Characterization tools estimate different endpoints: stability, solubility, aggregation, immunogenicity, affinity, kinetics, molecular descriptors, or ADMET properties. Define the endpoint and the decision it supports before choosing a predictor.
Match the endpoint to the input
| Question | Tools to investigate | Interpretation requirement |
|---|---|---|
| Could a mutation affect stability or binding? | ProteinMPNN-ddG, StaB-ddG | Correct reference, residue mapping, score direction, and supported mutation context |
| What property risks affect protein candidates? | Prot2Prop, DeepViscosity, BioPhi | Modality, property definition, training scope, and conditions |
| How might an enzyme perform? | RealKcat, CatPred, EpHod | Substrate identity, kinetic endpoint, units, and operating conditions |
| Which small molecules have property liabilities? | ADMET-AI, Admetica, RDKit descriptors | Chemical state, endpoint-specific applicability, and uncertainty |
Not every tool accepts the same input representation or predicts the same quantity. A kinetic range, a classification probability, and a continuous regression estimate require different comparisons.
Prepare a consistent candidate set
Keep a stable candidate identifier, exact sequence or chemical structure, reference variant where applicable, and required structural context. Preserve protonation and stereochemistry for compounds and chain/residue numbering for mutation-based tools.
Check that every candidate is within the model’s supported domain. Predictions for unusual modalities or far-out-of-distribution chemistry may be numerically valid but scientifically unreliable.
Interpret and combine evidence
Read each endpoint in its own units and direction. Record uncertainty or model disagreement when provided. Avoid averaging unlike scores into an unexplained composite ranking.
Compare predictions to controls or measured data representative of the same assay conditions. Agreement across correlated predictors is weaker evidence than genuinely independent methods or experiments.
Plan follow-up measurements
Use predicted liabilities to prioritize assays, reduce a library, or choose informative mutations. Do not present an ADMET prediction as a definitive safety determination or a stability score as guaranteed expression success.
Explore protein developability, ADMET, or fine-tuning when you have relevant measured data for a custom endpoint.