Friday, 13 April 2018

RDKit Reaction SMARTS

There's a been some papers using the RDKit for synthesis planning. If you're writing a paper and use the term "Reaction SMARTS" make sure you mean what everyone else thinks it means.

The SMILES, SMARTS, and SMIRKS line notations were created* by Daylight for storing, matching, and transforming connection tables.

  • SMILES describes a connection table to store molecule and reactions
  • SMARTS describes a pattern (or query) to match molecules and reactions
  • SMIRKS describes a transform (or "reaction") to modify molecules

RDKit uses the term "Reaction SMARTS" to mean "transform" (see RDKit Book). Unfortunately in Daylight's terminology Reaction SMARTS is a pattern not a transform.

Screenshot from the Daylight SMARTS theory manual.
Reactions SMARTS is primarily useful for searching reaction databases. For example this Reaction SMILES:


is matched by this Reaction SMARTS

[*:1][Nh2:2]>>[*:1][Nh0:2](~[OD1])~[OD1] amino to nitro

You can highlight the substructure:

Highlighting the SMARTS in the SMILES using CDK Depict

But that's a transform!

Yes but it's matching a transform (SMARTS) not applying one (SMIRKS), some may think you could read this unmodified as a SMIRKS but this is not the case. SMIRKS needs "real parts" after the second angled bracket as these are the parts created by the transform. Note that '*' is valid SMILES and in SMIRKS it kind of means "unmodified". This actually gives us the nice invariants:

All SMILES are valid SMARTS but not all SMARTS are valid SMILES
All SMIRKS are valid SMARTS but not all SMARTS are valid SMIRKS

Here is the SMIRKS transform for amino to nitro

[*:1][ND3:2]([H])([H])>>[*:1][N:2](=O)=O amino to nitro

In SMIRKS I can apply this SMIRKS to "molecules" and it will create "reactions". Note these molecules do not need to have atom-maps but they will come out with atom maps (see dt_transform)!


The output is


And another thing...

In general you can't run SMIRKS backwards. If I want to run a nitro to amino because the atoms/bonds we're adding need to be "real" we need to encode the reverse transform separately!

[*:1][ND3:2]([H])([H])>>[*:1][N:2](=O)=O amino to nitro
[*:1][ND3:2](~[OD1])(~[OD1])>>[*:1][N:2]([H])[H] nitro to amino

Although dt_transform specifies a direction this only controls whether the input molecules appear on the left or right of the output reaction.

*SMILES was created by Dave Weininger whilst at EPA

Saturday, 6 May 2017

Sharp Tools for Java Refactoring: Byte Code Analysis

I'm currently refactoring parts of the CDK core classes. As part of this I often need to find specific patterns/idioms that need to be changed. Whilst source code inspections and an IDE can make this task easy sometimes the tools aren't quite sharp enough.

I needed to find all occurrences of a reference (instead of value) comparison on a particular class. In Java there is no operator overload and so you can have situations like this:
Integer a = new Integer(25);
Integer b = new Integer(25);
if (a == b) {} // false
if (a.equals(b)) {} // true
I mentioned operating overloading but it's more subtle and is more about comparing reference vs. value comparison. In C/C++ we can have similar behaviour:
int aval = 25, bval = 25;
int *a = &aval;
int *b = &bval;
if (a == b) {} // false
if (*a == *b) {} // true
Most IDE's and code inspection programs will warn about common occurrences (for example Integer) but I wanted to find places where the CDK's classes were used like this. A simple text grep will find some but will have false positives and negatives requiring lots of manual checking. Daniel suggested the well known FindBugs might be able help.

Rather than analyze source code like PMD and Checkstyle, FindBugs analyses Java byte code with a set of defaults detectors to find often subtle but critical mistakes. FindBugs can be configured with custom detectors (see here), however the inspection I needed (RC: Suspicious reference comparison to constant) was almost there. After digging around in the source code I found you can provide a list of custom classes to detect. However it took a bit of trial and error to get what I needed.

