Skip to main content

Archived Comments for: Discovering biomarkers from gene expression data for predicting cancer subgroups using neural networks and relational fuzzy clustering

Back to article

  1. Current biomarker prediction scheme looks familliar to me!

    Junbai Wang, CU

    22 March 2007

    Dear Sir/Madam:

    I feel that the present published biomarker prediction method is quite similar to an old paper: Tumor classification and marker gene prediction by feature selection and fuzzy c-means clustering using microarray data: BMC BIOINFORMATICS 4: Art. No. 60 DEC 2 2003. If you look at the Figure (6) of Nikhil R Pal et al paper then look at the Figure (5) of that BMC Bioinformaitcs 2003 paper then you will have my thought too. In addition, the idea of applying K-means clustering/fuzzy k-means clustering on the neural networks/SOM for biomarker prediction is not new because it was published in another BMC Bioinformatics paper before (please refer to BMC Bioinformatics 2002, 3:36 ).

    Competing interests


  2. RE: Current biomarker prediction scheme looks familliar to me!

    Animesh Sharma, Fuzzylife

    20 May 2007

    This is in connection with the comments made by Mr. Junbai Wang. For the benefit of the readers we provide here the title and direct links to the two papers mentioned by Mr. Wang: Paper 1 - "Tumor classification and marker gene prediction by feature selection and fuzzy c-means clustering using microarray data" and Paper 2 - "Clustering of the SOM easily reveals distinct gene expression patterns: results of a reanalysis of lymphoma study".

    We read the above two papers and this is what we found:

    1. SOM - Paper 1 and 2 use optimally selected self-organizing maps (SOMs) while we do not use SOM.

    2. Clustering - Paper 1 uses Fuzzy c-means (FCM) while Paper 2 uses k-means. We use Non-Euclidean Relational Fuzzy c-Means (NERFCM).

    3. Perceptron based learning - Paper 1 and 2 do not use Perceptron based learning. We employ an OFS, a Multi-layer perceptron (MLP) to select features. It is a modified MLP developed by our group and has absolutely no relation to SOM.

    4. General Methodology - Paper 1 uses FCM with Euclidean distance to cluster tumor samples and then uses the FCM memberships to generate weighted SOMs. Paper 2 uses k-means (with Euclidean distance) to cluster SOM prototypes, the outcome of an unsupervised process. We cluster genes (selected by a supervised process) using NERFCM. It is NERFCM and not FCM. This makes a lot of differences, FCM or k-means is likely to place two negatively correlated genes in different clusters while with NREFCM we can have them in the same cluster.

    5. Feature Selection - Paper 1 uses Visual inspection of weighted SOM / pair-wise Fisher's linear discriminant. Paper 2 uses Clustering of SOM prototypes. We use neither of the techniques employed by the other 2 papers. We have used OFS and NERFCM (details in paper).

    6. Figure - Figure 5 of Paper 1 gives a flow chart of the steps in their method. Figure 6 of our paper gives a flow chart of our steps. Yes, they are similar only in the sense that both are flowcharts.

    Another point worth mentioning is that our online FSMLP based feature selection is a scheme that picks up features simultaneously with the learning of the multilayer perceptron network and thus can account for the interaction between features, the tool and the classification problem (class label) which is not done by unsupervised processes.

    N. R. Pal, K. Aguan, A. Sharma and S. Amari

    Competing interests

    None declared