BioNTech/InstaDeep posted a paper about their new machine learning model, which predicts if a new variant's spike protein sequence is concerning. (Hundreds of new variants are generated every day). It produces 5 scores, pertaining to potential for immune escape or transmissibility, which are aggregated into a single "Pareto" score. The underlying scores are all independently validated against in-vitro assays, which suggests the methodology has some utility.
But, for what? The model is not proven to accurately predict if a new variant will be of concern. Rather, when it was run on data from Sep 2020 to Nov 2021, and forced to flag 20 sequences every day, it eventually flagged all the WHO's variants of concern (except for Delta), usually over a month in advance. So, it is useful tool to focus surveillance resources, but not to produce a dashboard of "these will be scary".
Also, the t-tests in Table S.5 don't really indicate that the score is useful. Only 13 variants, out of thousands, are designated as concerning. If most variants are obviously not concerning, getting a low p-value is easy.