Participant:          Bach
ParticipantID:        IOKR_TanimotoGaussian
Category:	          4
Authors:              Eric Bach(1), Céline Brouard(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
Automatic pipeline:   yes
Spectral libraries:   no

Abstract
We used a recent machine learning approach, called Input Output Kernel Regression 
(IOKR), for predicting the candidate scores. IOKR has been successfully applied 
to metabolite identification [1]. 

In this method kernel functions are used to measure the similarity between MS/MS
spectra (input kernel) respectively between molecular structures (output kernel).
On the input side, we use several kernels defined on MS/MS spectra and fragmentation
trees, and combine them uniformly, i.e. we sum up the kernels with equal weights.
On the output side, we use a Gaussian kernel on Tanimoto features calculated 
from binary fingerprints representing the molecular structures in the candidate 
sets.

We train two separated IOKR models one for each ionization mode, i.e. positive
and negative. For the positive model we use ~14000 identified MS/MS spectra and 
for the negative model ~5800. Those spectra mainly are extracted from the GNPS 
and MassBank databases. We represent the candidate molecular structures using 
~7600 binary molecular fingerprints. 

For each challenge spectra we predict the molecular formula using Sirius [2] by
taking into account the possible molecule formulas based on the candidate sets.
The score we submitted for each candidate is the one corresponding to the most
likely molecular formula.

[1] Brouard, Cé.; Shen, H.; Dührkop, K.; d'Alché Buc, F.; Böcker, S. & Rousu, J.
    Fast metabolite identification with Input Output Kernel Regression
    Bioinformatics, 2016
[2] https://bio.informatik.uni-jena.de/software/sirius/

Participant:          Bach
ParticipantID:        MPIOKR_GaussianRFF
Category:	      4
Authors:              Eric Bach(1), Céline Brouard(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
Automatic pipeline:   yes
Spectral libraries:   no

Abstract

Magnitude-preserving Input Output Kernel Regression (MP-IOKR) is an
extension of the Input Output Kernel Regression (IOKR) method [1],
which has been successfully applied to metabolite identification
[2]. Magnitude-preserving IOKR uses a modified objective function,
which can exploit the knowledge about the molecular candidates for a
set of training MS/MS spectra.

IOKR objective function for the regression function h (prediction of
the feature vector representing a molecular structure):

(1) h = argmin_h sum_i ||h(x_i) - psi(y_i)||^2 + lambda ||h||^2
        
Magnitude-preserving modification of the objective function:

(2) h = argmin_h sum_i 1/n_i sum_j ||(h(x_i)-h(x_j))-(psi(y_i)-psi(y_j)))||^2 
                    + lambda ||h||^2
                    
with i in {1,...,l} being an iterator over the number of training data, j in {1,...,n_i} 
being an iterator over the number of molecular candidates of training example i. 
The x_i and y_i are the training MS/MS spectra respectively the training molecular 
structure. The x_j and y_j are the candidates' MS/MS spectra respectively the 
candidates' molecular structures. Equation (1) (IOKR) minimizes the prediction 
error between the training MS/MS spectra and the training molecular structure. 
In contrast to that, Equation (2) (MP-IOKR) learns a function h, which also preserves 
the magnitudes (differences) between the training molecular structure (y_i) and
all its candidate molecular structures (y_j). In that way we consider how the true
candidate relates to all the remaining candidates and include this knowledge into
our learning problem. It is important to node that we do not need the MS/MS 
spectrum of each candidate x_j, as we approximate the corresponding input feature 
vectors for each candidate using the molecular structure y_j.

MP-IOKR is a kernel method. Kernels measure the similarity between structured
objects, e.g. MS/MS spectra (input kernel) or molecular structures (output kernel).
On the input side, we use several kernels defined on MS/MS spectra and fragmentation
trees, and combine them uniformly, i.e. we sum up the kernels with equal weights.
On the output side, we use a Gaussian kernel calculated from binary fingerprints
representing the molecular structures in the candidate sets. As we cannot deal
with possibly millions of candidates (and O(million^2) kernel matrices). we 
approximate the Gaussian features and use those in our framework.

We train two separated MP-IOKR models one for each ionization mode, i.e. positive
and negative. For the positive model we use ~14000 identified MS/MS spectra +
~4Million candidates and their molecular structures. The MS/MS spectra are mainly
are extracted from the GNPS and MassBank databases. For the negative model we 
use ~5800 identified MS/MS spectra and ~1.5Million candidates. We represent the 
candidate molecular structures using ~7600 binary molecular fingerprints. 

For each challenge spectra we predict the molecular formula using Sirius [3] by
taking into account the possible molecule formulas based on the candidate sets.
The score we submitted for each candidate is the one corresponding to the most
likely molecular formula.

[1] Brouard, Cé.; Shen, H.; Dührkop, K.; d'Alché Buc, F.; Böcker, S. & Rousu, J. 
    Fast metabolite identification with Input Output Kernel Regression 
    Bioinformatics, 2016
[2] Brouard, Cé.; Bach, E.; Böcker, S. & Rousu, J.
    Magnitude-Preserving Ranking for Structured Outputs
    (submitted), 2017
[3] https://bio.informatik.uni-jena.de/software/sirius/