Cancer research is of great importance in life science and medicine and attracts research funds of thousands of millions dollars each year. With the explosion of biomedical research papers, it becomes more and more necessary to show the research trend in this spotlight area. In this paper, to provide a straightforward research atlas for the top killer cancers, Latent Dirichlet Allocation (LDA) is performed on the massive quantities of biomedical literatures. Moreover, Gibbs Sampling is used to make assessment on the parameters of the LDA model. The proposed evaluation carried out under multiple conditions with different Ks (the number of topics) for the top five cancers in recent five years. Additionally, a biomedical topic model was generated with the LDA model and delicate analysis was performed on the basis of that in order to explore the trending topic in cancer research. It can help the biology and medicine doctors quickly catch the frontiers of the cancer study, improve and expand their research programme, especially in today's era of “big data”.
1.1
Cancer is of great threat to human health, scientists and Medicians have been
continuously looking for effective treatment to conquer it, however, cancer research is
a very challenging field in today's life science. Over the past 5 years cancer research
has diverged enormously, partly based on the quickly development of biotechnology
and bioinformatics. It is not easy to summarize recent trends for different cancers’
study, and identify where the new findings are and therapies might come from [
The topic model is a probabilistic model of text mining appeared in recent years [
The rest of the article is organized as follows: we start in section 2 with a brief review of a topic model named as Latent Dirichlet Allocation and its related work. In section 3, the algorithm of Gibbs Sampling is introduced, and the framework of our exploration is described in detail. It presents the experimental methodology and results on Medline dataset in Section 4. At last, conclusions and future work are depicted in Section 5. 2
Hyper parameters are subject to Dirichlet distribution. We explain the symbol used to describe the model in Table.1. Now we have a text representation of the probability: p(wm , zm ,θ m , Φ |α ,β )
Nm =p(wm,n ∏ |ϕ zm,n ) p(zm,n |θ m ) ⋅ p(θ m |α ) ⋅ p(Φ | β ) n=1 (1)
The convergence of LDA is a more critical issue. We use the convergence function to solve it. Under the given model conditions, we choose the appearance probability of samples as the evaluation criterion of the model. The performance of the LDA model: (2) (3) (4) M p(w | M) = ∏ p(w m |θ m , Φ)
m=1 =∑K( ∏M ∏Nm nkt + β t ⋅ m=1 n=1 k=1 ∑ Vt =+β 1 nkt t
nmk + α k ∑ kK =+α 1 nmk k ) .
For the convenience of the calculation, we use log transform to the equation (1),
denoted as . Along with the iterative constantly, is used to determine the model
convergence [
Lˆ = - ∑M ∑Nm log2 ( ∑K ( nkt + β t ⋅ m =n=1 1 k =∑Vt 1 =+β 1 nkt t
nmk +α k ∑ kK =+α 1 nmk k )) 3 3.1
In fact, LDA is one of the probabilistic. Data are generally divided into two parts,
visible variables and latent variables. It is believed in the topic model that the data is
produced by the generation process, which defines the joint probability distribution of
visible random variables and latent random variables. For many modern probabilistic
models, including Bayesian statistics, the priori probability calculation is extremely
difficult. So the core research objective of modern probability modelling is to do
everything possible to obtain an approximate solution. Random sampling is a kind of
methods for solving the approximate solution with good performance. This article
describes a method commonly used sampling MCMC (Markov Chain Monte Carlo)
[
MCMC methods have Gibbs Sampling algorithm and Metropolis-Hastings (MH)
[
N-dimensional Gibbs Sampling: 1. Random initialization { xi : i =1, 2, ⋅ ⋅ ⋅, n } 2. For t
=0,1, 2, ⋅ ⋅ ⋅ loop in sampling x 1(t +1) ～ p(x1 | xt2 , xt3 , ⋅ ⋅ ⋅, xtn )
We choose the top 5 cancer from the top 10 deadliest cancer published by the
LiveScience [
Then we make a pre-process to extract the title and abstract for each paper and
make some hyphens for the entity name for a better segmentation. After that we use
WordNet [
Take the dataset of Pancreatic Cancer in 2010 as an example, the number of the abstract is 2088, includes 54004 words. The number of topic K is set to 100, 200, 300, 400, 500, considered for model convergence condition.
