BioSysBio:abstracts/2007/Shanfeng Zhu
- Add or delete the sections that you require.
Predicting Implicit Associated Cancer Genes from OMIM and MEDLINE by a New Probabilistic Model
Authors: Shanfeng Zhu, Yasushi Okuno, Gozoh Tsujimoto and Hiroshi Mamitsuka
Affiliations: Kyoto University, Japan
Contact:email: zhusf (at) scl.kyoto-u.ac.jp
Keywords: 'Cancer gene discovery' 'Machine Learning' 'Text Mining' 'Probabilistic model'
Abstract
Since the co-occurrence of biological entities has been shown to be a popular and efficient technique for identifying biological relationships in biomedical text mining, we extracted cancer-gene and cancer-cancer co-occurrence data from OMIM by a software tool CGMIM, and gene-gene co-occurrence data from MEDLINE. A new probabilistic model, which we call MAM, was employed to find implicit associated cancer genes by mining cancer gene co-occurrece data from OMIM and MEDLINE. Through a series of cross-validation experiments, the accuracy of predicting cancer-associated genes was shown to be significantly improved by incorporating gene-gene co-occurrence pairs from MEDLINE into cancer-gene co-occurrence pairs in OMIM, especially in the case of a small size of training data. Furthermore, after training the MAM with all three types of co-occurrence data, some implicit associated cancer genes were predcited. The detailed result was presented on line (http://www.bic.kyoto-u.ac.jp/pathway/zhusf/CancerInformatics/Supplemental2006.html) for the reference of interested researchers and further validation by biologists.
Introduction
Discovering cancer associated genes can facilitate the understanding of tumor pathogenesis, the medical diagnoses and the treatment of patients. In addition to cytogenetic and molecular genetic techniques, we can resort to bioinformatics approaches to discover cancer associated genes by analyzing high-throughput biological data, such as genomics, transcriptomics and metabonomics data [1]. Here we mined OMIM and MEDLINE to discover implicitly associated cancer genes by applying a new probabilistic model, mixture aspect model (MAM) [2], on cancer gene co-occurrence data in OMIM and MEDLINE. Analyzing the co-occurrence of biological entities in the literature is a simple, comprehensive and popular technique to characterize the association of these entities [3]. MAM was proposed by us to mine implict "chemical compound-gene" relations by integrating three types of co-occurrence data (compound-compound, gene-gene and compound-gene) in the literature [2]. It was extended from a clssical probabilistic latent variable model, which is generally called aspect model (AM), which has wide applications in information retrieval and collaborative filtering in E-commerce [4]. The main advantage of MAM is the ability of integrating different type of co-occurrence data from heterogenous data sources. MAM was first estimated by an EM algorithm to fit the existing co-occurrence data of cancer and gene, and then was used to predict the likelihood of the association of an unobserved pair of a cancer and a gene. Through cross-validation experiments, the accuracy of predicting associated cancer genes was shown to be improved by incorporating gene-gene co-occurrence pairs from MEDLINE into cancer-gene co-occurrence pairs in OMIM. Furthermore, some implict associated cancer genes were predcited and analyzed preliminarily. The result was presented on line for further validation by biologists.
Data
We extracted cancer-gene and cancer-cancer co-occurrence pairs from OMIM, a human curated knowledgebase on human genes and inherited diseases. A software tool CGMIM was used to extract the description section of OMIM to obain cancers and assocated genes [5]. This software maps genetic disorders into 21 different types of cancers. The gene-gene co-occurrence pairs were extracted from MEDLINE. To avoid the difficulty of recognizing gene names, we extracted a human curated database, Locuslink, to obtain a subset of high quality MEDLINE records, where we obtained gene-gene co-occurrence data. The size of co-occurrence datasets is shown in Table 1.
