Pocket-ligand Representation Alignment
This is a research question note, not an implementation result. It defines the object, split, and metric boundary needed to compare pocket-ligand representations.
Motivation
Protein-ligand modeling often mixes heterogeneous objects: protein sequence, pocket geometry, ligand graph, ligand conformer, and interaction pattern. A useful representation should preserve what each object contributes while still making cross-object comparison and prediction possible.
Question
Can protein pocket, ligand conformer, and interaction features be aligned into a shared representation space without hiding the difference between sequence-level, graph-level, and 3D structure-level evidence?
Hypothesis
Pocket-ligand alignment may improve retrieval, docking candidate ranking, and affinity-related pretraining when the objective explicitly separates object identity, geometry, and interaction evidence. The same setup can fail if the contrastive pairs are defined by noisy docking output, protein-family leakage, or ligand scaffold shortcuts.
Scope
| Axis | Include | Exclude |
|---|---|---|
| Object | protein, pocket, ligand, complex | private targets or unpublished assay details |
| Representation | sequence embedding, molecular graph embedding, conformer embedding, pocket embedding | one score treated as a universal representation |
| Learning | contrastive learning, masked modeling, multi-task pretraining | benchmark-only leaderboard chasing |
| Architecture | Transformer, GNN, equivariant model | architecture names without input/output contract |
| Evaluation | retrieval, ranking, pose quality, split robustness | single aggregate metric without denominator |
Method Axis
| Axis | Notes |
|---|---|
| Entity contract | Define whether the sample is a ligand, pocket, complex, or interaction label. |
| Positive pair | Use experimentally supported or carefully filtered protein-ligand relation when possible. |
| Negative pair | Avoid trivial negatives that only teach scaffold, protein-family, or dataset-source shortcuts. |
| Geometry | Keep 3D evidence separate from 1D sequence evidence until the claim requires fusion. |
| Calibration | Check whether representation similarity is meaningful as a ranking score or only as a retrieval feature. |
Evidence Plan
- Compare ligand-only, pocket-only, and pocket-ligand joint representations on the same public split.
- Track scaffold split, protein-family split, and pocket-similarity split separately.
- Report failure cases where high embedding similarity does not imply plausible binding pose.
- Separate retrieval quality, docking pose quality, and affinity ranking as different claims.
- Record the denominator: filtered complexes, failed conformer generation, failed docking, and invalid molecules.
Project Handoff
If this becomes an implementation task, the project artifact should be a public-safe benchmark notebook or pipeline that takes public complexes, builds representation pairs, and reports split-aware retrieval/ranking metrics. That belongs in Projects, while this note keeps the research question and evaluation boundary.