Chapter 19Proteomic Disease Classification Algorithm

In this chapter, we study an algorithm [1] to automatically classify the status of a patient to be healthy or to have a particular disease, based on the mass spectrum (MS) data of his or her blood serum. We have been studying proteins in this book using atomic-resolution biophysical methods such as NMR and X-ray crystallography. MS, in contrast, allows us to assay and compare the masses of many proteins at the same time.

Proteomic Disease Classification

Mass spectrometry (MS) provides ultrahigh-resolution mass information. A mass spectrum consists of a set of m/zm/z (“mass-to-charge”) values and corresponding relative intensities that are a function of all ionized molecules present with that m/zm/z ratio; mm is mass and zz is charge; m/zm/z is plotted on the horizontal axis, and the corresponding relative intensity is plotted on the vertical axis of a typical mass spectrum. Mass spectrometry classification algorithms (MSCAs) are designed to discriminate one condition from another by comparing their mass spectra. For example, an MSCA can tell whether or not a patient has a particular disease from MS data of his peripheral blood.

Methods

Spectral space Each spectrum is represented as a point (p1,…,pr)(p_1, \dots, p_r) in an rr -dimensional space, where rr is the number of m/zm/z values for which relative intensities pip_i are recorded per spectrum.

Training set A set of data-class pairs (r,c)(r, c) , where rr is the data and cc is the class that rr belongs to. For example, rr is the mass spectrum of a patient, and c∈{diseased,healthy}c \in \{diseased, healthy\} indicating the health status of the patient. In this study, diseased can only mean one particular disease, for example, ovarian cancer. An extension was considered in Lilien et al. [1] to multiple classes (more than two).

Classifier A classifier is a mapping from data to classes. Usually, the classifier is first trained on a training set, then, for any data (points in the spectral space), it outputs a class for the data (for example, diseased or healthy).

Q5: An MSCA Algorithm

An MSCA called Q5 was introduced in Lilien et al. [1]. In the training step, it first reduces the dimension of the data in the training set by PCA. The dimension is reduced to the intrinsic dimensionality of the problem, which is the number of samples (in this case, patient blood samples) minus the number of classes. Then, a hyperplane HH is computed by linear discriminant analysis (LDA) so that the projections of data in the training set onto HH best discriminate the two classes, in terms of the ratio of the in-class variance to the between-class variance.

In the testing step, for a point pp in the spectral space, Q5 first reduces the dimension of pp by projecting it to the same subspace computed in the training step. Then, the reduced-dimension data is further projected to HH . Finally, a probabilistic classifier is applied to compute which class pp belongs to.

More precisely, Q5 works as follows:

Training Stage

Input Data spectrum represented as column vectors. 𝒙∈X\mathbf{x} \in X (healthy) and 𝒚∈Y\mathbf{y} \in Y (disease). Let nx=|X|n_x = |X| , ny=|Y|n_y = |Y| , n=nx+nyn = n_x + n_y .

PCA

  • Compute all-class-mean μ′=1n(∑𝒙∈Xx+∑𝒚∈Yy)\mu' = \frac{1}{n} (\sum_{\mathbf{x} \in X} x + \sum_{\mathbf{y} \in Y} y) .
  • Let X′X' and Y′Y' be the zero-mean matrices obtained by subtracting μ′\mu' from XX and YY , respectively. Let 𝑷=[X′Y′]\mathbf{P} = [X'Y'] be the r×nr \times n matrix.
  • Let 𝑪=𝑷𝑷T\mathbf{C} = \mathbf{P}\mathbf{P}^T . Compute the eigenvalues and normalized eigenvectors of 𝑪\mathbf{C} , denoted by λi\lambda_i and viv_i (i=1,…,wi = 1, \dots, w), respectively. That is, for any i=1,…,wi = 1, \dots, w , 𝑪vi=λivi\mathbf{C}v_i = \lambda_i v_i .
  • Sort λ1,…,λw\lambda_1, \dots, \lambda_w and discard the eigenvectors corresponding to small eigenvalues. Let the remaining eigenvectors be 𝑽\mathbf{V} .
  • Reduce the dimension of 𝑷\mathbf{P} by 𝑽\mathbf{V} . For any 𝒙′∈X′\mathbf{x}' \in X' , let 𝒙p=𝑽𝒙′\mathbf{x}_p = \mathbf{V}\mathbf{x}' ; for any 𝒚′∈Y′\mathbf{y}' \in Y' , let 𝒚p=𝑽𝒚′\mathbf{y}_p = \mathbf{V}\mathbf{y}' .

