Data Mining and Bioinformatics Some Challenges Qiang Yang

Data Mining and Bioinformatics: Some Challenges Qiang Yang, Computer Science and Engineering HKUST Thanks: HKUST RPC Project -Ben Niu, -Can Yang -Prof. Hannah Xue -Prof. W. Yu http: //www. cse. ust. hk/~qyang/ 1

State of Art: DM for Bio l We know how to classify biological sequences l SVM, Neural Nets, Decision Trees, Rules l We know how to cluster biological entities l Bi-clustering, K-means, hierarchical l We l know how to select features PCA, LDA, SVM-RFE http: //www. cse. ust. hk/~qyang/ 2

Data Mining: Challenges in Bio 1. Non-traditional Feature Selection l l 2. Explainable and Accurate Data Mining Methods l 3. NN, SVM Rules? Transfer Learning l 4. When the number of attributes >> number of samples? Highly imbalanced Can knowledge learned from one set of samples help data mining on another sample? Exploiting the network structure l Individual i. i. d type of classification vs social networks? http: //www. cse. ust. hk/~qyang/ 3

Challenge 1: Non-traditional Feature Selection: Question: which (few) genes lead to diseases? l l l # of attributes >> # samples Dataset name Train Test Attributes Source ALL-AML leukemia 38 34 7129 ‘Molecular Classification of Cancer: Class Discovery and Class Prediction by Gene Exressoin Monitoring’, Science, Vol. 286, 1999. Breast cancer 78 19 24481 ‘Gene Expression Profiling Predicts Clinical Outcome of Breast Cancer’, Nature, Vol. 415, 2002. Central nervous system 30 30 7192 ‘Prediction of Central Nervous System Embryonal Tumour Outcome Based on Gene Expression’, Nature, Vol. 415, 2002. MLL_Leukem ia 57 15 12582 ‘MLL Translocations Specify A Distinct Gene Expression Profile that Distinguishes A Unique Leukemia’, Nature Genetics, Vol. 30, 2002. Large B-cell Lymphoma 24 23 4026 ‘Distinct types of diffuse large B-cell lymphoma identified by gene expression profiling’. Nature, Vol. 403, 2000. # of positive << # negative l KDD CUP 2002 task 2 l Yeast Gene Regulation Prediction Train Test Positive set Negative set 122 2896 1489 Traditional feature selection methods fail: overfitting, singularity of covariance matrix http: //www. cse. ust. hk/~qyang/ 4

Non-traditional Feature Selection (2) l Some potential solutions l ‘Characterization of a family of algorithms for generalized discriminant analysis on undersampled problems’, Journal of Machine Learning Research. Vol. 6, 2005. l l l Singularity problem is solved by splitting the subspace into the regular and the irregular parts. Irregular part (null space) of the within-class scatter matrix is fully utilized to extract the discriminant info. ‘Two-Dimensional PCA: A New Approach to Appearance. Based Face Representation and Recognition’, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 26, 2004. l l l High dimensional data in 2 D arrays are projected directly onto the subspaces. Size of covariance matrix can be reduced significantly. Singularity is avoided. http: //www. cse. ust. hk/~qyang/ 5

Non-traditional Feature Selection (3) l Other approaches: l Manifold learning l l Manifold learning methods, e. g. , Isomap, LLE, maintain the local patterns of distribution during transform, Extract features suitable for k-NN classifiers Can be used to reduce the dimensionality of Bio. Data. Semi-supervised learning l l l What if we have 10% labeled data, but the rest 90% are unlabelled? Build clusters around the labeled samples. Samples in the same cluster are labeled as from the same class, assuming they follow the normal distributions. http: //www. cse. ust. hk/~qyang/ 6

Challenge 2: Explainable and Accurate Data Mining Methods l Current methods, such as SVMs, discriminant analysis, neural networks, are ‘black box’ models. l l The learned knowledge is hard to understand by biologists. Some potential solutions l l l Logic based method, e. g. , decision trees and variants may be better in giving the ‘IF-THEN’ like rules that explicitly define the epigenetic logics in cancer and stem cell development. DNA methylation rules can be learned by using SVM based recursive feature elimination and fuzzy logics. [Gene selection for cancer classification using support vector machines’, Machine Learning, 2002. ] http: //www. cse. ust. hk/~qyang/ 7

