RDKit (cheminformatics) Neo4j Integration Mentors: Christian - - PowerPoint PPT Presentation

RDKit (cheminformatics) Neo4j Integration

Mentors: Christian Pilger (BASF) Greg Landrum (RDKit) Stefan Armbruster (Neo4j) Presenter - Evgeny Sorokin

Motivation

• Neo4j = useful tool to map knowledge
• Chemical/pharmaceutical R&D:

○ Required: mapping data of completely different nature (recipe, process, application test, chemical structures) ○ Knowledge graphs are frequently a good choice here over other data models ○ Problem: Neo4j does not support chemical structures

• RDKit

○ is a widely used Open Source tool to deal with chemical structures ○ has proven its value in conjunction with Postgres

• Idea: enrich Neo4j's capabilities by combining it with RDKit => GSoC project

Chemical structure representation

• Not intended: “dissolve” atoms and bond as nodes and relations into the graph!
• Intended: use available structure representation as node properties

○ SMILES format: c1ccccc1 (single line ASCII representation --> exact search via string matching) ○ MOL format: (3D coordinates: richer format, more details --> advantages in sub-structure searches) name: benzene formula: C6H6 SMILES: c1ccccc1

Chemical structures example

name: benzene formula: C6H6 SMILES: c1ccccc1

Requirements

• Basic Functionality :

○ Exact chemical search (“find the molecule benzene") ○ Chemical substructure search (“find all molecules that contain a benzene moiety")

• Typical application scenarios in Graph context

○ Find entry points into the graph ○ Filter paths during graph traversal with chemical structure conditions

How was it implemented - storage in a graph

• A new node with labels :Chemical:Structure is processed by RDKit event handler
• From either smiles or mdlmol property a list of 7-8 properties is created per node
• A full text index is created for fingerprint property
• canonical_smiles
• inchi
• formula
• molecular_weight
• fp - bit-vector fingerprint
• fp_ones - count of positive bits
• mdlmol
• smiles [optional]

How was it implemented - exact search

Simple case: compare two canonical smiles with each other, find a match. SMILES O=S(=O)(Cc1ccccc1)CS(=O)(=O)Cc1ccccc1 Canonical SMILES O=S(=O)(CC1=CC=CC=C1)CS(=O)(=O)CC1=CC=CC=C1

How was it implemented - SSS

Chemical fingerprint is a unique pattern for the presence of a particular molecule. Bitvector and count-based fingerprints

How was it implemented - SSS

1. Each of the structures is encoded as bitvector fingerprint 2. Bitvectors are transformed into a string of positive indexes 3. Fulltext index is applied to transformed bitvectors (numbers -> words) 4. Search is done with constraints regarding specific properties.

How was it implemented - SSS

Chemical reactions’ relationships

What are possible applications

2.) expand path apoc.path.expand

Resources

• https://github.com/rdkit/neo4j-rdkit
• https://www.rdkit.org/UGM/2012/Landrum_RDKit_UGM.Fingerprints.Final.pptx.pdf
• https://www.rdkit.org/
• https://neo4j.com/docs/cypher-manual/current/schema/index/
• http://tiny.cc/mol_block_definition
• @evgerher via telegram

Hunger games Q&A

1) Hard: what format can resolve a situation when chemical structure has chirality property (ex.: Lactic acid)?

a) SMILES b) MOL block c) All of above

2) Medium: what is the difference between bitvector and count-based fingerprints?

a) Harder to store b) Does not support similarity search c) Does not keep track of occurence amount

3) Easy: transformation of bitvector [1 0 1 0 1 1 0 0 1] is:

a) “1 3 5 6 9” b) “2 4 6 7 9” c) “1 3 5 6 9”