Folie 1

1
Spectroscopic Molecular Pathology and Diagnosis
Matthew J Baker EPSRC LSI Fellowship Review
Meeting 20th March 2009
2
Proposed Research from the EPSRC LSI Proposal

• Milestones
• Raman spectra and images taken of cell lines and
  stem cells.
• Development of multivariate statistical analysis
  to turn hyperspectral data into specific
  knowledge about pathological processes.
• Identification of spectral components
  responsible for discrimination of Raman spectra.

3
Content
Prostate cancer tissue pathology investigated by
FTIR imaging
Preliminary study using single cell Raman
spectroscopy to investigate populations of cells
from PC-3 cell lines
4
(No Transcript)
5
RWPE Cell Lines

• Cell lines have proven to be interesting models
  to study utilising FTIR spectroscopy due to the
  more homogenous population in contrast to
  heterogenous tissue
• Cell lines have previously been discriminated
  using vibrational spectroscopy. (Stone et al.
  BJC)
• However those cell lines were from different
  anatomical positions so the reason for
  discrimination is uncertain.

Transformed by Ki-Ras
Transformed by N-methyl-N-nitrosurea (MNU)
6
RWPE Cell Lines - Aims

• To observe if FTIR spectroscopy can be combined
  with machine learning and multivariate analysis
  to enable discrimination of cell lines derived
  from the same anatomical position
• To investigate the underlying causes of cancer
  and invasiveness using RWPE cell lines
• To move towards standardised selection of
  pre-processing techniques

7
RWPE Cell Lines - Method

• Cell lines were cultured according to identical
  ATCC protocols.
• Cells were cultured onto MirrIR slides until 80
  confluent then fixed in 4 formalin and air
  dried. 30 slides per cell line representing 30
  different cultures were prepared.
• 10 FTIR spectra were acquired from each slide
  (each culture). 256 co-added scans and 4 cm-1
  resolution.
• After quality testing spectra were split into
  training set (150 spectra per cell line),
  validation set (30 spectra) and a doubly blind
  test set (at least 70 spectra per cell line (530
  spectra in total)).
• Spectra were then subjected to a genetic
  algorithm fed support vector machine to find the
  optimum pre-processing steps based upon correctly
  classified validation set.
• Then the model is tested by the doubly blind
  set.

8
RWPE Cell Lines Genetic Algorithm (GA) fed
Support Vector Machine (SVM)

• Genetic algorithm (GA) based upon Darwinian
  evolution. (Jarvis et al. Bioinformatics, 2005,
  Vol21(7)).
• We have used it to assess which is the best
  pre-processing technique or combination of
  techniques.

9
RWPE Cell Lines Genetic Algorithm (GA) fed
Support Vector Machine (SVM)

• SVM models can be used as supervised
  classification methods to classify non linear
  problems.
• The SVM transforms from an input space (original
  variables) into a feature space which can have
  infinite dimensions.

• The model seperates the classes by inserting a
  hyperplane.
• The variables on the margins of the hyperplane
  are called support vectors.
• If a variable is on the wrong side of the
  hyperplane it contributes to a penalty function.
• O Ivanciuc. Reviews in Computational Chemistry
  Vol 23 (2007)

10
RWPE Cell Lines - Results
Blind set overall sensitivity 97.38 and overall
specificity 99.41
11
RWPE Cell Lines - Results
SVM Results CV Accuracy 98.99 Blind Set
Accuracy 95.47
12
RWPE Cell Lines - Conclusions

• To observe if FTIR spectroscopy can be combined
  with machine learning and multivariate analysis
  to enable discrimination of cell lines derived
  from the same anatomical position
• To investigate the underlying causes of cancer
  and invasiveness using RWPE cell lines
• To move towards standardised selection of
  pre-processing techniques

• Expanding the pre-processing genetic algorithm
  options e.g new developments in spectral
  artefacts (P Gardner 1050 am) and region
  selection.
• Further assays and research into the use of RWPE
  cell lines as models of prostate cancer. Can
  biological tests differ between different methods
  of transformation?
• Investigation into the use of spectroscopic
  imaging for the differentiation of co-cultured
  cell lines

