Machine learning in the life sciences with knime

Machine Learning in the Life Sciences... with KNIME!

Gregory Landrum NIBR Informatics

Novartis Institutes for BioMedical Research, Basel

Cartoon machine learning

Training Data

Training Model

New Items

Model Predictions

Training a model:

Using a model:

The data introducing vocabulary

Descriptors End point

A typical life-sciences problem

Training Data

Training Model

New Items

Model Predictions

Training a model:

Using a model:

Literature molecules active for an interesting

protein target

New molecules we are thinking about making.

Prioritized list

A problem... Here’s what our input looks like:

All data taken from ChEMBL (https://www.ebi.ac.uk/chembl/)

Good luck training a model with that!

One solution: Molecular Fingerprints

§  Idea : Apply a kernel to a molecule to generate a bit vector or count vector (less frequent)

§  Typical kernels extract features of the molecule, hash them, and use the hash to determine bits that should be set

§  Typical fingerprint sizes: 1K-4K bits.

...

The toolbox: Knime + the RDKit §  Open-source RDKit-based nodes for Knime providing cheminformatics

functionality

+

§  Trusted nodes distributed from knime community site

§  Work in progress: more nodes being added (new wizard makes it easy)

What’s there?

Let’s build a model! Step 1, getting the data ready

Detail: we’re using atom-pair fingerprints

100 actives

~83K assumed inactives

Detail: we’re using Histamine H3 actives

Let’s build a model! Step 2, training

For this example I use 70% of the data (randomly selected) to train the model

Detail: the model is a depth-limited random forest with 500 trees

Let’s build a model! Step 3, testing

Test with the 30% of the data that was not used to build the model

The model is 99.9% accurate. Unfortunately it’s saying “inactive” almost all the time. This makes sense given how unbalanced the data is

Adjusting the model for highly unbalanced data Is there a signal there?

Test with the 30% of the data that was not used to build the model

Obviously a strong signal there, we just need to figure out how to use it.

Adjusting the model for highly unbalanced data Is there a signal there?

Test with the 30% of the data that was not used to build the model

Obviously a strong signal there, we just need to figure out how to use it. How about changing the decision boundary?

Find the model score that corresponds to this point in the ROC curve for the training data

Adjusting the model for highly unbalanced data Shifting the decision boundary

Set decision boundary here

Now we’ve got a >99% accurate model that does a good job of retrieving actives without mixing in too many inactives.

Training data ROC

Wrapping up

§  We were able to build very accurate random forests for predicting biological activity by adjusting the decision boundary for models built using highly unbalanced data

§  The same thing works with the Knime “Fingerprint Bayesian” nodes.

§  Acknowledgements: • Manuel Schwarze (NIBR) •  Sereina Riniker (NIBR) •  Nikolas Fechner (NIBR) •  Bernd Wiswedel (Knime) •  Dean Abbott (Abbott Analytics)

Advertising

3rd RDKit User Group Meeting 22-24 October 2014

Merck KGaA, Darmstadt, Germany

Talks, “talktorials”, lightning talks, social activities, and a hackathon on the 24th. Announcement and (free) registration links at www.rdkit.org We’re looking for speakers. Please contact [email protected]