The Fragment Network: A Chemistry Recommendation Engine Built Using a Graph Database

Transcription

1 Supporting Information The Fragment Network: A Chemistry Recommendation Engine Built Using a Graph Database Richard J. Hall, Christopher W. Murray and Marcel L. Verdonk. Contents This supporting information contains a detailed description of the algorithm to generate the nodes and edges of the fragment network. The document also contains details on inserting nodes and edges into a Neo4j graph database and some example network queries. The additional supporting information contains data that can be inserted into a Neo4j graph database to create an illustrative Fragment Network that is based around the 4-hydroxy-biphenyl example presented in Figure 2 of the manuscript. S1

2 Generating Nodes and Edges for the Fragment Network To generate a set of nodes and edges for a molecule, we first generate a node that represents the molecule itself. The molecule is represented as a nonisomeric smiles string. As well as the smiles string, the heavy atom count (HAC) and ring atom count (RAC) are stored as node attributes, as is a simplified representation of the molecular graph. This simplified graph representation uses the daylight function dt_molgraph to set all bond orders to one, remove aromaticity, set the hydrogen count, remove charges and set masses to zero. Additionally we set the element type of every ring atom to carbon. For 4-hydroxy-biphenyl, this first node has the following data {SMILES: "Oc1ccc(cc1)c2ccccc2", HAC: 13, RAC: 12 RINGSMILES: "OC1CCC(CC1)C2CCCCC2"} (1) Next, a set of molecular components is generated. The SMARTS pattern "[*;R]-;!@[*]" is used to find single acyclic bonds to a ring atom. For each matching bond b, the start and end atom are determined. The bond b is deleted and isotopically labelled xenon atoms are bonded to the start and end atom. The isotopic label is incremented for each matching bond, to provide unique labelling. The components for 4-hydroxy-biphenyl are ['O[100Xe]', '[101Xe]c1ccccc1', '[100Xe]c1ccc([101Xe])cc1'] (2) The bond between the hydroxyl and the phenyl ring is broken and the oxygen and carbon atoms are attached to xenon atoms with the isotopic label 100. The bond between the phenyl rings is broken and the carbon atoms are attached to xenon atoms with the isotopic label 101. Next, this set of molecular components is combined to generate child nodes. Each child node is created by leaving out one of the components. For example, leaving out the O[100Xe] component from (2), we create a child node by considering the two phenyl ring components. These components are combined by locating pairs of matching xenon atom labels. The two phenyl ring components contain one pair of 101Xe labels, as well as an unmatched 100Xe label. A bond is added between the atoms attached to the paired xenon atoms. The bonds to the paired xenon atoms and the xenon atoms themselves are then deleted. The rebuilt molecule will contain the unmatched xenon atoms that indicate the attachment point(s) of the excluded component. In the example there is a single xenon atom that shows the location of the excluded hydroxyl component. If the excluded component is a ring or linker the rebuilt molecule will contain multiple xenon atoms and will be disconnected. We generate a smiles string for the child node by replacing the remaining xenon atom with hydrogen. For the components in (2), the combinations are given in table S1. Exclude Leaving Rebuilt As Child Node [100Xe]c1ccc([101Xe])cc1 ['O[100Xe]', '[101Xe]c1ccccc1'] O[Xe].[Xe]c1ccccc1 O.c1ccccc1 [101Xe]c1ccccc1' ['O[100Xe]', Oc1ccc([Xe])cc1 Oc1ccccc1 '[100Xe]c1ccc([101Xe])cc1'] O[100Xe] ['[101Xe]c1ccccc1', '[100Xe]c1ccc([101Xe])cc1'] [Xe]c1ccc(cc1)c2ccccc2 c1ccc(cc1)c2ccccc2 Table S1. The set of child nodes created from components of 4-hydroxy-biphenyl An edge is created to join each child node to the parent. Each edge is labelled with a set of attributes relating to the parent and child nodes. For example, the edge that links 4-hydroxybiphenyl to biphenyl has the following attributes the type of the excluded component: 'FG' (we used to refer to substituents as functional groups) the type of the rebuilt combination: 'RING' (unused) the nonisomeric smiles of the excluded component: O[Xe] S2

