experimental limits of detection as model parameters

[wsdewitt]

Observed functional scores are lower bounded by experimental detection limits, while the predicted values aren't. The effect of this can be observed in a scatter plot:

If the limit of detection is known, we can tell the model about it in __init__(), and it can clamp the predictions that would otherwise come out below the limit (in the forward() method for the model).

[wsdewitt]

My suggestion above to clamp output based on the limit of detection is not right—the model wouldn't train because the gradient of the output layer will be 0 for all pre-clamp values less than the limit.