Table 1: The size of co-occurrence datasets Item Gene type Gene-Gene Cancer type Cancer-Cancer Cancer-Gene Size 2,017 3,118 21 206 3,743
Results
We evaluated the performance of MAM by cross-validation on predicting associated cancer-gene pairs. In addition to training AM on cancer-gene pairs, we trained three other types of MAM by incorporating different type of co-occurrence data. 2MAM (CG+CC) and 2MAM (CG+GG) were built by adding cancer-cancer pairs and gene-gene pairs, respectively. In addition, 3MAM was built by incorporating all three types of co-occurrence data. To explore the effect of the size of the training data set on the performance of the probabilistic model, we set three different ratios of the size of training to test datasets, 3:1, 1:1 and 1:3, in the cross-validation experiment. The negative test examples were randomly generated, and it was assured that no negative test example would appear in either training or postive test data. We carried out 50 rounds of this cross-validation to reduce possible biases occurring in only a few rounds and averaged the results obtained. After estimating the probability parameters of a probabilistic model from training data, we computed the likelihood of each cancer-gene pair in test data and ranked all pairs according to their likelihoods. Then it would be evaluated by AUC (Area Under the ROC curve). Please note that the larger the value of AUC, the better the performance of the model. The t-value was also computed to check the statistical significance of the different performance by two models. Here if the t-value is greater than 3.50 (2.36), the difference is more than 99.9% (98%) statistically significant. As illustrated in Table 2, 3MAM outperforms all other models statistically signficant in all cases. The improvement over AM ranges from 2 to 9%, and is especially significant in the case of a small size of training data.
Table 2: AUCs and t-values obtained in the cross-validation experiment on cancer-gene association prediction. Model The ratio of training to test data 3:1 1:1 1:3 3MAM(CG+CC+GG) 76.1 74.6 73.2 2MAM(CG+CC) 75.8(2.56) 74.2(2.44) 71.8(12.9) 2MAM(CG+GG) 73.9(17.2) 71.4(22.5) 68.3(38.0) AM(CG) 74.1(14.7) 70.5(26.3) 64.9(55.1)
After training 3MAM with all three types of co-occurrence data, we can compute the likelihood of all other cancer-gene pairs that are unknown in the OMIM. From top 20 implicit associate cancer gene pairs we found some famous oncogenes, such as BCL2, TP53 and TNF, which lead to an unsurprising result. Thus we further calculated the likelihood of a gene associated with a specific cancer. For each type of cancer, we present the top implicit gene in the Table 3. One interesting result is the top implict assocated gene specific to the colorectal cancer, CYP1A1, which was already verfied by Hou et al [6]. More detailed results can be accessed from http://www.bic.kyoto-u.ac.jp/pathway/zhusf/CancerInformatics/Supplemental2006.html
Table 3: For each type of cancer, the top implicit associated gene specific to this cancer by 3MAM Cancer Type Gene Name BLADDER GAGE6 BRAIN MMP17 BREAST ZAP70 CERVIX PTGER1 COLORECTAL CYP1A1 ESOPHAGUS TMEM4 KIDNEY SDHD LARYNX TPR LEUKEMIA NOTCH4 LUNG CST6 LYMPHOMA LMO1 MELANOMA MYO5A MYELOMA PDCD2 ORAL FOXO1A OVARY INHBB PANCREAS NR5A2 PROSTATE KLK10 STOMACH FUT6 TESTIS GAGEB1 THYROID SSTR3 BODY_OF_UTERUS TIMP2
Methods
To explore the co-occurrence of cancer and genes in literature, let [math]\displaystyle{ G }[/math] be an observable random variable taking values [math]\displaystyle{ g_1, g_2, \cdots, g_S }[/math], each of which stands for a specific gene, and let [math]\displaystyle{ C }[/math] be an observable random variable taking value [math]\displaystyle{ c_1, c_2, \cdots, c_T }[/math], each of which stands for a specific type of cancer. Similarly, let [math]\displaystyle{ Z }[/math] be a discrete-valued latent variable taking on values [math]\displaystyle{ z_1,\cdots,z_H }[/math], each of which corresponds to a latent cluster, where [math]\displaystyle{ H }[/math] is the number of clusters. Let [math]\displaystyle{ \theta }[/math] be a set of parameters for the model to be optimized in the learning process, and let [math]\displaystyle{ \pi }[/math] be a mixture parameter (ie weight) of a component of our model that the users can specify. Let [math]\displaystyle{ D }[/math] be a set of all examples.
Here we present the E-Step and M-Step for 3MAM. For further details on MAM, please refer to [2].
