Recently I’m learning conformal prediction and today I used the ready to use model for target prediction which is trained by chembl_24 (little bit old) with lightGBM.
ChEMBL team provides nice package. You can get the model as docker image. URL is below.
And original article is here.
To use the image in your local environment it is very easy just type following code.
docker run -p 8080:8080 chembl/mcp
After typing the code, docker image will run and server will be launched. Now target prediction server is ready to use. ;)
Original github repo describes how to use the server from terminal with curl command.
BTW, I would like to use it from python. It is easy to do it by using requests package.
My example is below.
The model predicts that the compound has activity against Xanthine dehydrogenase. (Maybe the compound was included training data…)
It is interesting point that conformal prediction returns not only active/non active label but also both/empty label. It means the model predicts posi and nega with uncertainly.
To curate data and build model requires several steps and time. So ready to use system is user friendly. If one is allowed to wish so much I would like to build model with new version of chembl so I will be happy if source trainig code will be available. ;)