According to the evaluation function of the convergence mentioned in Section 2, which shows the cost of compression of the text using the model, the smaller the better. Figure.2 below respectively under different K values for the iterative convergence condition of iteration times of 400 and 1000.
We measure the similarity and differences between the 5 cancer research by their
word vector cosine coefficients similarity [
We further analysis the output of the LDA Gibbs Sampling, particularly in the
topic words file (this file contains topic words most likely words of each topic). We
first transform the topic-word matrix to a topic word vector representing the selection
and the frequency of topic words to describe one cancer research. Then we use feature
scaling normalization [
Unlike the cosine coefficients which give a numerical description on the trend and topic of the 5 cancers, the common words are easier for readers to have an intuition on the 5 cancers topics. We do the both analysis of the results so that we may get a comprehensive answer. 4
To examine the behavior and the performance of LDA, the experiments are illustrated
on the widely used Medline dataset. In order to straightforwardly compare the trend
and topic resulted from the LDA model, the visual results to get the entire recognition
of the topic words of different cancers are shown, which uses the Word Cloud tool
supported by Tagul.com [
The publicly available Medline dataset is provided by NCBI PubMed, which contains a title and an abstract for each paper on the five fatal cancers. In order to get a better visual understanding of the data, the 5 cancers with different colors (in details, breast cancer blue, colon cancer orange, lung cancer green, pancreatic cancer red, prostate cancer purple) are listed in Table 2. An outline of the selected dataset is shown below (actually we collect the 2014 data before October 2014, that is why it seems all the 2014 paper decrease from the previous years): After introducing the Gibb Sampling and transforming the topic-word matrix to the vector, we draw all the five cancer topic words in the recent five years. A snapshot of the collection of the topic words clouds are shown in Figure.3.
Filtering the topic words for each cancer into the common words collection, we can easily discover the main topic for each cancer and new findings and focus for each year. Take breast cancer as an example shown below:
In the Figure.4 above, we can find out the big words in it like cell, breast cancer, tumor, patient, woman and so on, these words are simply the most common used to elaborate breast cancer, which means if you talk about breast cancer and you just can’t avoid mentioning them. They are obviously from the different topics, therapy, factor, risk and other aspects of the breast cancer. The common words are defined as the words which appear every year for certain cancer research during 2010-2014. They are shown in Figure.5 below, and the shape of the words represent their frequency. 4.3
The Figure.6 above shows the contrast result of the vector cosine coefficient between two cancers in the same year. The higher coefficient, the more similar. We can easily find out that all the similarities contrast with colon cancer (see colon&lung, colon & pancreatic, breast&colon, colon&prostate) are higher than the other cancer pairs. It indicates colon cancer research is more related with the others. It is because that: lung cancer may easily metastasize to colon; colon cancer and pancreatic cancer are both belong to lower digestive cancers; lack of exercise and sitting for long time are the common causes of breast and colon cancer; long-term androgen deprivation therapy for prostate cancer may increase the risk of colon cancer.
And we find it interesting that almost all the five cancer research diverse from each other in 2012, because they have a concave at the time in the figure. Fig.7 shows the topic words changes for each cancer research during the recent five years. We can learn from the figure that breast cancer, lung cancer, prostate cancer, these 3 cancer research only change a little, and pancreatic cancer research changes a lot. It is because that the pancreatic incidence rate increase fast recently and it is named as “the king of cancer” because the lowest survival rate. Many scientists engage in the research of pancreatic cancer. In this paper, we first apply LDA Gibbs Sampling model on the analysis of the top 5 deadliest cancer research trends, which is extended from cosine coefficient using the vector space model after transforming topic-word matrix into topic word vector. Then we generate the common topic word collection for each cancer research, in order to get the trending topic words. We further explore the trend by comparing and contrasting topic words for each cancer and their cosine coefficients. It is found that the trending topic words for the 5 cancers research from 2010 to 2014, which are depicted in the words clouds. Moreover, it is found that numerical trends of the four cancers research are as follows: breast cancer, lung cancer, and prostate cancer are of little change, but pancreatic cancer is changing a lot in the recent five years. But what cause all five cancer research diverse in 2012, how to visualize the allocation of the topics for each cancer research, and how to make a computational evaluation for the trend results, are all what need to be explored in our future work.
The authors are grateful to the support of the NSFC (61272207, 61472158, 61103092) and the Science Technology Development Project from Jilin Province (20130522106JH, 20140520070JH）