A well-tuned nearest-neighbor baseline matches a multitask neural net for target prediction, but blending the two beats both
- Authors built a scaffold-disjoint ChEMBL benchmark covering many human protein targets, with hyperparameters fixed on a validation split.
- A tuned nearest-neighbor similarity search and a multitask neural network ranked targets about equally well on the held-out set.
- A weighted blend of both gave a small but statistically significant gain; evidence is retrospective only, with no experimental validation.
💡 Points Based on the abstract
A scaffold split still leaves close analogs across the boundary, so a tuned nearest-neighbor matching the neural net says as much about the benchmark as the models. The blend's gain is small, so check effect size per target, not just significance. A prospective test would be more telling: how do both rank targets for a novel chemotype with sparse ChEMBL coverage?
Integrating classical similarity and multitask neural representations ↗