of Hong Kong, posted to arXiv in September 2024. Node features are frozen ESM-1b embeddings summed with a learned residue-identity embedding, so the sequence signal a protein language model already captures is inherited rather than relearned.

#Key Features

• Region proposals before prediction: Functional residue detection is a stage of the pipeline, not a post-hoc explanation of it, so localization improves accuracy rather than merely describing it.
• Domain knowledge in the pooling layer: Pruning is scored jointly by hierarchy-aware hyperbolic attention, DSSP secondary structure and inverse-distance proximity, not by a learned importance score alone.
• Structure-preserving pruning: A connected-components penalty in the loss discourages the selected functional nodes from fragmenting, keeping proposals as coherent clusters.
• Contrastive representation learning: A supervised contrastive loss groups proteins sharing GO terms while an InfoNCE term matches each protein against a perturbed copy of its own structure.

#Technical Details

The region proposal module is pretrained on PDBSite: 4,723 experimentally characterized active sites spanning 197 functions across 603 PDB structures, split 80:20. Those same 603 structures supplied the motivating analysis — functional residues cluster spatially, and a protein of several hundred residues carries only 1 to 30 of them, hence the multi-stage pruning. On the pretraining task the module reaches 0.95 ROC on anchor functionality prediction and 0.85 on pruning. The full framework is then trained on the HEAL dataset, an adaptation of the DeepFRI benchmark comprising 36,629 sequences from the PDB and 42,994 from SWISS-MODEL, under a combined cross-entropy, contrastive and connected-components loss. Under the CAFA metrics, ProteinRPN reaches protein-centric Fmax of 0.618 (BP), 0.691 (CC) and 0.754 (MF) against 0.581, 0.673 and 0.708 for HEAL; the roughly 7% Fmax improvement the paper headlines covers BP and MF, with a smaller 2.7% gain on CC. AUPR reaches 0.344, 0.459 and 0.683 and Smin falls to 0.495, 0.458 and 0.335, ahead of BLAST, FunFams, DeepGO, DeepFRI and HEAL on every metric. Ablations confirm that both the contrastive losses and the domain-knowledge scoring contribute.

#Applications

The model annotates a solved or predicted structure with GO terms and, in the same pass, marks the residues it believes carry each function — useful for choosing mutagenesis targets in a protein whose active site is uncharacterized, or for triaging UniProt entries lacking experimentally supported annotation. On chain 2BCC-B (422 residues, 10 functional) it recovers 8 within proposals covering 28 residues; on 2CHG-A (226 residues, 11 functional) it recovers 9 within 43.

#Impact

ProteinRPN is, by the authors' account, the first application of graph region proposals to protein function prediction, and the idea generalizes to any graph task whose label is carried by a small coherent subgraph. The caveats: evaluation is entirely computational, with no wet-lab confirmation of the predicted functional residues; the authors note the analysis rests on a limited set of structures; and while the training and evaluation code is public, it carries no license and no trained checkpoints are distributed, so using the model means retraining it.