Epigenetic Analysis: A Case Study l Epigenetic events dominate the growth of cancer and embryonic stem cells l l Genes can be turned on/ off through Cytosine methylation or Histone modifications l l These two type of cells are of great importance The logics of DNA methylation underlie the cells’ behaviors l Traditional methods, SVMs, ANNs are l l l ‘black box’ models Knowledge are trained connection weights, or Support Vectors. Hard to understand for biologists Wish to Know: Methylation status of Cp. G sites l l Cp. G islands/ promoter regions in DNA sequence Cancer prediction http: //www. cse. ust. hk/~qyang/ 8

Adaptive Cascade Sharing Trees (ACS 4) Niu et al. 2007 (tmr’s talk) Objective: learn human understandable rules that define the epigenetic process in cancer and embryonic stem cells l Idea: l Adaptively partition the numeric attributes into a set of the linguistic domains, e. g. , ‘high’, ‘very high’, ‘Medium’, ‘Low’, ‘Very Low’ l l l Method: clustering Train a committee of trees to select the most salient features and predict by voting l Method: tree learning http: //www. cse. ust. hk/~qyang/ 9

ACS 4 method (2) http: //www. cse. ust. hk/~qyang/ 10

ACS 4 method (3) l Dataset: l l l 37 h. ESC, 33 non-h. ESC, 24 cancer cell lines, 9 normal cell lines. 1, 536 attributes Result l l l Just 2 attributes are enough to separate the 3 cell types No need of 40 attributes by using fisher’s score in [1]. Wet lab cost can be reduced by testing on 2 attributes only, instead of 40. Accuracy is better, except when compared with SVM, but SVM cannot tell us ‘why’. Rules can be easily understood to biologist to conceive new biological experiments seeking in wet lab proof. 40 attributes: [1] ‘Human embryonic stem cells have a unique epigenetic signature‘, Genome Research, Vol. 16, 2006 11 http: //www. cse. ust. hk/~qyang/

Challenge 3: Transfer Learning l In real life, data are hard to obtain l l l Can we leverage the knowledge learned in one task/domain/data set for prediction of another? l l l Biological experiments are expensive However, biological data are related Humans often do this: having learned one language, find it easier to learn another In Web mining, having learned to classify one web site, use the abstract knowledge to help classify another web site Challenge: can we leverage the knowledge learned from one data set to classify/cluster/predict another? http: //www. cse. ust. hk/~qyang/ 12

Transfer Learning (Examples) l Problem: how to Propagate the classification knowledge? l Difficulty: old and new data may have different distributions Night time period Day time period http: //www. cse. ust. hk/~qyang/ Time 13

Transfer Learning to Classify Web l [Dai, et al, 2007] 20 newsgroups (20, 000 documents, 20 data sets) Old comp. graphics (comp) comp. os. mis-windows. misc (comp) sci. crypt (sci) sci. electronics (sci) l SRAA New comp. sys. ibm (? ) comp. windows. x (? ) sci. med (? ) sci. space (? ) (A. Mc. Callum, 70, 000 articles) Old sim-auto (auto) sim-aviation (aviation) http: //www. cse. ust. hk/~qyang/ New real-auto (? ) real-aviation (? ) 14

Document-word co-occurrence [Dai, et al. 2007] Old Di Knowledge transfer New Do http: //www. cse. ust. hk/~qyang/ 15

Transfer Learning: Related Works l Semi-supervised Learning l [Zhu, Survey, Blum and Mitchell “co-training”, Nigam et al, “EM-based”, Zeng et al “clustering”, Joachims, “transductive”] l Distributions of training and test data are usually assumed to be the same l Multi-task Learning l [Caruana, MLJ] l l l multiple Dis exist Domain specific knowledge jointly learned to benefit each other. Focused on how multiple tasks helping each other l Semi-supervised Clustering l Same distribution assumption, but can be relaxed when must-links are few http: //www. cse. ust. hk/~qyang/ 16

