I believe better featurizations are required, but there is a limit to physiological activity prediction only from ligand information 'cause it's a molecular recognition.

MoleculeNet: A Benchmark for Molecular Machine Learning

Results for biophysical and physiological datasets are currently weaker than for other datasets, suggesting that better featurizations or more data may be required for data-driven physiology to become broadly useful.

What features are multitask modering shared in physiological activity prediction ?

Multitask modeling relies on the fact that some features have explanatory power that is shared among multiple tasks.

I'm guessing it's a liposoluble feature such as logP expressed in Hansch-Fujita equation.