First up we turn off all inspections except for the one we're looking for (we need to fix many others reported but I was looking for something specific). To do this we create an XML config that will only run the specific inspection (RC for Reference Comparison):
<?xml version="1.0" encoding="UTF-8"?>
    <Bug code="RC"/>
We then run findbugs with this configuration and provide the frc.suspicious property.
Running findbugs
$> findbugs -textui \
            -include findbugs-include.xml \
            -property "frc.suspicious=org.openscience.cdk.interfaces.IAtom" \
This produces an accurate report of all the places the references are compared. Here's a sample:
H C RC: Suspicious comparison of org.openscience.cdk.interfaces.IAtom references in org.openscience.cdk.Bond.getOther(IAtom)  At[line 253]
H C RC: Suspicious comparison of org.openscience.cdk.interfaces.IAtom references in org.openscience.cdk.Bond.getConnectedAtom(IAtom)  At[line 265]
H C RC: Suspicious comparison of org.openscience.cdk.interfaces.IAtom references in org.openscience.cdk.Bond.getConnectedAtoms(IAtom)  At[line 281]
H C RC: Suspicious comparison of org.openscience.cdk.interfaces.IAtom references in org.openscience.cdk.Bond.contains(IAtom)  At[line 300]

Monday, 3 April 2017

CDK AtomContainers are Slow - Let's fix that

The core class for molecule representation in CDK is the AtomContainer. The AtomContainer uses an edge-list data structure for storing the underlying connection table (see The Right Representation for the Job).

Essentially this edge-list representation is efficient in space. Atoms can be shared between and belong to multiple AtomContainers. Therefore querying connectivity (is this atom connected to this other atom) is linear time in the number of bonds.

The inefficiency of the AtomContainer can really sting. If someone was to describe Morgan's relaxation algorithm you may implement it like Code 1. The algorithm looks reasonable however it will run much slower than you expected. You may expect the runtime of this algorithm to be ~N2 but it's actually ~N3. I've annotated with XXX where the extra effort creeps in.
Code 1 - Naive Morgan-like Relaxation (AtomContainer/AtomIter)
// Step 1. Algorithm body
int[] prev = new int[mol.getAtomCount()];
int[] next = new int[mol.getAtomCount()];
for (int i = 0; i < mol.getAtomCount(); i++) {
  next[i] = prev[i] = mol.getAtom(i).getAtomicNumber();
for (int rep = 0; rep < mol.getAtomCount(); rep++) { // 0..numAtoms
  for (int j = 0; j < mol.getAtomCount(); j++) {     // 0..numAtoms
    IAtom atom = mol.getAtom(j);
    // XXX: linear traversal! 0..numBonds
    for (IBond bond : mol.getConnectedBondsList(atom)) {
      IAtom nbr = bond.getConnectedAtom(atom); 
      // XXX: linear traversal! 0..numAtoms avg=numAtoms/2
      next[j] += prev[mol.getAtomNumber(nbr)]; 
  System.arraycopy(next, 0, prev, 0, next.length);

A New Start: API Rewrite?

Ultimately to fix this problem correctly, would involve changing the core AtomContainer representation, unfortunately this would require an API change, optimally I think adding the constraint that atoms/bonds can not be in multiple molecules would be needed**. This would be a monumental change and not one I can stomach right now.

Existing Trade Off: The GraphUtil class

In 2013 I added the GraphUtil class for converting an AtomContainer to a more optimal adjacency list (int[][]) that was subsequently used to speed up many algorithms including: ring finding, canonicalisation, and substructure searching. Each time one of these algorithm is invoked with an IAtomContainer the first step is to build the adjacency list 2D array.

Code 2 - GraphUtil usage
IAtomContainer mol = ...;
int[][]        adj = GraphUtil.toAdjList(mol);

// optional with lookup map to bonds
EdgeToBondMap  e2b = EdgeToBondMap.withSpaceFor(mol);
int[][]        adj = GraphUtil.toAdjList(mol, e2b);

Although useful the usage of GraphUtil is somewhat clunky requiring passing around not just the adjacency list but the original molecule and the EdgeToBondMap if needed.
Code 3 - GraphUtil Depth First Traversal
void visit(IAtomContainer mol, int[][] adj, EdgeToBondMap bondmap, int beg, int prev) {
  mol.getAtom(beg).setFlag(CDKConstants.VISITED, true);
  for (int end : adjlist[beg]) {
    if (end == prev)
    if (!mol.getAtom(end).getFlag(CDKConstants.VISITED))
      visit(mol, adj, bondmap, end, beg);
      bondmap.get(beg, end).setIsInRing(true); // back edge

Using the GraphUtil approach has been successful but due to the clunky-ness I've not felt comfortable exposing the option of passing these through to public APIs. It was only ever meant as an internal optimisation to be hidden from the caller. Beyond causing unintentional poor performance (Code 1) what often happens in a workflow is GraphUtil is invoked multiple times. A typical use case would be matching multiple SMARTS against one AtomContainer.