13
Prostate cancer tissue pathology investigated by
FTIR imaging
14
FTIR Pathology of Prostate Tissue

• Prostate Cancer (CaP) is a worldwide problem,
  the fifth most diagnosed cancer and the most
  diagnosed gender specific cancer.
• Current methods used to diagnose CaP are the
  prostate specific antigen test and Gleason
  grading system. (J Shanks 1440 )
• Both of these techniques are flawed. The PSA
  test is related to the volume of the prostate and
  as such is not specific for CaP and the Gleason
  grading system has been shown to have inherent
  intra and interobserver variability.
• The pathological use of FTIR imaging has many
  applications and could aid current diagnostic
  techniques.

15
FTIR Pathology of Prostate Tissue - Aims

• To collect spectra from CaP tissue blocks to
  build a spectral database based upon
  histopathological assignments.
• To collect FTIR images from CaP tissue blocks
  and observe whether the use of the spectral
  database applied through an Artificial Neural
  Network (ANN) can differentiate the
  histopathological assignments.
• To compare and contrast the ANN approach with
  Hierarchical Cluster Analysis (Wards Algorithm).

16
FTIR Pathology of Prostate Tissue - Method

• Adjacent tissue sections were cut with one
  placed onto a CaF2 disc for FTIR analysis and the
  other onto a microscope slide for HE staining.
• Dewaxing.
• Spectra were collected from 80 patients (20 BPH,
  20 GSgt7, 20 GS7 and 20 GSlt7).
• These spectra represent 128 co-added scans at a
  6 cm-1 resoution with every 5th spectrum a
  background spectrum. Quality tested.
• FTIR images were collected from previously
  unanalysed samples. (64 co-added)
• Spectra and images were pre-processed and
  analysed using OPUS software, Cytospec and
  Synthon Neurodeveloper. (HCA, ANN and K means
  clustering)
• Due to residual paraffin spectral region 1800
  1000 cm-1 was used.

17
FTIR Pathology of Prostate Tissue - HCA

• Hierarchical Cluster Analysis (HCA) is non
  subjective. (It cant be used for projection)
• A similarity matrix based upon Euclidean
  Distances is constructed.

18
FTIR Pathology of Prostate Tissue Artificial
Neural Networks
Spectral Features

• ANNs are information processors inspired by the
  way the nervous system processes information.
• The network is structured so that it contains a
  large number of highly connected elements.
• We are using an ANN in a supervised learning
  fashion. Based upon a training set, validation
  set and blind set.
• The optimum ANN is described by the minimum
  summed squared error of the validation set

Training Process
19
FTIR Pathology of Prostate Tissue Artificial
Neural Networks
Top Level Net
Sub Net
Top Level Net 1st Der 13 SP 84 Spectral Inputs 1
Hidden Layer 3 Neurons Training Set 1995
(33) Validation Set 486 (8) Blind Set 4521
(39) Accuracy 82.1
Sub Level Net 1st Der 15 SP 93 Spectral Inputs 1
Hidden Layer 3 Neurons Training Set 1180
(33) Validation Set 231 (8) Blind Set 1173
(39) Accuracy 84.1
SPECTRAL INPUTS
Stroma
Lymphocyte
Cancer
Cancer and BPH
BPH
20
FTIR Pathology of Tissue - Results
H E Optical Image
1070 µm
GS6
1070 µm
GS7
21
FTIR Pathology of Tissue - Results
H E Optical Image
1070 µm
GS7
1070 µm
BPH
22
FTIR Pathology of Tissue - Results
H E Optical Image
2140 µm
2140 µm
23
FTIR Pathology of Tissue - Conclusions

• To collect spectra from CaP tissue blocks to
  build a spectral database based upon
  histopathological assignments.
• To collect FTIR images from CaP tissue blocks
  and observe whether the use of the spectral
  database applied through an Artificial Neural
  Network (ANN) can differentiate the
  histopathological assignments.
• To compare and contrast the ANN approach with
  Hierarchical Cluster Analysis (Wards Algorithm)

• To collect larger images from larger tissue
  samples.
• Research the construction of the ANN (using
  every possible combination of training,
  validation and blind).
• After further analysis and examination of data
  increase the number of classes in the ANN.
  Research the use of alternative MVA e.g LDA
  (Manfait Analyst)
• Analyse the results based upon Dr Shanks
  microscopic pathology.

