A cascade model is a sequential workflow in which one model’s output becomes the next model’s input. In drug discovery, this can connect structure prediction, docking, and property prediction. section 13.4.3
Structure-to-docking cascade
ESMFold can produce a protein structure in PDB format, which a docking model such as DiffDock consumes alongside a ligand in SDF format. A property predictor may then consume canonicalized SMILES. Conversion failures, incompatible conventions, and explicit-hydrogen or bond-order requirements can break the pipeline. section 13.4.3
Evaluate cascades as complete workflows because:
- Errors propagate between stages; two independently 90%-accurate models yield at most 81% combined accuracy under the stated assumption. section 13.4.3
- Correlated errors, especially around predicted binding pockets, can make real performance worse than independent-error estimates. section 13.4.3
- Reproducibility requires tracking component versions, training datasets, preprocessing, and inference parameters. section 13.4.3