A New Public API: Atom and Bond References

I wanted something nicer to work with and came up with the idea of using object composition to extend the existing Atom and Bond APIs with methods to improve performance and connectivity checks.

Essentially the idea is to provide two classes, and AtomRef and BondRef that reference a given atom or bond in a particular AtomContainer. An AtomRef knows about the original atom it's connected bonds and the index, the BondRef knows about the original bond, it's index and the AtomRef for the connected atoms. The majority of methods (e.g. setSymbol, setImplicitHydrogenCount, setBondOrder) are passed straight through to the original atom. Some methods (such as setAtom on IBond) are blocked as being unmodifiable.

Code 4 - AtomRef and BondRef structure
class AtomRef implements IAtom {
  IAtom         atm;
  int           idx;
  List<BondRef> bnds;

class BondRef implements IBond {
  IBond         bnd;
  int           idx;
  AtomRef       beg, end;

We can now re-write the Morgan-like relaxation (Code 1) using AtomRef and BondRef. The scaling of this algorithm is now ~N2 as you would expect.
Code 5 - Morgan-like Relaxation (AtomRef/AtomIter)
// Step 1. Initial up front conversion cost
AtomRef[] arefs = AtomRef.getAtomRefs(mol);

// Step 2. Algorithm body
int[]   prev  = new int[mol.getAtomCount()];
int[]   next  = new int[mol.getAtomCount()];
for (int i = 0; i < mol.getAtomCount(); i++) {
  next[i] = prev[i] = mol.getAtom(i).getAtomicNumber();
for (int rep = 0; rep < mol.getAtomCount(); rep++) {
  for (AtomRef aref : arefs) {
    int idx = aref.getIndex();
    for (BondRef bond : aref.getBonds()) {
      next[idx] += prev[bond.getConnectedAtom(aref).getIndex()];
  System.arraycopy(next, 0, prev, 0, next.length);

The depth first implementation also improves in readability and only requires two arguments.
Code 6 - AromRef Depth First (AtomRef/AtomFlags)
// Step 1. Initial up front conversion cost
void visit(AtomRef beg, BondRef prev) {
  beg.setFlag(CDKConstants.VISITED, true);
  for (BondRef bond : beg.getBonds()) {
    if (bond == prev)
    AtomRef nbr = bond.getConnectedAtom(beg);
    if (!nbr.getFlag(CDKConstants.VISITED))
      visit(nbr, bond);
      bond.setIsInRing(true); // back edge


I like the idea of exposing the AtomRef and BondRef to public APIs. I wanted to check that the trade-off in calculating and using the AtomRef/BondRef vs the current internal GraphUtil. To test this I wrote a benchmark that implements some variants of a Depth First Search and Morgan-like algorithms. I varied the algorithm implementations and whether I used, IAtomContainer, GraphUtil, or AtomRef.

The performance was measured over ChEMBL 22 and averaged the run time performance over 1/10th (167,839 records). You can find the code on GitHub ( Each algorithm computes a checksum to verify the same work is being done. Here are the raw results: depthfirst.tsv, and relaxation.tsv.

Depth First Traversal

A Depth first traversal is a linear time algorithm. I tested eight implementations that varied the graph data structure and whether I used an external visit array or atom flags to mark visited atoms. When looking just at initialisation time the AtomRef creation is about the same as GraphUtil. There was some variability between the different variants but I couldn't isolate where the different came from (maybe GC/JIT related). The runtime of the AtomRef was marginally slower than GraphUtil. Both were significantly faster (18-20x) than the AtomContainer to do the traversal. When we look at the total run-time (initialisation+traversal) we see that even for a linear algorithm, the AtomRef (and GraphUtil) were ~3x faster. Including the EdgeToBondMap adds a significant penalty.

