There will always be a trade off between sensitivity and specificity in this type of analysis. If you want to predict the docking ranking ( high/low)from the descriptor ranking(high/low) , high sensitivity will be gained at the cost of low specificity.

I am not an expert in this subject area and do not know your aims in doing this.

Is it cheaper to do chemical descriptors than docking experiments?

If you want to be sure of getting all high ligand rankers (high sensitivity) you would want the chemical descriptor to be more generous,with resultant low specificity.

this might necessitate more confirmatory docking experiments,at cost.

A confusion matrix is typically an n by n matrix which contains counts that are related to some sort of classification scheme. The set of actual values and the set of predicted values require the same list of classification levels. Numbers on the diagonal (down, lft to rt) are counts were the observed classification level agrees with the predicted classification level. Numbers outside the diagonal indicate counts were a datum for an observed classification level was predicted as another, specific, classification level. The statistical model generates the predictions from the actual values. Any thresholds are part of the statistical model not the confusion matrix. As such thresholds are adjusted by changing the statistical model. Nevertheless, changing a threshold may alter the predictions and, thereby, require changes in the confusion matrix counts. These are paired values, because each observation is paired with its prediction. I hope this helps.

James Leigh and Michael Theobald thank you very much. Your answers helped me a lot, now it is clear what was my problem!