Presentation on theme: "1 Tobias Kind FiehnLab at UC Davis Genome Center November 2006 Statistics evaluation and graphics with ChemAxon tools and Statistica and WEKA towards QSPR."— Presentation transcript:
1 Tobias Kind FiehnLab at UC Davis Genome Center November 2006 Statistics evaluation and graphics with ChemAxon tools and Statistica and WEKA towards QSPR and QSAR development Free Academic Licenses for JChem and Instant JChem provided by ChemAxon Academic License for Statistica Dataminer provided by Statsoft GNU general public license for WEKA provided by WEKA Machine Learning Project
2 Metabolomics - The science of the small molecules Compound Classes: sugars amino acids steroids fatty acids lipids phospholipids organic acids... Molecules under investigation 3D model of a molecule with surface plot Visit us!
3 Techniques and tools Analytical techniques (LC-MS, GC-MS, FT-MS, NMR, IR) BioInformatics, ChomoInformatics Liquid Chromatography LC-MS Gas Chromatography GC-MS BioInformatics and Cheminformatics Statistics (Statistica Dataminer) Open Source Tools
4 ChemAxon JChem has now PCA and PLS 1)Create new library with JCHEM Manager GUI (testcase here: fingerprints) 2)Exctract fingerprints and do dimension reduction with principal component analysis (PCA) with command line tool PCA.bat or pca.sh
5 ChemAxon JChem Principal Component Analysis (PCA) PCA -d "sun.jdbc.odbc.JdbcOdbcDriver" -u "jdbc:odbc:jchem-z" -l test -p test -P JChemProperties -q "SELECT cd_fp1, cd_fp2, cd_fp3, cd_fp4, cd_fp5, cd_fp6, cd_fp7, cd_fp8, cd_fp9, cd_fp10, cd_fp11, cd_fp12, cd_fp13, cd_fp14, cd_fp15, cd_fp16 FROM nci99 WHERE cd_id <= " -o PCA-scores.txt -t PCA-Eigenvalues.txt TimeThis : Command Line : run-pca.bat TimeThis : Start Time : Mon Nov 27 17:02: TimeThis : End Time : Mon Nov 27 17:19: TimeThis : Elapsed Time : 00:17: Start PCA by getting information from DB (here Access, but can be Oracle, Derby, MySQL) Test case chemicals from NCI DB PCA can be done from any descriptor, chemical fingerprints, BCUT etc. This is just a simple example made from the 16 standard fingerprints. Be sure only to select descriptors you want (and not the molecule ID) Problem here: A) JDBC extraction not tuned - DB extraction of values nearly 2 minutes. B) PCA calculation time too long - 15 minutes for a matrix x 16 The current PCA algorithm needs to be changed, its very inefficent (faster matrix routines exist for JAVA) Database extraction time with Statistica: 8 seconds. The same PCA with Statistica is finished in: 1 second (no joke – thats a factor of 1:900).
6 JChem PCA output The PCA results matrix is inverted and values *(-1) from Statistica. Problem: Currently no graphics. But multivariate statistics lives from graphics. Follwing simple graphic examples are made with Statistica or WEKA via DB query.
7 Following slides What could be in the future. or What can be done right now. Check the pretty comprehensive statistics link
8 PLS Machine Learning (KNN) Feature selection Tree model Neural Network Cluster Analysis Response curves Machine Learning and statistic tools We use Statistica Dataminer as a comprehensive statistics work tool. WEKA or YALE are free but (not yet :-) as powerful as the Statistica Dataminer.
9 Connection of a JCHEM molecule DB via JDBC with Statistica Time for query + copy of 4,000,000 values with 250k molecules 16 fingerprints = 8 seconds. Test system JChem 3.2 with MS Access with Statistica Dataminer 7.1 Dual Opteron 2.8 GHz
10 Statistica with JChem data
11 PCA Scree plot – determine optimal factors to retain Visible Step Four factors can be retained. The 16 dimensional space can be compressed into a 4-dimensional space. (Scree plot is not optimal here) Statistica Dataminer 7.1
12 Which of the 16 fingerprints are similar? Those who cluster together are similar (fp_11 and fp_14). The variables fp_5 and fp_16 influence factor 1 in the same way. Variables inside or near the center (0,0) have no discrimination power. Remember PCA is no cluster analysis! PCA Loadings plot – which variables are influential? Statistica Dataminer 7.1
13 PCA Scores plot – picture of the reduced dimensionality. The 16 fingerprints are compressed into 2D. We can use other high dimensionality descriptors for enhanced examples. Cases (molecules) which cluster together may have same properties or functional groups (depending on input). Here we see the KOW molecule set covers the whole NCI dataset based on 16 pfs. Statistica Dataminer 7.1
14 The 16 fingerprints are compressed into 3D. The KOWWIN test set does not cover the whole molecules space of important silicon containing molecules. You can also do an Overlap Analysis (compare two databases) within the all-new Instant-JChem. PCA Scores 3D plot – KOWWIN versus silicon compound test set Statistica Dataminer 7.1
15 Statistica – Random Forest Machine learning 1024-DIM FC descriptor space Chemical fingerprint descriptors generated with JCHEM GenerateMD GenerateMD performance 1800 molecules/second for 1024 dimensional fp On Dual Opteron 2,8 GHz (one core used only). Statistica generates all graphical output + SQL code
16 CART tree method for QSPR and QSAR Classification trees, boosting trees, random forest, regression trees and honest trees and adaptive trees – lots of wood and forests - did you hear about them?
17 Model classspecific model# Generalized Linear Models (GLM) General Discriminant Analysis1 Binary logit (logistic) regression2 Binary probit regression3 Nonlinear model Multivariate adaptive regression splines (MARS) 4 Tree models Standard Classification Trees (CART)5 Standard General Chi-square Automatic Interaction Detector (CHAID) 6 Exhaustive CHAID7 Boosting classification trees8 Neural Networks Multilayer Perceptron neural network (MLP)9 Radial Basis Function neural network (RBF)10 Machine Learning Support Vector Machines (SVM)11 Naive Bayes classifier12 k-Nearest Neighbors (KNN)13 Other machine learning techniques from Statistica Dataminer we use Most of them work for classification and regression
18 Now with open source datamining tool WEKA URL SQL Data Easy: enter DB URL, enter SQL statement, import data. Try free AquaStudio for SQL!AquaStudio Yellow = OK
19 WEKA - Machine learning algorithms in Java
20 WEKA – fingerprint visualization
21 Conclusions regarding statistics: JChem PCA and PLS output (Eigenvalues, scores, loadings) are provided only as textfile. More univariate and multivariate tools needed. 1)JChem PCA and PLS results must have graphical output. (They must) 2)JChem PCA must be made faster (factor ) by using math routines. 3)Integration into Instant-JChem would be good or ChemAxon provides enhanced bundled statistics tools. 4)Currently JDBC query from JChem to other statistical packages like WEKA or Statistica or R or MATLAB or YALE is perfect. Each package works best in the field it was designed for.