AI drug discovery has been around for a long time, but in many cases, it is not as accurate as expected in actual use. This is due to the well-known problem that AI models do not predict well for compounds that are far from the training data. We have attempted to solve this problem by using a large scale training data set of one billion compounds. We show the usefulness of our method by predicting measured initial toxicity and pharmacokinetic parameters on real-world data, of which the distribution is different from the training data. Furthermore, we will show how our model can be modularly added to existing and new AI drug discovery projects to easily develop highly accurate models.