Participant:	Vaniya
Authors:	Vaniya, Arpana [1], Stephanie N. Samra [1], Sajjan S. Mehta [1], 
		Diego Pedrosa [1], Hiroshi Tsugawa [2], and Oliver Fiehn [1]
Affiliations: 	[1] Genome Center, University of California, Davis 
		[2] RIKEN Center for Sustainable Resource Science (CSRS), Wako, Japan

ParticipantID:	avaniya003
Category:	Category 2
Automatic methods: Yes

Abstract: 

MS-FINDER developed by H.Tsugawa et al. was used as an in silico software for unknown
compound identification in Category 2. MS-FINDER version 1.62 was used. MS/MS spectra
were uploaded to MS-FINDER in msp format. Precursor m/z, ion mode, mass accuracy of
instrument, and precursor type were used as metadata. Each candidate file was
converted to a structure database file which can be read by MS-FINDER. Each file was
saved in the software folder in order for it to be called by MS-FINDER. This file was
changed after each calculation in order to match the challenge data. A search of the
challenge msp against the challenge candidate list was performed on the top 500
candidates. Up to 500 top candidates structures were exported as a text file from
MS-FINDER. Final scores and SMILES were reported for submission to CASMI 2016.
Multiple candidates were submitted for each challenge.

Participant:	        Duehrkop
Authors: 		Dührkop, Kai (1) and Shen, Huibin (2) and Meusel, Marvin (1)
			and Rousu, Juho (2) and Böcker, Sebastian (1)
Affiliations:         	(1) Chair of Bioinformatics, Friedrich-Schiller University, Jena
			(2) Department of Computer Science, Aalto University

ParticipantID:	      csifingerid
Category:	      category2
Automatic pipeline:   yes
Spectral libraries:   yes (for training)

Abstract

We processed the peaklists in MGF format using a command line version
of CSI:FingerId 1.0.1. Fragmentation trees were computed with Sirius 3.1.4 
using the Q-TOF instrument settings. We computed trees for all
candidate formulas in the given structure candidate list.  Only the
top scoring trees were selected for further processing: Trees with a
score smaller than 80% of the score of the optimal tree were
discarded. Each of these trees was processed with CSI:FingerId as
described in [1]. We predicted for each tree a molecular fingerprint
(with platt probability estimates) and compared them against the
fingerprints of all structure candidates with the same molecular
formula. The resulting hits were merged together in one list and were
sorted by score. A constant value of 10000 was added to all scores to
make them positive (as stated in the CASMI rules). Ties of compounds
with same score (and sometimes also with same 2D structure) were
ordered randomly.

The machine learning method was trained on 7352 spectra (4564
compounds) downloaded from GNPS [2] and Massbank [3]. As our training
dataset contains only spectra in positive ion mode (there are too few
spectra with negative ion mode in GNPS), we ommited all challenges
with negative ion mode; As long as there are not enough public
available reference spectra measured in negative ion mode our method
will be only able to process positive ion mode spectra.

We observed for 67 challenges that the top scoring structure candidate
was a compound which is also contained in our training set. If we
evaluate our method on spectra from compounds we have already trained
on we usually reach a performance comparable to spectral library
search. To avoid an overestimation of the performance of our method,
we removed all of these top scoring candidates from our training set
and retrained our classifiers. To compensate the removed spectra, we
added the training spectra that are provided by CASMI. The submission
with the ParticipantID csifingerid_leaveout contains the search
results of these newly trained classifiers.

[1] Kai Dührkop, Huibin Shen, Marvin Meusel, Juho Rousu and Sebastian
    Böcker Searching molecular structure databases with tandem mass
    spectra using CSI:FingerID.  Proc Natl Acad Sci U S A,
    112(41):12580-12585, 2015.

[2] https://gnps.ucsd.edu

[3] Horai H, et al. MassBank: a public repository for sharing mass 
    spectral data for life sciences.  J Mass Spectrom 45(7):703–714, 2010.

Participant:		Verdegem
Authors:		Verdegem, Dries and Ghesquière, Bart
Affiliation:		Vesalius Research Center, VIB/KULeuven, Leuven, Belgium

ParticipantID:		dverdegem
Category:		category2
Automatic method:	yes

Abstract
For all assignments, we used the MAGMa+ software [1].

MAGMa+ uses MAGMa [2] under the hood. It runs MAGMa twice with two
different, fine-tuned parameters of which the values depend on the
ionization mode. MAGMa+ then determines the molecular class of the top
ranked metabolites returned by both MAGMa runs. This latent molecular
class is determined by a trained two-class random forest
classifier. Depending on the most prevelant molecular class, one of
both MAGMa outcomes (the one from the run with the parameters
corresponding to the most prevelant class) is returned to the user.

As structure database, the possible solution list provided in the
contest was used. We did not perform any prefiltering.

[1] Verdegem, Dries, et al. "Improved metabolite identification with
    MIDAS and MAGMa through MS/MS spectral dataset-driven parameter
    optimization." accepted for publication in Metabolomics
[2] Ridder, Lars, et al. "Substructure‐based annotation of
    high‐resolution multistage MSn spectral trees." Rapid
    Communications in Mass Spectrometry 26.20 (2012): 2461-2471.

Participant:          Allen
Authors:              Felicity Allen, Tanvir Sajed, Russ Greiner, David Wishart
Affiliations:         Department of Computing Science
		      University of Alberta, Canada

ParticipantID:        FelicityAllenCFMOrig
Category:             category2
Automatic pipeline:   yes
Spectral libraries:   no

Abstract

We processed the list of molecules and provided candidates using cfm-id.
The original  CFM positive and negative models were used, which were trained 
on data from the Metlin database.  Mass tolerances of 10ppm were used
and the Jaccard score was applied for spectral comparisons. The input spectrum
was repeated for the low, medium and high energies.

Participant:	      Brouard
Authors:              Céline Brouard(1,2), Huibin Shen(1,2), Kai Dührkop(3), 
		      Sebastian Böcker(3) and Juho Rousu(1,2)
Affiliations:         (1) Department of Computer Science, Aalto University, Espoo, Finland
                      (2) Helsinki Institute for Information Technology, Espoo, Finland
                      (3) Chair for Bioinformatics, Friedrich-Schiller University, 
		      Jena, Germany

ParticipantID:        IOKRAlignf
Category:	      category2
Automatic pipeline:   yes
Spectral libraries:   no

Abstract

We used a recent machine learning approach, called Input Output Kernel
Regression, for predicting the candidate scores. In this method, the
similarities between the MS/MS spectra and the molecular similarities
are encoded using two kernel functions. In input, we computed
different kernels based on MS/MS spectra and on fragmentation
trees. In output we built a gaussian kernel based on molecular
fingerprints. We used approximately 6000 molecular fingerprints from
OpenBabel. We combined the different input kernels using the Alignf
algorithm, which searches to maximize the alignment between the
combined kernel the output kernel.

We trained separate models for the MS/MS spectra in positive mode and
the MS/MS spectra in negative mode.  We considered additional MS/MS
spectra from GNPS and MassBank for training the models.