The current design of AptaTransPipeline appears inconsistent with the expected scikit-learn–style API since it lacks fit and predict methods. Without these common interfaces, it’s unclear how benchmarking is intended to operate, especially given that AptaNetPipeline does follow this pattern: there’s no shared method that can be called on both classes for fitting and generating predictions. Hence, we should add fit and predict methods to AptaTransPipeline.

Additionally:

1. The description of prot_words should be standardized and rewritten in a clear, generic way across all instances where it appears, as its current form is overly confusing.
2. The predict function should explicitly document the datatype returned within the Tensor; at present, users have to inspect the evaluate function to infer what is being returned.