Graph Relaxation

A more interesting test is a Morgan-like relaxation, as a more expensive algorithm (N2) it should emphasise any difference between the AtomRef and GraphUtil. The variability in this algorithm is whether we relax over atoms (AtomIter - see Code 1/5) or bonds (BondIter). We see a huge variability in AtomContainer/AtomIter implementation. This is because the algorithm is more susceptible to difference in input (molecule) size.

Clearly the AtomContainer/AtomIter is really bad (~80x slower). Excluding this results shows that as expected the AtomRef/AtomIter is slower than the GraphUtil/AtomIter equivalent (~2x slower). However because the AtomRef has a richer syntax, we can do a trick with XOR number storage to improve performance or iterate over bonds (BondIter) giving like-for-like speeds.


The proposed AtomRef and BondRef provide a convenience API to use the CDK in a natural way with efficient connectivity access. The conversion to an AtomRef is efficient and provides a speedup even for linear algorithms. The encapsulation facilities the passing as a public API parameter, users will be able to compute it ahead of time and pass it along to multiple algorithms.

I'm somewhat tempted to provide an equivalent AtomContainerRef allowing a drop-in replacement for methods that take the IAtomContainer interface. It is technically possible to implement writes (e.g. delete bond) efficiently in which case it would no longer be a 'Ref'. Maybe I'll omit that functionality or use a better name?


  • ** My colleague Daniel Lowe notes that OPSIN allows atoms to be in multiple molecules and know about their neighbours but it's a bit of a fudge. It's certainly possible with some extra book keeping but prevents some other optimisations from being applied.

Sunday, 10 July 2016

Generic Structure Depiction

Last week I attended the Seventh Joint Sheffield Conference on Chemoinformatics. It was a great meeting with some cool science and attendees. I had the pleasure of chatting briefly with John Barnard who's contributed a lot to the representation, storage, and retrieval of generic (aka Markush) structures (see Torus, Digital Chemistry - now owned by Lhasa).

At NextMove we've been doing a bit on processing sketches from patents (see Sketchy Sketches). I learnt a few things about how generic structures are typically depicted I thought be interesting to share.

Substituent Variation (R groups)

The most common type of generic feature is substituent variation, colloquially known as R groups. The variation allows concise representation with an invariant/fixed part of a compound and variable/optional part.

wherein R denotes

That is: anisole, toluene, or ethylbenzene.

Substituent Labels

Multiple substituent labels may be distinguished by a number R1, R2, ... Rn. However in reality, any label can and will be used. This can be particularly confusing when they collide with elements, examples include: Ra (Radium), Rg (Roentgenium) B (Boron), D (Deuterium), Y (Yttrium), W (Tungsten). The distinction between the label Ra and Radium may be semantically captured by a format but lost in depiction.

To distinguish such labels we can style them differently. By using superscripting and italicizing the label the distinction becomes clear and also somewhat improves the aesthetics of numbered R groups. We avoid subscript due to ambiguities with stoichiometry, for example: –NR2.

Attachment Points

For substituents there are different notation options. In writing, radical nomenclature is used, for the above example we'd say: methyl-oxyl (-OMe), ethyl (-Et), or methyl (-Me). However this doesn't translate well to depictions: .

The CTfile actually does stores substituents this way and specifies the attachment point (APO) information separately.
  2  1  0  0  0  0            999 V2000
    1.9048   -0.0893    0.0000 O   0  0  0  0  0  0  0  0  0  0  0  0
    2.6192    0.3232    0.0000 C   0  0  0  0  0  0  0  0  0  0  0  0
  1  2  1  0  0  0  0
M  APO  1   1   1
  1  0  0  0  0  0            999 V2000
    1.9940   -1.2869    0.0000 C   0  0  0  0  0  0  0  0  0  0  0  0
M  APO  1   1   1
  2  1  0  0  0  0            999 V2000
    1.8750   -2.3286    0.0000 C   0  0  0  0  0  0  0  0  0  0  0  0
    2.5895   -1.9161    0.0000 C   0  0  0  0  0  0  0  0  0  0  0  0
  1  2  1  0  0  0  0
M  APO  1   1   1
Alternatively we may use a virtual or 'null' atom. We can convert to/from CTfile format although it's slightly easier to delete the null atom that add it on, due to coordinate generation. A disadvantage of this is the atom count isn't accurate, however the labelled group is also a type of null atom and already distorts the atom count. There are unfortunately different ways of depicting this null atom.
Don't use a dative bond style! You have to fudge the valences and just doesn't work, how would I show a double bond attachment?