24
Preliminary study using single cell Raman
spectroscopy to investigate populations of cells
from PC-3 cell lines
25
Single Cell Raman Spectroscopy - Aim

• Increasing evidence to suggest that aberrant
  cellular proliferation begins at stem cell level.
• Stem cells have been shown to play crucial roles
  in acute myeloid, lymphoid and chronic myeloid
  leukaemia.
• Prostatic stem cells have been implicated in the
  aetiology of prostate cancer.
• Recently we have been able to isolate a Hoescht
  3342 side population which in the haemotopoetic
  system (blood production) is highly enriched for
  stem cells.
• Both benign and malignant prostatic side
  populations have been isolated.
• However the number of cells are too few for
  conventional analysis.

Brown et al. The Prostate 671384-1396 (2007)
26
Single Cell Raman Spectroscopy - Aim
BPH
Cell Volume
CaP

• The small sample size obtained from benign and
  malignant tissue make the samples valuable
• As a test, side population (SP) and non-side
  population (NSP) cells have been collected from
  the PC-3 cell line and analysed by Raman
  spectroscopy.

27
Single Cell Raman Spectroscopy - Method

• Cells obtained from FACs sorting using Hoesch
  3342 and split into side and non side
  populations.
• Cells fixed in 4 formalin in PBS and shipped
  to Berlin.

28
Single Cell Raman Spectroscopy - Results Non Side
Population
2981 2952 cm-1 CH3 stretch
2834 2910 cm-1 CH2 sym stretch
1545-1598 cm-1 CC Protein
29
Single Cell Raman Spectroscopy - Results Non Side
Population
30
Single Cell Raman Spectroscopy - Results Non Side
Population
Overlay
CC Amide
Sym CH2
31
Single Cell Raman Spectroscopy - Results Side
Population
CH2 Asym
C-C (Lip), C-N (Pro), PO2-
32
Single Cell Raman Spectroscopy - Conclusions

• This was a preliminary experiment designed to
  prove the technique for analysis of valuable
  human samples (inherently low cell numbers).
• Shown viability of experiment
• Possible sensitivity of Confocal Raman
  Microscopy for stem cell differentiation

• Further experimental preparation on cell line
  populations.
• Moving onto human prostatic stem cell enriched
  side populations and non side populations.
• Heterospectral analysis combination of
  ToF-SIMS and Confocal Raman Microscopy

33
Achieved Research from the EPSRC LSI Proposal

• Achieved Milestones
• FTIR and Raman spectra and images taken of cell
  lines and stem cells.
• Development of multivariate statistical analysis
  to turn hyperspectral data into specific
  knowledge about pathological processes.
• Identification of spectral components
  responsible for discrimination of FTIR and Raman
  spectra.
• The use of FTIR imaging spectroscopy for
  prostate cancer tissue diagnostics.

34
EPSRC LSI Proposal

• Aim of fellowship is not to just conduct
  research.
• Gain experience of working in international
  laboratories.
• Initiate contacts which will hopefully be kept
  throughout academic career.
• Learn from experts in your field.

Thank you to Prof. Dr. Naumann and everybody at
the Robert Koch for making me feel very welcome
and helping with my research
Grant Success Awarded a 3 800 Royal Society
International Travel Grant to allow me to spend a
short research period at the Centre for
Biospectroscopy, Monash University, Melbourne in
Developing 4D Imaging Molecular Pathology and
to attend ICAVS.
35
Thank you for listening