3 the nonisomeric smiles of the rebuilt molecule: [Xe]c1ccc(cc1)c2ccccc2 the simplified graph of the excluded component: O[Xe] the simplified graph of the rebuilt molecule: [Xe]C1CCC(CC1)C2CCCCC2 As with the node attribute, the simplified graph representation for the edge uses the daylight function dt_molgraph to set all bond orders to one, remove aromaticity, set the hydrogen count, remove charges and set masses to zero. Additionally we set every ring atom element to carbon. This means that our simplified graph will be the same for all rings of the same size, irrespective of the aromaticity or the heteroatomic composition of the original component. We can also define equivalencies between rings of different type using our simplified graph representation. This allows us to treat eg a 1,4 substituted 6 membered ring as equivalent to a 1,3 substituted 5 membered ring. These equivalencies are derived from a set of common ring scaffolds in the Astex registry and ChEMBL. When we search the graph, if the excluded component is marked as type RING, we can use our ring equivalence dictionary to match rings of different sizes. For compounds with more than two substituents, the order of the substituents in the ring equivalence dictionary is relevant. For this reason, we add canonicalized isotopic labels to the simplified graph. The canonicalization uses the lexicographical ordering of the smiles pattern for the excluded components. Figure S1 illustrates this. All compounds have common substituents, methyl, fluoro and chloro. Removing the ring system from compounds a-d means that all compounds will have an edge that joins to the node C.Cl.F. Compounds a and b share the same simplified graph for the excluded ring component and can be grouped as replace ring, equivalent vectors. Compound c has a different arrangement of substituents and will have a different simplified graph. This means compound c will be placed in the replace ring, different vectors group. Compound d has a different ring system but the same arrangement of substituents. We can therefore add an equivalency rule between the simplified graph of ring d and the simplified graph of rings a and b that will allow compound d to also be classified as replace ring, equivalent vectors. (a) (b) (c) (d) Figure S1. Related compounds with equivalent and non-equivalent simplified graphs Once the edge data has been recorded, each new node is itself used as input for the algorithm to recursively create additional nodes and edges. Note that the algorithm keeps track of the set of nodes that have been generated. If a node is already in the set the algorithm returns immediately without recomputing child nodes and edges (since these will have been generated in a prior iteration). This provides a significant reduction in the amount of work required to build nodes and edges for a large database. We apply additional rules to deal with the special case of ring-ring bonds and spiro systems. For each ring ring bond in the compound, we add xenon to the bond atoms and delete the bond. For 4- hydroxy-biphenyl, we generate Oc1ccc([Xe])cc1.[Xe]c1ccccc1 S3

