Thermostability predictions are generated based on a number of structural, residue and alignment features for a given protein. We use a gradient boosting model trained to predict the thermostability to deliver the probability variants to 3DM. The thermostability prediction value indicates how likely it is that a given variant results in a net positive ΔΔG value, thus resulting in a more stable protein on the whole.
Predictor performance
The overall accuracy of the predictor is 71% on proteins it was not trained on. Because most variants will not have a thermostabilizing effect it is easier to predict non-stabilizing variants (Fig. 1) and more difficult to correctly predict thermostabilizing variants. Therefore this number is not representative for identifying thermostabilizing variants.
On the subset of thermostabilizing variants the accuracy is around 40%. This means that when you select the top 10 thermostabilizing variants, around 4 will have a thermostabilizing effect on the protein. Thus, this predictor has a much higher chance to identify stabilizing variants compared with random mutagenesis.
.png?version=1&modificationDate=1549019942588&cacheVersion=1&api=v2&height=250)
Figure 1. Most variants (below 0.5) have a neutral or destabilizing effect.