ProteinExt3: Adaptive Fusion of Protein Language Models for Protein Function and Target Annotation

Prospective protein function prediction requires evaluation designs that control for temporal leakage and sequence homology while accounting for the strong imbalance of Gene Ontology (GO) labels. We developed ProteinExt3 to integrate complementary sequence- derived evidence for GO-based protein function prediction under a temporally separated and homology-controlled evaluation protocol.

ProteinExt3 is a late-fusion framework that combines representations extracted from ESM- 2 layers 20 and 33, ProtT5 representations, 63 sequence-derived physicochemical features, and BLASTp homology transfer. Fusion weights were learned from out-of-fold (OOF) predictions, and final scores were propagated through the GO directed acyclic graph. The benchmark used experimentally supported GO annotations, a temporal split based on the earliest experimental annotation date, and MMseqs2 filtering between training and prospective test proteins.

ProteinExt3 reached Fmax values of 0.226 for Biological Process (BP), 0.547 for Cellular Component (CC), and 0.669 for Molecular Function (MF). Among the evaluated learning-based methods, ProteinExt3 achieved the highest Fmax in all three namespaces. For CC, the frequencybased Naive baseline retained the highest Fmax (0.569), whereas ProteinExt3 achieved higher AUPR and lower Smin. Ablation results showed that the effects of homology transfer, handcrafted features, and pooling differed across namespaces and metrics.

Most of the OOF gain came from combining prediction branches; optimizing their weights provided a smaller improvement. Component effects varied across GO namespaces and evaluation metrics.

ProteinExt3 combines complementary sequence-derived evidence to support protein function annotation and downstream target prioritization.