The first time I'd encountered attachment points was in ChEBI where and R group means 'something attaches here' (CHEBI:58314, CHEBI:52569), whilst a 'star' label means 'attaches to something' (CHEBI:37807, CHEBI:32861). This actually a nice way of thinking about it, like two jigsaw pieces the asymmetry allows the substituent to connect to the labelled atom.

The 'star' atom used by ChEBI is tempting to use as there is a star atom in SMILES.

However a '*' in SMILES actually means 'unspecified atomic number', some toolkits impose additional semantics. ChemAxon reads a 'star' to mean 'any atom', whilst OEChem, Indigo, and OpenBabel actually read more like an R Group, with [*:1] and [*:2] being R1 and R2 etc. ChemAxon Extended SMILES allows us to explicitly encode attachment points.
*OC |$_AP1$|
*C |$_AP1$|
*CC |$_AP1$|
I opted to implement the wavy line notation in CDK which is preferred by IUPAC graphical representation guidelines.
A major disadvantage of this notation is mis-encoding by users mistaking it for a wavy up/down stereo bond. I talk more about this in the poster (Sketchy Sketches) but the number of times you see the following drawn:
The captured connection table for that sketch does not have null atoms but instead uses carbon:

Saturday, 21 November 2015

Bringing Molfile Sgroups to the CDK - Rendering Tips

In the last but one post I gave a demonstration of S(ubstance)group rendering in the CDK. Now I want to give some implementation insights.

Abbreviations (Superatoms)

Abbreviations contract part of a structure to a line formula or common abbreviation.

Abbreviating too much or using unfamiliar terms (e.g. maybe using CAR for carbazole) can make a depiction worse. However some structures, like CHEMBL590010, can be dramatically improved.
One way to implement abbreviations would be by modifying the molecule data structure with literal collapse/contract and expand operations. Whilst this approach is perfectly reasonable, deleting atoms/bonds is expensive (in most toolkits) and it somewhat subtracts the "display shortcut" nature of this Sgroup.
For efficiency abbreviations are implemented by hiding parts of the depictions and remapping symbols. Just before rendering we iterator over the Sgroups and set hint flags that these atoms/bonds should not be included in the final image. If there is one attachment (e.g. Phenyl) we remap the attach point symbol text to the abbreviation label ('C'->'Ph'). When there are no attachments (e.g. AlCl3) we add a new symbol to the centroid of the hidden atoms.
Hide atoms and bonds Symbol Remap Abbreviated Result
For two or more attachments (e.g. SO2) you also need coordinate remapping.

Multiple Group

Multiple groups allow, contraction of a discrete number of repeating units. They are handled similarly to the abbreviations except we don't need to remap parts.
All atoms are present in the data structure but are laid out on top of each other (demonstrated below). We have a list of parent atoms that form the repeat unit. Therefore to display multiple groups we hide all atoms and bonds in the Sgroup except for parent atoms and the crossing bonds.
It's worth mentioning that hidden symbols are still generated but simply excluded from the final result. This allows bond back off for hetero atoms to be calculated correctly as is seen in this somewhat tangled example:


Polymer and Multiple group Sgroups require rendering of brackets. Encoded in the molfile (and when laid out) brackets are described by two points, a line. It is therefore up to the renderer to decide which side of the line the tick marks should face.
I've seen some implementations use the order of the points to convey bracket direction. Another method would be to point the brackets at each other. As shown for CHEBI:59342 this is not always correct.

Poor bracket direction Preferred bracket direction
I originally thought the solution might involve a graph traversal/flood-fill but it turns out there is a very easy way to get the direction correct. First we consider that brackets may or may not be placed on bonds, if a bracket is on a bond this information is available (crossing bonds).
  • For a bracket on a crossing bond exactly one end atom will be contained in the Sgroup, the bracket should point towards this atom.
  • If a bracket doesn't cross a bond then the direction should point to the centroid of all atoms in the Sgroup.