PRELIMINARY ARTIFICIAL NEURAL NETWORK ANALYSIS OF SELDI MASS SPECTROMETRY DATA FOR THE CLASSIFICATIO

1
PRELIMINARY ARTIFICIAL NEURAL NETWORK ANALYSIS OF
SELDI MASS SPECTROMETRY DATA FOR THE
CLASSIFICATION OF MELANOMA TISSUE

• Lee Lancashire

2
Melanoma(1)

• Serious form of skin cancer.
• Begins in the melanocytes (skin pigment melanin).
• Accounts for just 4 of all skin cancer cases but
  causes most skin cancer related deaths.
• Incidences are increasing, in the US since 1973,
  incidence rate per 100,000 people/year has risen
  from 5.7 14.3.

3
Melanoma(2)

• Most commonly defined in four stages
• Stage I Presence of mole or growth on top layer
  of skin.
• Stage II Growth deeper but not spreading.
• Stage III Spread to neighbouring tissue.
• Stage IV Melanoma has spread to other, more
  distant areas of the body.

4
Melanoma(3)

• ABCD system can help tell a normal mole from one
  that could be melanoma

• A Asymmetry melanoma lesions are typically
  asymmetrical.
• B Border melanoma lesions frequently have
  uneven or irregular borders
• C Colour melanoma lesions often contain
  multiple shades of brown or black.
• D Diameter early melanoma lesions are often
  more than 6 mm in diameter

http//www.melanoma.com/diagnosing/look/
5
Melanoma(4)

• If detected in its early stages, melanoma is
  often treatable and curable.
• Treatment usually involves surgery followed by
  either
• Chemotherapy
• Radiation Therapy
• Immunotherapy
• Combination of all 3

6
SELDI-MS(1)

• Surface Enhanced Laser Desorption Ionisation Mass
  Spectrometry.
• Involves on chip separation of complex mixtures
  together with mass spectrometry.
• Able to rapidly analyse samples containing vast
  amounts of proteins.
• Generates patterns that these proteins produce.
• Shows differences between these patterns for
  proteins expressed in different tissues, or in
  tissues during different disease states.

7
SELDI-MS(2)
? Sample is prepared in a molar excess
(10001) of a matrix compound. ? Matrix absorbs
light at the wavelength of the laser. ? Sample
and matrix molecules ejected into the gas phase,
where charge transfer occurs.
To mass spectrometer
8
SELDI-MS(3)

• This laser desorbs the proteins on the chip,
  causing them to be launched as ions.
• The time-of-flight (TOF) of the ion before
  detection by an electrode is a measure of the m/z
  value of the ion.
• Peptides with a larger m/z move more slowly down
  the flight tube and therefore have a longer TOF.

9
SELDI-MS(4)
10
SELDI-MS(5)
11
SELDI-MS(6)

• Mass intensity spectra is average of 5 reads.
• Data exported for ANN analysis.

12
Identification of potential biomarkers/therapeutic
targets

• 2-30 KDa. mass range consists of approx. 20,000
  data points.
• Identification of ions/proteins that may have any
  clinical relevance requires large sample size
  (n50at least).
• This suggests at least 1 million data points
  would have to be screened to find potential
  markers.
• Essential to implement a bioinformatic approach
  to cope with this mass of data.

13
Artificial Neural Networks

• Recent years have shown an increasing application
  of techniques such as ANNs for biological
  problems, in particular, cancer.
• Examples include prostatic, cervical, lung,
  ovarian and breast.
• Other uses include
• Prediction of rehospitilization
• Progression of glaucoma
• Classification of bacterial growth
• Plant response to ozone

14
Aims

• The identification of ions important in the
  correct classification of tumour grade which may
  serve as potential biomarkers representative of a
  specific disease state.
• To achieve this, a multi-layer perceptron (MLP)
  ANN with a back propagation (BP) algorithm and 2
  hidden nodes was used to model for 72 melanoma
  serum samples (36 stage I, 36 stage IV).

15
Parameterisation of models(1)

• Allows us to determine the importance of the
  inputs in a model.
• Eliminates unimportant inputs, removing noisy
  data and reduces model complexity.
• Data split into blocks, with each block being
  used in a separate model.
• Data blocks
• 2-5, 2.5-5.5, 3-6, 3.5-6.5, 4-7, 4.5-7.5, 5-8,
  5.5-8.5, 6-9, 6.5-9.5..27-30 KDa.

16
Parameterisation of models(2)

• Blocks trained over 50 random training/test/produc
  tion subsets (bootstrapping- high level of
  confidence).
• During training, ANN model is optimised against
  test set, then validated against production
  (validation) set.
• Model convergence determined by a failure of
  model to improve mean squared error (MSE) of the
  test data for 20,000 training events.
• Relative importance values for each input
  recorded and used for initial screening.

17
Parameterisation results (2-10 KDa.)
18
Parameterisation results (10-20 KDa.)
19
Parameterisation results (20-30KDa.)
20
Parameterisation of models(3)

• This relative importance analysis used to reduce
  the number of inputs in the model.
• The top 1,000 ions of greatest relative
  importance selected, and training process
  repeated.
• Same process used to determine top 500, 300, 200,
  100, 50, 30 and finally top 20 ions from initial
  set of 20,000.

21
Relative Importance of top 1000 ions
22
Additive approach(1)

• To identify the minimum number of ions from the
  top 20 which were capable of accurately
  predicting tumour grade.
• All ions taken sequentially and used as a single
  input in the model (creation of 1-ion model).
• 100 training/test/production subsets of each
  model used so that all tumour samples were
  treated as unseen a number of times.

23
Additive approach(2)

• MSE calculated and ion with the lowest error was
  selected for further training.
• Remaining 19 ions were added sequentially to this
  creating the 2-ion model which was trained as
  before.
• Process repeated in creating a 3 and 4-ion model.

24
Results Top 20 ions
25
Results
26
Receiver Operating Characteristic (ROC) curves

• Represents the values of the true positive ratio
  (sensitivity) and false positive ratio
  (specificity) at different possible prediction
  thresholds.
• Used in this study to assess model performance.
• Area under the curve (AUC) measures model
  performance
• A perfect test has an AUC of 1

27
ROC curve results
AUC results 1 ion model 0.574 (poor) 2 ion
model 0.748 (fair) 3 ion model 0.809 (good) 4
ions model 0.854 (good)
28
Comparison of models unseen data
29
Summary

• Parameterisation of models
• Identifies importance of inputs in a model
• Removes noisy data from system
• Reduces complexity from model
• Top 20 ions (from initial 20,000) of importance
  identified
• Additive approach employed to create a 4-ion
  model which predicts tumour grade with a gt80
  accuracy.

30
Conclusions

• By combining ANNs and SELDI-MS, essential ions
  involved in the classification of tumour grade
  can be found.
• These models are currently being developed
  further to deduce how many ions the optimal model
  contains for this data set.

31
Future Work

• Developing methods for the analysis of
  interactions between ions/proteins within the
  system.
• Sequencing of important proteins.
• These may have clinical relevance, important in
  establishing diagnostic markers.

32
Acknowledgements

• The Nottingham Trent University
• Dr. Graham Ball
• Dr. Shahid Mian
• Prof. Robert Rees
• Universitätsklinikum Mannheim
• Prof. Dirk Schadendorf
• Fifth Framework Programme