4 The xenon atoms are replaced with hydrogen to give the child node Oc1ccccc1.c1ccccc1 An edge between this child node and the parent is added with the following attributes: the type of the excluded component: 'FG' the type of the rebuilt combination: 'RING' (unused) the nonisomeric smiles of the excluded component: [Xe] (indicates a zero length linker) the nonisomeric smiles of the rebuilt molecule: Oc1ccccc1.c1ccccc1 the simplified graph of the excluded component: [Xe] the simplified graph of the rebuilt molecule: OC1CCCCC1.C1CCCCC1 The child node is then passed into the node and edge generating algorithm. A similar approach is used to break spiro ring systems. One can envisage how the algorithm might be extended to deconstruct fused ring systems, although we have not chosen to do so in this implementation. Another suggested enhancement is to treat cyclopropyl rings as both a ring and a substituent. By modifying the smarts pattern that finds ring bonds we could exclude cuts at a cyclopropyl ring. An attribute label will be added to the Oc1ccc(cc1)c2ccccc2 node, to indicate that this compound has the identifier in the EM (emolecules) database. Additional attributes are added to this node, for example a label to mark this node as ChEMBL record ATTR Oc1ccc(cc1)c2ccccc2 EM ATTR Oc1ccc(cc1)c2ccccc2 CHEMBL A complete set of nodes and edges and attributes for 4-hydroxy-biphenyl is listed below. NODE Oc1ccc(cc1)c2ccccc OC1CCC(CC1)C2CCCCC2 0 NODE O.c1ccccc1 7 6 O.C1CCCCC1 1 NODE O 1 0 O 2 EDGE O.c1ccccc1 O RING c1ccccc1 C1CCCCC1 FG O O NODE c1ccccc1 6 6 C1CCCCC1 2 EDGE O.c1ccccc1 c1ccccc1 FG O O RING c1ccccc1 C1CCCCC1 EDGE Oc1ccc(cc1)c2ccccc2 O.c1ccccc1 RING [Xe]c1ccc([Xe])cc1 [100Xe]C1CCC([101Xe])CC1 RING O[Xe].[Xe]c1ccccc1 O[100Xe].[Xe]C1CCCCC1 NODE Oc1ccccc1 7 6 OC1CCCCC1 1 EDGE Oc1ccccc1 O RING [Xe]c1ccccc1 [100Xe]C1CCCCC1 FG O[Xe] O[Xe] EDGE Oc1ccccc1 c1ccccc1 FG O[Xe] O[Xe] RING [Xe]c1ccccc1 [100Xe]C1CCCCC1 EDGE Oc1ccc(cc1)c2ccccc2 Oc1ccccc1 RING [Xe]c1ccccc1 [100Xe]C1CCCCC1 RING Oc1ccc([Xe])cc1 OC1CCC([Xe])CC1 NODE c1ccc(cc1)c2ccccc C1CCC(CC1)C2CCCCC2 1 EDGE c1ccc(cc1)c2ccccc2 c1ccccc1 FG [Xe] [Xe] RING [Xe]c1ccccc1 [100Xe]C1CCCCC1 EDGE c1ccc(cc1)c2ccccc2 c1ccccc1 FG [Xe] [Xe] RING [Xe]c1ccccc1 [100Xe]C1CCCCC1 NODE c1ccccc1.c1ccccc C1CCCCC1.C1CCCCC1 2 EDGE c1ccccc1.c1ccccc1 c1ccccc1 FG [Xe] [Xe] RING c1ccccc1 C1CCCCC1 EDGE c1ccccc1.c1ccccc1 c1ccccc1 FG [Xe] [Xe] RING c1ccccc1 C1CCCCC1 EDGE c1ccc(cc1)c2ccccc2 c1ccccc1.c1ccccc1 FG [Xe] [Xe] RING c1ccccc1.c1ccccc1 C1CCCCC1.C1CCCCC1 EDGE Oc1ccc(cc1)c2ccccc2 c1ccc(cc1)c2ccccc2 FG O[Xe] O[Xe] RING [Xe]c1ccc(cc1)c2ccccc2 [100Xe]C1CCC(CC1)C2CCCCC2 NODE Oc1ccccc1.c1ccccc OC1CCCCC1.C1CCCCC1 1 EDGE Oc1ccccc1.c1ccccc1 Oc1ccccc1 RING c1ccccc1 C1CCCCC1 RING Oc1ccccc1 OC1CCCCC1 EDGE Oc1ccccc1.c1ccccc1 O.c1ccccc1 RING [Xe]c1ccccc1 [100Xe]C1CCCCC1 RING O[Xe].c1ccccc1 O[100Xe].C1CCCCC1 EDGE Oc1ccccc1.c1ccccc1 c1ccccc1.c1ccccc1 FG O[Xe] O[Xe] RING [Xe]c1ccccc1.c1ccccc1 [100Xe]C1CCCCC1.C1CCCCC1 S4