Transfer to Classify Web l Co-clustering is applied between words and out-of-domain documents (new tasks) l Word clustering is constrained by the labels of in-domain (Old) documents l The word clustering part in both domains serve as a bridge http: //www. cse. ust. hk/~qyang/ 17

A Biological Transfer Problem l ‘Promoter prediction analysis on the whole human genome’, Nature Biotechnology, Vol. 22, 2004. l l Most of the promoter prediction programs are effective on individual chromosomes, e. g. , Chr 21, Chr 22, But inadequate to generalize to the whole genome scale l l only 65% of accuracy rate on average too low Can we build a unifying model for transferring the learned knowledge to other chromosomes l l l to predict across the whole genome? to cluster other genes and protein arrays? to classify related sequences? http: //www. cse. ust. hk/~qyang/ 18

Challenge 4: Exploiting the network structure l We are short of labeled data l l The matrix structures are very sparse if we only have several hundred samples and a huge number of attributes Classification accuracy cannot be improved much l l l Gene expression data: tens or low hundreds of samples, but tens of thousands of attributes (? ) Accuracy ~ less than 80% Challenge: can we leverage the network structure? http: //www. cse. ust. hk/~qyang/ 19

Social Network Mining Citation (Paper 2) Title Conference Name l Very large scale computational analysis of gene and social networks. l Social networks: a social structure made of nodes (individuals or organizations) tied by one or more specific types of relations. l Relations: financial exchanges, friends, web links, disease transmission (epidemiology), or gene interactions. l Small world phenomenon: chain of social acquaintances required to connect on arbitrary person to another arbitrary person anywhere in the world is generally short, five to seven separation steps are sufficient. l Centrality Eigenvector: measure the importance of node in a network. Author (Paper 1) • Collective Classification • Collective Recommendation http: //www. cse. ust. hk/~qyang/ 20

Social Net Mining: Engineering meets Science l ‘Empirical Analysis of an Evolving Social Network’, Science, Vol. 311, 2006. l l l A dynamic social network comprising 43, 553 students, faculty, and staff at a large University. Interactions between individuals are inferred from time-stamped e-mail headers recorded over one academic year and are matched with affiliations and attributes. Findings: l when two students are in the same class, they are on average 3 times more likely to interact if they also share an acquaintance l l Netflix Challenge and KDDCUP 2007 Blog Evolution (NEC Work) l Can we deduce the actors’ roles/functions/attributes from the topology of the network? http: //www. cse. ust. hk/~qyang/ 21

Using Network Structure in Biology l ‘Adaptive Response of a Gene Network to Environmental Changes by Fitness-Induced Attractor Selection’, Plos One, 2006 l The gene network is formulated as differential equations, given some initial state the network stabilized at some attractors, corresponding to the different cell types. l l The complex dynamics of the gene networks can explain the high diversity of the species. Given some perturbations, how will the state of the gene networks change to adjust the levels of gene expression to environment factors? l The dynamics of a gene network are described by differential equations, e. g. , a simplified network involving only two gene nodes is formulated as: l where m 1 and m 2 are the gene expression levels. l S(A) and D(A) are the rate coefficients of synthesis and degradation. They depend on A, which represents cellular activity. l g 1 and g 2 represent the noises in gene expression. http: //www. cse. ust. hk/~qyang/ 22

‘Adaptive Response of a Gene Network to Environmental Changes by Fitness-Induced Attractor Selection’, Plos. One, 2006. Given the initial condition, the gene expression levels stabilize at the attractors determined by the coefficients of equations. Real world gene networks can be much more complex, involving thousands of genes, leading to the complex patterns of attractors and cell activities. http: //www. cse. ust. hk/~qyang/ 23

Conclusions l I have listed four challenges 1. 2. 3. 4. l l Non-traditional Feature Selection Explainable and Accurate Data Mining Methods Transfer Learning Exploiting the network structure There are more… Solving these challenges requires biologists and computer scientists to work hand in hand http: //www. cse. ust. hk/~qyang/ 24