You
know that in the ChEMBL group, we love to play with the data we collect!! Back
in April 2014, we started to work on a target
prediction tool. Wow! This was almost 5 years ago! Since then,
we have continued to update the tool for each new ChEMBL release, providing you
with the actual models and the result of the prediction on the ChEMBL website
for the drug molecules. The good news is that these target predictions are not
dead and a successor is on its way!
First,
we would like to introduce you some closely related work. You may have heard
about conformal prediction (CP). If not, it is a machine learning framework
developed to associate confidence to predictions. I personally consider this as
a requirement for decision making. Basically, you train a model as you would do
in QSAR but then you first predict a so-called calibration set, for which you
know the actual values. For each of these observations you obtain two
probabilities: one for the active and one for the inactive class (in a typical
classification scheme). Now that you have this information, each time you
predict a new compound you compare its probabilities to those of the
calibration set (the non-conformity scores as they are called) and you derived
p-values for each class. Based on your predefined significance level, the
compound can be assigned in different categories: only active or only inactive,
but also both active and inactive or none of them. I am sure you can start
seeing here the added value of CP!
Here
I have briefly detailed how it works for classification models but CP can also
be applied to regression models. If you want to know more about conformal
prediction, I recommend you to read this book and
also this very nice example
of the application in drug discovery. Having learnt how to build conformal
predictors, we were intrigued to know how well they perform against traditional
QSAR models with our ChEMBL data!
With
this in mind, we decided to build a panel of models using a substantial data
set from ChEMBL. With our new protocol, we were able to build models for 788
targets (550 of them human targets). For the descriptors we used RDKit Morgan
fingerprint (2048 bits and radius 2) and 6 physicochemical descriptors. For the
machine learning part we used the good old Random Forests as implemented in
Scikit-learn version 0.19. For the QSAR models, this is all that is needed, but
for CP you need a framework and this was provided by the very nice library
provided by Henrik
Linusson.
The
next part consisted of training the models and checking their internal
performance, but we went a bit further and decided that with our models trained
on ChEMBL_23 data, it would be interesting to see how they perform with new
data in ChEMBL_24 in a so-called temporal validation. All the details, results
and conclusion are presented in the recently accepted article!
The
dataset for each target is already available here and
you can find the models ready to use there.
Feel
free to take a look and to share your opinion in the comments.
Now, you remember that I started this post mentioning our good old target predictors. So does it mean a new generation of ChEMBL models using conformal prediction is ready to be launched for our users? Well, unfortunately not yet, but stay tuned!
