Protein representation learning
APObind binding-site experiments
Residue-level ligand-binding-site experiments combining ESM-2 protein-language embeddings with graph attention over protein structure.
A protein can adopt different structures with and without a ligand. This work explores whether sequence-derived representations and structural message passing can identify which residues form a binding site.
Do protein-language-model embeddings become more useful for residue-level binding prediction when they are refined over a structural graph?
My contribution
- Developed the biomedical machine-learning experiment around ESM-2 embeddings and Graph Attention Networks.
- Worked with the recorded APObind collection of 10,599 apo–holo pairs.
- Structured evaluation around PR-AUC, ROC-AUC, F1, MCC and explicit class-imbalance handling.
Signature view
Trace the evidence.
Move across the research chain. The case study below remains complete and readable without this interaction.
Question
Locate the residues a ligand may recognise
Sequence representations contain useful context, but ligand binding is also a three-dimensional neighbourhood problem.
Method, as a sequence of decisions
- 01
ESM-2 converts protein sequences into contextual residue embeddings.
- 02
Residues become graph nodes; structural proximity defines message-passing edges.
- 03
Graph attention learns which neighbouring residues are informative for each node.
- 04
A residue classifier estimates binding-site likelihood under strong class imbalance.
What the record actually supports
Recorded apo–holo pairs
Dataset scale reported in recent CV materials.
Prediction granularity
The target is a label for each amino-acid residue, not a whole-protein class.
Evaluation plan
ROC-AUC, PR-AUC, F1, MCC and class-imbalance analysis; final values are not claimed.
Evidence: recent CV records and achievement inventory. Dataset scale and experimental scope are reported; publication, repository and final metrics remain unverified.
Choices and trade-offs
Why combine sequence and structure?
Protein language models capture evolutionary and sequence context; a graph can add local three-dimensional relationships that sequence order alone misses.
Why prioritise PR-AUC and MCC?
Binding residues are sparse. Accuracy and even ROC-AUC can appear strong under imbalance, so precision-recall behaviour and balanced correlation matter.
What has not been established
- Final benchmark values, split protocol and a public repository were not recovered, so no performance claim is published here.
- Apo and holo structures can introduce pairing, leakage and conformational-bias questions that require careful split design.
- Predicted sites would require structural and experimental interpretation before biological use.
What I would investigate next
- Q1How does performance change under sequence-identity-aware splits?
- Q2Which graph edges and attention patterns remain stable across apo/holo conformations?
- Q3What calibration is needed for residue probabilities to support experimental triage?
No public paper or repository is attached yet. The case study is intentionally limited to the evidence in the supplied record.
Peer notes / APObind
See a missed assumption?
Research suggestions, methodological critique and relevant collaboration ideas are particularly useful.