Question 1

Why is random splitting inadequate for binding prediction?

Accepted Answer

Chemical series contain many near-identical ligands and protein sets contain close homologues, so a random split puts near-duplicates on both sides and the model can succeed by recognition rather than generalisation. Scaffold splits, protein-family splits and time-based holdouts each test a different generalisation claim, and the one reported should match the intended use.

Question 2

How should heterogeneous affinity labels be handled?

Accepted Answer

Values measured by different assay formats are not interchangeable, and pooling them introduces noise that no architecture can overcome. Restricting to one measurement type, or modelling assay identity explicitly, keeps the target coherent. Estimating experimental reproducibility from repeated measurements gives an upper bound against which model error can be judged honestly.

Question 3

Which baselines must a deep model be compared against?

Accepted Answer

A nearest-neighbour predictor using chemical similarity, a descriptor-based regression model, and where structures exist a conventional docking score all provide meaningful references. Many published gains vanish against these once splits are corrected. Reporting them alongside the proposed model is the clearest way to show what the added complexity actually buys.

Question 4

Can attention or attribution maps be treated as binding-site evidence?

Accepted Answer

Not without validation. These maps show what the model weighted, which may reflect dataset artefacts rather than physical contacts. Comparing highlighted residues with resolved interaction data, or with mutagenesis results where available, converts an appealing visualisation into a supported claim. Agreement should be quantified across many examples, not shown for one case.

Question 5

How do predicted protein structures change what is possible?

Accepted Answer

They extend structure-aware models to targets without experimental coordinates, but confidence varies considerably by region and side-chain placement in binding pockets remains uncertain. Filtering by local confidence scores, and testing whether performance depends on that confidence, distinguishes genuine structural signal from a model relying mainly on sequence features.

Question 6

What counts as prospective validation in this field?

Accepted Answer

A prediction made before the measurement exists, then tested experimentally or against data released afterwards. Retrospective performance on any held-out portion of an existing collection does not qualify, however carefully split. Even a small prospective set, arranged with a collaborating laboratory, carries more evidential weight than a large retrospective benchmark.

Biotechnology

Biotechnology

Research Frontiers in Deep Learning Drug-Target Binding