Paremeters are estimated to maximize the log-likelihood of data D:
[math]\displaystyle{ \theta^{ML} = arg \max_{\theta} \log p(D ; \theta) }[/math]
Specifically in 3MAM,
[math]\displaystyle{ \log p(D) = \pi_{CG} \sum_{i,j} \frac{N_{i,j}}{N_{CG}} \log \sum_h p(c_i|z_h) p(g_j|z_h) p(z_h) + \pi_{GG} \sum_{j,j'} \frac{M_{j,j'}}{N_{GG}} \log \sum_{h} p(g_{j}|z_{h}) p(g_{j'}|z_{h}) p(z_{h}) }[/math]
[math]\displaystyle{ + \pi_{CC} \sum_{i,i'} \frac{L_{i,i'}}{N_{CC}} \log \sum_{h} p(c_{i}|z_{h}) p(c_{i'}|z_{h}) p(z_{h}) }[/math]
E-step:
[math]\displaystyle{ p(z_h|c_i,g_j) = \frac{p(c_i|z_h) p(g_j|z_h) p(z_h)}{\sum_{h'} p(c_i|z_{h'})p(g_j|z_{h'})p(z_{h'})} }[/math]
[math]\displaystyle{ p(z_h|g_j,g_{j'}) = \frac{p(g_j|z_h)p(g_{j'}|z_h)p(z_h)}{\sum_{h'} p(g_j|z_{h'})p(g_{j'}|z_{h'})p(z_{h'})} }[/math]
[math]\displaystyle{ p(z_h|c_i,c_{i'}) = \frac{p(c_i|z_h)p(c_{i'}|z_h)p(z_h)}{\sum_{h'} p(c_i|z_{h'})p(c_{i'}|z_{h'})p(z_{h'})} }[/math]
M-step:
[math]\displaystyle{ \hat{p}(c_i|z_h) \propto \pi_{CG} \sum_{j} \frac{N_{i,j}}{N_{CG}} p(z_h|c_i,g_j) + \pi_{CC} \sum_{i'}\frac{L_{i,i'}}{N_{CC}} p(z_h|c_i,c_{i'}) }[/math]
[math]\displaystyle{ \hat{p}(g_j|z_h) \propto \pi_{CG} \sum_{i} \frac{N_{i,j}}{N_{CG}} p(z_h|c_i,g_j) + \pi_{GG} \sum_{j'}\frac{M_{j,j'}}{N_{GG}} p(z_h|g_j,g_{j'}) }[/math]
[math]\displaystyle{ \hat{p}(z_c) \propto \pi_{CG} \sum_{i,j} \frac{N_{i,j}}{N_{CG}} p(z_h|c_i,g_j) + \pi_{GG} \sum_{j',j''}\frac{M_{j',j''}}{N_{GG}} p(z_h|g_{j'},g_{j''}) + \pi_{CC} \sum_{i',i''}\frac{L_{i',i''}}{N_{CC}} p(z_h|c_{i'},c_{i''}) }[/math]
Conclusion
We mined OMIM database and MEDLINE to discover implicitly associated pairs of cancers and genes by applying a new probabilistic model, mixture aspect model(MAM), on the data of co-occurrence of cancers and genes, using OMIM and MEDLINE. By integrating the co-occurrence data from heterogenous sources, MAM outperformed the original AM in the cross-validation experiment. Here to aviod the difficulty of recognziing gene names, we actually used a subset of MEDLINE indexed in LocusLink. It would be interesting to extract gene-gene co-occurence data directly from MEDLINE.
References
[1] Giallourakis C, Henson C, Reich M, Xie X and Mootha VK.(2005) Disease gene discovery through integrative genomics. Annu. Rev. Genomics Hum. Genet 6:381-406
[2] Zhu S, Okuno Y, Tsujimoto G. and Mamitsuka H.(2005) A probabilistic model for mining implicit chemical compound-gene relations from literature. Proc. of ECCB2005 (Bioinformatics 21 Supplement 2): ii245-ii251
[3] Jenssen,T. et al. (2001) A literature network of human genes for high-throughput analysis of gene expression. Nat. Genet.,28:21-28.
[4] Hofmann T. (2001) Unsupervised learning by probabilistic latent semantic analysis. Machine Learning. 42:177-196.
[5] Bajdik CD, Kuo B, Rusaw S, Jones S and Brooks-Wilson A. (2005) CGMIM: Automated text-mining of Online Mendelian Inheritance in Man (OMIM) to identify genetically-assocaited cancers and candidate genes. BMC Bioinformatics, 6:78-84
[6] Hou T et al. (2005) CYP1A1 Val462 and NQO1 Ser187 polymorphisms, cigarette use, and risk for colorectal adenoma. Carcinogenesis, 26(6):1122-1128.