5 EDGE Oc1ccc(cc1)c2ccccc2 Oc1ccccc1.c1ccccc1 FG [Xe] [Xe] RING Oc1ccccc1.c1ccccc1 OC1CCCCC1.C1CCCCC1 ATTR Oc1ccc(cc1)c2ccccc2 EM Loading Data into Neo4j A set of nodes, edges and attributes for the compounds that are listed as results in Figure 2 can be found in the supporting information files jm7b00809_si_002.txt (nodes), jm7b00809_si_003.txt (edges), jm7b00809_si_004.txt (attributes). These data files can be loaded into a Neo4j graph database (our implementation uses Neo4j vesion 2.1.6). Neo4j has a SQL like syntax called cypher. The following cypher commands can be used to load the nodes, edges and attribute into the graph database. USING PERIODIC COMMIT LOAD CSV FROM 'file:///jm7b00809_si_002.txt ' AS line FIELDTERMINATOR ' ' MERGE (:F2 { smiles: line[1], hac: toint(line[2]), chac: toint(line[3]), osmiles: line[4]}); USING PERIODIC COMMIT LOAD CSV FROM 'file:///jm7b00809_si_003.txt ' AS line FIELDTERMINATOR ' ' MATCH (n1:f2 { smiles: line[1]}), (n2:f2 { smiles: line[2]}) MERGE (n1)-[:f2edge{label:line[3]}]- >(n2); USING PERIODIC COMMIT LOAD CSV FROM 'file:///jm7b00809_si_004.txt ' AS line FIELDTERMINATOR ' ' MATCH (n:f2 { smiles: line[1]} ) set n:mol, n:em, n.em=toint(line[3]); Querying Neo4j The cypher query used to find and categorise 'medium' paths of length one from 4-hydroxy-biphenyl to a commercially available compound is match p = (n:f2{smiles:'oc1ccc(cc1)c2ccccc2'})-[nm]-(m:em) where abs(n.hac-m.hac) <= 3 and abs(n.chac-m.chac) <= 1 return split(nm.label, ' ')[4], split(nm.label, ' ')[1], nm.label, m.hac-n.hac, m.smiles, m.em order by split(nm.label, ' ')[4]; This query is somewhat complicated by the decision to store edge metadata as a single attribute. The metadata is pipe separated and needs to be split to construct the query. Anyone wishing to implement a similar network might choose to store each piece of edge metadata as a separate attribute. Our current implementation splits and groups the paths using python code after the paths are returned from Neo4j - the cypher query is for information only. By grouping the matches on the change in atom count and selected attributes of the edge label, we can classify the results as deletions or additions at a specific position. To sort these groups, we use the values from a lookup dictionary that is keyed on the substituent attribute of the edge label. Any substituent that is not in the dictionary will be given a weight of zero. Any ties are broken using the the heavy atom count and the lexicographical sort order of the smiles string. A similar cypher query can be used to find and categorise medium paths of length two between 4- hydroxy-biphenyl and a commercially available compound MATCH (sta:f2 {smiles:"oc1ccc(cc1)c2ccccc2"})-[n4:f2edge]-(n3:f2)-[n2:f2edge]-(end:em) where S5

6 abs(sta.hac-end.hac) <= 3 and abs(sta.chac-end.chac) <= 1 and sta.smiles <> end.smiles RETURN split(n4.label, ' ')[4], split(n4.label, ' ')[2], split(n2.label, ' ')[2], split(n2.label, ' ')[1], end.em, end.smiles order by split(n4.label, ' ')[4], split(n2.label, ' ')[2]; The first column in the output can be used to group similar transformations. The second and third columns can be used to compare equivalencies between simplified graphs. The fourth column lists the replacement. The fifth and sixth columns are the identifier and the smiles string of the related compound. Again the query is for illustration purposes; our implementation splits the edge metadata after the query results are returned. S6