Proteins constitute the biomolecular machinery of all living
organisms. Their function crucially depends on the proper
folding of the polypeptide chain into stable conformation.
Proteins evolved towards a specific function optimized to
maximal performance near the physiological temperature of their
host organisms. Consequently, most proteins are only marginally
stable which hinders their use in biotechnological or
pharmaceutical industries, where extreme conditions are often
required. Moreover, unstable proteins are prone to aggregation,
which can lead to their aberrant function and the production of
toxic species. Several dozens of human pathologies are linked
to protein misfolding and amyloid deposition, which represent a
high socio-economic burden. Accurate prediction of structural
and energetic changes upon mutation and the ability to
stabilize a protein in a fast and reliable manner is therefore
highly desired. Stability and aggregation predictors often
exploit machine learning. However, their performance is still
suboptimal largely due to the poor quality and/or size of
available training datasets. Collection and appropriate
analysis of high-quality data from (un)folding studies of
proteins from diverse families are therefore essential for
their further improvement.