1. Use of Machine Learning in Chemoinformatics
Irene Kouskoumvekaki
Associate Professor
February 15th, 2013

2. Major Aspects of Chemoinformatics
Databases: Development of databases for storage and retrieval of small molecule structures and their properties.
Machine learning: Training of Decision Trees, Neural Networks, Self Organizing Maps, etc. on molecular data.
Predictions: Molecular properties relevant to drugs, virtual screening of chemical libraries, system chemical biology networks…

3. Machine Learning
Tries to teach the computer to draw conclusions based on previous experience.

4. Akinator, the Web Genius Akinator the Genius can read your mind and tell you who you're thinking of by answering a few questions.

18. Machine learning classifiers

19. Clustering: Self Organizing Maps
Distinguishing molecules of different biological activities and finding a new lead structure

20. Clustering: Self Organizing Maps
Distinguishing molecules of different biological activities and finding a new lead structure

21. Clustering: Self Organizing Maps
Distinguishing molecules of different biological activities and finding a new lead structure

22. Clustering: Self Organizing Maps
Distinguishing molecules of different biological activities and finding a new lead structure

23. Machine Learning

24. Machine Learning Molecular Structures Properties Molecular Descriptors
QSAR
Virtual Screening
Clustering
Classification
Molecular
Structures
Properties
Molecular Descriptors

25. Different descriptor types
• Simple feature counts (such as number of rotatable bonds or molecular weight)
• Fragmental descriptors which indicate the presence or absence (or count) of groups of atoms and substructures
• Physicochemical properties (density, solubility, vdWaals volume)
• Topological indices (size, branching, overall shape)

26. Major Aspects of Chemoinformatics
Databases: Development of databases for storage and retrieval of small molecule structures and their properties.
Machine learning: Training of Decision Trees, Neural Networks, Self Organizing Maps, etc. on molecular data.
Predictions: Molecular properties relevant to drugs, virtual screening of chemical libraries, system chemical biology networks…

27. Quantitative Structure-Activity Relationships (QSAR)
In QSAR models structural parameters (descriptors) are fitted to experimental data for biological activity (or another given property, P)

28. Prediction of Solubility, ADME & Toxicity
Give guidelines
Filter compounds that enter the lab
Speed up the drug discovery process

29. Virtual screening
Computational techniques for a rapid assessment of large libraries of chemical structures in order to guide the selection of likely drug candidates.

30. Similarity Search
Similar Property Principle – Molecules having similar structures and properties are expected to exhibit similar biological activity.
Thus, molecules that are located closely together in the chemical space are often considered to be functionally related.

31. Fingerprints-based Similarity Search
widely used similarity search tool
consists of descriptors encoded as bit strings
Bit strings of query and database are compared using similarity metric such as Tanimoto coefficient
MACCS fingerprints: 166 structural keys
that answer questions of the type:
Is there a ring of size 4?
Is at least one F, Br, Cl, or I present?
where the answer is either
TRUE (1) or FALSE (0)

32. Tanimoto Similarity
or 90% similarity

33. Similarity Search

34. Example: Virtual Screening of PubChem
Go to
Search PubChem for compounds that are similar to this structure with Tc>0.95:
How many similar compounds do you find?
Click on BioActivity Analysis
How many of them are biologically active? On how many bioassays have they been tested on?
Click on Structure-Activity
Which compound is your query? Which compounds are most similar to your query? Are they active on the same bioassays?

40. Questions?