Topology aware graph attention structure-based protein function prediction

Author

• Jacob Tanchak, University of North FloridaFollow

Year

2026

Season

Spring

Paper Type

Master's Thesis

College

College of Computing, Engineering & Construction

Degree Name

Master of Science in Computer and Information Sciences (MS)

Department

Computing

NACO controlled Corporate Body

University of North Florida. School of Computing

Committee Chairperson

Dr. Anirban Ghosh

Second Advisor

Dr. Indika Kahanda

Rights Statement

http://rightsstatements.org/vocab/InC/1.0/

Third Advisor

Dr. Ayan Dutta

Fourth Advisor

Dr. Yongzhao Wang

Department Chair

Dr. Nan Niu

College Dean

Dr. William Klostermeyer

Abstract

Protein function prediction remains challenging because sequence and structural data now scale far faster than experimental functional annotation. This work studies Gene Ontology prediction as a hierarchical multi-label learning problem and investigates whether topology can provide useful structural signal beyond strong sequence and graph-based representations. We introduce TopoGO, a topology-aware framework that combines residue-level ESM2 embeddings, AlphaFold-derived residue graphs, and persistent-homology summaries of protein structure within a unified model. Persistent-homology features are con- verted into fixed-length embeddings through a learnable topology vectorizer and injected into graph learning through Feature-wise Linear Modulation (FiLM), allowing global structural summaries to influence residue-level message passing. TopoGO is evaluated in a CAFA6-based setting using a top-five-taxa benchmark with identity filtering to limit redundancy between training and test proteins. The results show that TopoGO is a strong and competitive structure-aware predictor, with its clearest advantage in Molecular Function, where it matches the strongest baseline in Fmax and achieves the best reported precision-recall performance. Across the benchmark, TopoGO also remains consistently stronger than PANDA-3D and DeepFRI, while internal ablations indicate that topology contributes a meaningful complementary signal rather than serving as the dominant source of predictive power. The observed effects are strongly branch- specific, suggesting that topology is most effective when used as a targeted conditioning modality rather than as a universal replacement for sequence or structural learning. Overall, these findings support topology-aware conditioning as a promising direction for improving structure-based protein function prediction.

Suggested Citation

Tanchak, Jacob, "Topology aware graph attention structure-based protein function prediction" (2026). UNF Graduate Theses and Dissertations. 1410.
https://digitalcommons.unf.edu/etd/1410