There's no easy answer because complex stability hasn't been systematically studied and can be defined in different ways and is also subject to many influences. The notion of stability you want is one that is relevant for the type of experiment you've used to derive your data. For example detection is biased towards stable complexes in AP-MS experiments but that may not be the case if you derive complexes from some other type of experiments like yeast two-hybrid. Note also that it is usually assumed that complexes found in AP-MS experiments are stable. This is without even considering the definition of what constitutes a complex (see for example here).
Anyway, here are some ideas:
- use complex member protein stability as a proxy, the idea being that if a member is unstable, then the complex would also be unstable (probably not true in many cases though because being in a complex may stabilize a protein),
- use information on the protein-protein interfaces e.g. binding energy,
- compare complexes found in AP-MS experiments with and without cross-linking (which would stabilize transient interactions).
- train a machine learning model to predict stability based on a training data set of curated complexes.