Even with only the random seed being different, the prediction results vary greatly.

With the same training data and the same training parameters, but only changing the random seed, the following situations occur:

The ROC does not change much, but the top-20 accuracy varies greatly. Is this reasonable?

For the compounds with the highest predicted probabilities, some seeds can reach 83%, while others only reach 40%.

Using the model that had the best top-prediction accuracy on the first test set to predict the second test set, the top performance on the second test set is very poor.

Seeking help!!! Is there anyone in the field who can discuss this with me?
Facing these situations, what can I do? My goal is to use a model with high top-prediction accuracy to perform virtual activity screening on a new large compound library.