LDA

  • Compute the within-class means. Let μx=1nx∑𝒙p∈Xp𝒙p\mu_x = \frac{1}{n_x} \sum_{\mathbf{x}_p \in X_p} \mathbf{x}_p and μy=1ny∑𝒚p∈Yp𝒚p\mu_y = \frac{1}{n_y} \sum_{\mathbf{y}_p \in Y_p} \mathbf{y}_p .
  • Compute the within-class scatter matrix 𝑺w\mathbf{S}_w by

𝑺w=𝑴x𝑴xT+𝑴y𝑴yT,\mathbf{S}_w = \mathbf{M}_x \mathbf{M}_x^T + \mathbf{M}_y \mathbf{M}_y^T, where the columns of 𝑴x\mathbf{M}_x are 𝒙p−μx\mathbf{x}_p - \mu_x ; the columns of 𝑴y\mathbf{M}_y are 𝒚p−μy\mathbf{y}_p - \mu_y .

  • Compute the between-class scatter matrix 𝑺b\mathbf{S}_b by

𝑺b=nx(μx−μ)(μx−μ)T+ny(μy−μ)(μy−μ)T.\mathbf{S}_b = n_x (\mu_x - \mu) (\mu_x - \mu)^T + n_y (\mu_y - \mu) (\mu_y - \mu)^T.

  • Let 𝒆\mathbf{e} be the normalized generalized eigenvector of 𝑺b\mathbf{S}_b and 𝑺w\mathbf{S}_w corresponding to the largest generalized eigenvalue λ\lambda . That is, 𝑺b𝒆=λ𝑺w𝒆\mathbf{S}_b \mathbf{e} = \lambda \mathbf{S}_w \mathbf{e} .

Let p1p_1 (respectively p2p_2) be the mean of the (𝒙p)T𝒆(\mathbf{x}_p)^T \mathbf{e} (respectively (𝒚p)T𝒆(\mathbf{y}_p)^T \mathbf{e}).

Output 𝑽\mathbf{V} , 𝒆\mathbf{e} , p1p_1 , and p2p_2 .

Testing or Prediction Stage

Input zz in the spectral space, and a threshold t∈[0.5,1]t \in [0.5, 1] .

Let zd=(𝑽(z−μ′))T𝒆z_d = (\mathbf{V}(z - \mu'))^T \mathbf{e} .

Probabilistic Classifier

  • Let qq be the midpoint of p1p_1 and p2p_2 . Choose σ\sigma such that exp⁡[−(q−p1)2/σ2]=12\exp[-(q - p_1)^2/\sigma^2] = \frac{1}{2} .
  • Compute Pr(zd∈Ci)=exp⁡[−(zd−pi)2/σ2]Pr(z_d \in C_i) = \exp[-(z_d - p_i)^2/\sigma^2] (i=1,2i = 1, 2). If Pr(zd∈Ci)>tPr(z_d \in C_i) > t for some ii , then output zdz_d is in class CiC_i . Otherwise output ambiguous.

Results and Discussion

Q5 was demonstrated on several MS data sets for patients with ovarian and prostate cancer. In all cases, excellent accuracy was obtained in disease classification. In Lilien et al. [1], an extension was described to identify biomarkers as well. In addition to impressive experimental accuracy, Q5 has guarantees, as an algorithm, of efficiency, soundness, and completeness. Such algorithms can probe the network of proteins that interact in cancer processes, and may help translate MCSA diagnosis toward the clinic.

This chapter shows how geometric techniques from statistical estimation and machine learning can be used to develop an algorithm for cancer proteomics, which uses data from a mass spectrometer to distinguish between healthy and diseased blood in humans. While geometry pervades the algorithms in this book, Lilien et al. [1] can be viewed as an investigation of geometry at a larger scale—instead of considering a single protein (or a small number of proteins) and their interactions, this chapter discussed algorithms for classifying and characterizing the geometry of the oncoproteome (i.e., the space of “all” expressed cancer proteins) as projected onto the mass-to-charge ratios of their proteolytic digest.

Reference

[1] R. Lilien, H. Farid and B. Donald. Journal of Computational Biology. 10;6(2003):925–946.