How to apply Enformer on de novo mutations

Hi Enformer developer, thanks for the great tool! I notice that the output of Enformer is the difference caused by a variant on the 5313 tracks. According to the nat meth paper, if we want to predict whether the variant could impact gene expression, we need to train a new classification model (random forest or LASSO) based on these 5313 tracks. But for de novo mutations, there is no reasonable training set to train this classification model. So, given 5313 features on my de novo mutations, is there a way to tell whether they could impact gene expression? Other metrics like CADD or spliceAI would draw a manual cutoff and state that mutations over the threshold are likely influential, and I'm wondering whether there is a reasonable threshold for Enformer output. Thanks in advance for your help!