Why ChemSpider? Arbitrary. Done PubChem in the past already. Moreover, ChemSpider has the largest database of molecular structures and in that sense important to my research.

Why KNIME? Played with Taverna in the past, and expect to do much more work on Taverna in the coming year (see also this and this ). Moreover, KNIME got a CDK plugin already, and the KNIME developers contributed valuable feedback to the CDK project in the last year. It was about time that I contributed something back, though the current functionality is quite limited. KNIME has a better architectural design than Taverna1, but will face though competition with Taverna2, due next year.

The presentations

Heringa gave a presentation on network analysis, and discussed the scale-free network, hub nodes, etc, after which he gave an example on the 14-3-3 PPI family which both have promoting and inhibiting capabilities. Fraser presented work on improving microarray data analysis, by reducing non-random background noise. Schroeter presented the use of Gaussian process modeling in QSAR studies, which allows estimation of error bars (see DOI:10.1002/cmdc.200700041. I did not feel the results were very convincing, though, but the method sounds interesting. Larhlimi presented research on network analysis of metabolic networks. His approach finds so-called minimal forward direction cuts, which identifies critical parts in the network if one is interested in repressing certain metabolic processes. Hofto presented some work on the use of DFT for proteins, and picked up that one has to do things critically to be able to reproduce binding affinities. Combinations of DFT or MM with QM are becoming popular to model binding sites. Van Lenthe presented such an approach of the second day of CompLife.

By far the most interesting talk at the conference, was the insightful presentation by Paulien Hogeweg. She apparently coined the term bioinformatics. Anyway, she had a exciting presentation on feed-forward loops in relation to evolution, and showed correlation between jumps in FFL motifs with biodiversity. She also warned us for the Monster of Loch Ness syndrome, where computational models may indicate large underlying processes, which are not really existing. But that should be a problem that most of my readers should be aware of. She introduced evolutionary modeling, to put further restrictions on the models, to reduce the chance of finding monsters.

Hussong had an interesting presentation too, if one is interested in analysis of GC/MS or LC/MS data. He introduced a hard-modeling approach for proteomics data using wavelets technology. His angle on this was to use a wavelet that represents the isotopic pattern of a protein mass spectrum. Interestingly, the wavelet had negative intensities, something which one will never find in mass spectra. However, I seem to recall a mathematical restriction on wavelets that would forbid taking the squared version of the function. He indicated that the code is available via OpenMS.

Jensen, finally, presented his work at the UCC on Markov models for protein folding, where he uses the mean first passage time as observable to analyze of processes in folding state space. This allows him to compare different modeling approaches and, for example, to predict how many time steps are needed to reach folding. Being able to measure characteristics of certain modeling methods, one is able to make a objective comparison. Something which allows a fair competition .

The Monday/Tuesday after that, I will present CDK-Taverna integration I worked on in 2005 (see e.g. Taverna on Classpath and CDK-Taverna fully recognized ) at the Taverna meeting, before Thomas continued on that, leading to the cdk-taverna.de plugin website. If time permits, I will prepare an example workflow from metabolomics. Unlike previous times I went to Cambridgeshire, I won’t fly in on Stansted, but take the EuroStar instead. I am very much looking forward to that. Unfortunately, I will not have time to visit Cambridge itself, this time :(

Today, Miguel Rojas and Stefan Kuhn discussed their research. Miguel showed the state of mass spectrum prediction using the CDK and the MEDEA plugin for Bioclipse. Stefan demonstrated the NMRShiftDB and a new lab systems for NMR experiment scheduling and management system based on that. Dr. Ott (Nijmegen, Netherlands) showed the BioMeta Database which contains metabolite and reaction information derived from the KEGG, but which fixes a set of chemical problems in the latter (see also the article, DOI:10.1186/1471-2105-7-517).

The afternoons of CDK workshops traditionally have discussion sessions and hackathons. Two groups were formed: one consisted of the KNIME guys who, together with Miguel and Federico focused in QSAR descriptor calculations in KNIME, while Stefan, Martin and me looked at the fingerprinter peculiarities that Martin found (see also this CDK News article), and came up with a possible further performance improvement of the AllRingsFinder. Because one class of molecules that is causing trouble consist of two ring systems connected by a long linker, like Choloyl-CoA (below), we anticipate that splitting the molecule up into ring systems prior to using the SSSR algorithm should speed up the complete all-ring finding process.

Currently, the spanning tree is calculated before deciding on using the SSSR finder, which, we think, can be used to partition the molecule into separate ring systems. On each of them, then, the further steps of the ring search can be applied.

After dinner (pasta/pizza), during the Spanish-German handball game, we continued the hacking and discussions, now focusing as a whole group on QSAR descriptors in KNIME. We looked at each descriptor and decided if it should go into a QSAR calculator node, or even in a node of its own.

Bugs found

I won’t close this blog entry without giving a list of problems we found in the current CDK; some minor and small, some more troublesome. Here goes: typos all over the place; the OrderQueryBond lack a return statement in an else clause; the Mol2Reader does not mark atom and bond aromaticity properly and reads a single bond as aromatic, and an aromatic bond as single; the Renderer2D does not always highlight both atoms when hovering over a bond; SmilesGenerator.parseBond() should output bond orders correctly; the SSSR finder seems to have a messed up if-else statement for the ringBondCount limit of 37; the BondCount descriptor should count all bonds by default, not just the single bonds; IDescriptor.getParameters() should return null instead of Object[0]; several programs use the SYBYL atomtype S.o2, while the specification and the CDK config defines S.O2; the IP descriptor now returns a variable length descriptor.

One research presentation compared a number of fingerprint implementations in a QSAR study, and CDK came out very well, beating a few commercial programs. The free software session was full of CDK, however, with AMBIT, iBabel, Bioclipse and KNIME mentioning the CDK.

The latter is really interesting: it’s a workflow program just like Taverna or PipeLine Pilot, which is using the Eclipse RCP as starting point, just like Bioclipse. And like the other two, KNIME has CDK integration, at least for displaying structures.