Introduction to Bioinformatics

1
Introduction to Bioinformatics
Larry Hunter Center for Computational
Pharmacology Larry.Hunter_at_uchsc.edu 2817b
SoM 315-1094
2
Overview

• Bioinformatics in the post-genomic era
• Data sources tapping public science
• Some useful databases you should know
• Metadata and integrated sources
• Inference transforming data into knowledge.
• Principles and foundations
• Example applications
• Advice for your own efforts

3
Bioinformatics now

• Computational methods routine (although not quite
  as integrated as statistical analysis)
• High throughput assays increase the stakes.
  Pharma has undergone radical change!
• Competitive advantages
• Getting more out of your data
• Finding other relevant information faster.
• Exploratory, hypothesis-generating analyses

4
Some challenges

• Staying current which techniques are accepted?
  Are best?
• Finding external data and algorithms. Which web
  sites? Which commercial offerings?
• Managing your own data. Beyond the spreadsheets
  eyeballing
• Finding and managing collaborative relationships,
  especially with computer people.

5
Taking advantage of public data

• An enormous amount of high quality data is
  available free.

6
How to find public data

• Start with http//www.ncbi.nlm.nih.gov
• Consult the Jan 2001 issue of Nucleic Acids
  Research http//nar.oupjournals.org/content/vol28
  /issue1
• Metadata sites (listings of databases)
• The NAR issue has an associated site
• http//www.genome.ad.jp/kegg/kegg4.html
• http//bioinformatics.weizmann.ac.il/mb/molecular_
  biol_databases.html
• Commercial portals, e.g. www.biolinks.com,
  www.doubletwist.com

7
NCBI Ground Zero

• The National Center for Biotechnology Information
  is the first place to go. Sequences, structures,
  PubMed, taxonomy, medical genetics, etc.
• Spend some time learning all it has to offer.
  There are good online tutorials at
  http//www.ncbi.nlm.nih.gov/Education/ Look at
  the site map, not just the front page!
• Check out PROW (Protein Reviews on the Web), a
  journal/reference source at NCBI.

8
An Abundance of Specialized Data

• Gene sequences and protein structures are not all
  there is!
• Metabolic, regulatory and signaling pathway data
  is growing rapidly
• Carbohydrates, drugs, lipids, diseases,
  organisms, etc. all have their own public
  databases

9
(No Transcript)
10
(No Transcript)
11
Integrated data sources

• Like the data, the shear volume of databases can
  be overwhelming.
• Integrated systems offer organized summaries of
  diverse datasets.
• An excellent starting place for information about
  human genes are GeneCards http//bioinformatics.w
  eizmann.ac.il/cards/

12
(No Transcript)
13
Inference from biological data

• Goal is to move from raw data to meaningful
  conclusions.
• Examples detecting remote homologues,
  identifying coregulated genes, predicting binding
  affinities
• Broadly applicable computational techniques
  clustering, discrimination/regression density
  estimation

14
Clustering

• Begin with a set of instances (e.g. gene
  sequences, protein structures) and a distance
  metric.
• Create a collection of groups of the instances
  which are more similar to each other than they
  are to instances in other groups. Groups can be
  hierarchically clustered themselves
• Examples
• Building taxonomic trees from aligned sequences
• Identifying coregulated genes from expression
  arrays

15
Discrimination/Regression

• Induce a predictor of some aspect (the label) of
  an instance from other aspects. Numeric
  predictions are regression, class predictions
  discrimination.
• Beginning with a training set of labeled
  instances
• Produce a model which accurately predicts the
  labels of other (unlabeled) instances.
• Examples
• Protein secondary structure prediction
• Prediction of drug response from gene expression

16
Density estimation

• Produces a method for assessing the probability
  of an observation. Like a histogram
• Uses a set of observations (and, optionally, a
  distance function)
• Examples
• Recognition of members of protein families
• Evaluation of diversity of compound libraries

17
Particular applications

• Those broad computational approaches have many
  particular instantiations and applications
• Two examples
• Hidden Markov models for multiple sequence
  alignment and homologous family discrimination
• Analysis of gene expression array data
• Finding genes that vary significantly
• Estimating the number of clusters
• Finding high-order discriminators

18
Hidden Markov models

• Technique from speech understanding is now widely
  used in sequence analysis
• Good software and tutorials on the web
  http//www.cse.ucsc.edu/research/compbio/ismb99.tu
  torial.html
• HMMs infer unobserved states that influence the
  probability distribution of observed states
• Most common use is to model sequence families.
  Unobserved states are indel events.

19
HMM example

• States are amino acid distributions, insertions
  and deletions. Train with sequences from a
  single protein family. Resulting model
  recognizes other members of the family, including
  distant homologs

20
Training an HMMThe E/M Algorithm

• Start with random state and transition
  probabilities
• Use model to ?parse? training data, assigning
  each to most likely path
• Change the state and transition probabilities to
  reflect the assignments
• Iterate until convergence

21
Trained HMM is a very sensitive discriminator for
families

• HMM trained with 5 proteins has 20 fewer errors
  than single-sequence probabilistic
  Smith-Waterman using ten sequences gives 51
  fewer errors.
• Weighting Hidden Markov Models for Maximum
  Discrimination. R. Karchin and A. Hughey,
  Bioinformatics, 14(9)772--782, 1998.
• pfam.wustl.edu (2000 pretrained models)
• www.cse.ucsc.edu/research/compbio/sam.html

22
Expression Array Analysis

• Gene expression arrays are a popular new
  technique for assaying the expression level of
  tens of thousands of genes simultaneously
• Many problems arise in analyzing this data
• We are in collaboration with many groups for
  analysis, and developing our own tools

23
Expression Analysis Issues

• Identifying genes that changed significantly over
  a set of observations. How much change is
  enough? What's wrong with 2-fold?
• Estimating the number of expression clusters.
  How many groups of genes are there?
• Finding discriminators based on expression levels
  (e.g. for response to drugs)

24
Variation filters for expression

• Most basic question in array analysis still
  open Which genes changed significantly?
• Several challenges
• Unknown variances, so hard to compare small
  numbers of observations
• Noise is inversely correlated with signal
• Small (30) changes known to be biologically
  important.

25
A novel approach to detecting variable genes

• Assume most genes do not change, so use median
  variance as estimate of noise
• Ratio of gene (N-1)s2 to median s2 is c2
  distributed with N-1 degrees of freedom gene
  variance test
• Problem is that overall variance may be low, but
  one difference can be large
• Investigating statistics on distances.
• Use database for estimates of s2 for each gene

26
Estimating the number of clusters

• Required input for many clustering methods
  (e.g., k-means, SOM)
• Hartigan test Ratio of sum squared error for K
  vs. K-1 clusters is F distributed (if K K-1
  differ by a single split class)
• Assume multivariate Normal for each cluster
  -2logPr(K)/Pr(K-1) is c2 distributed.
• Monte Carlo cross-validation estimates, too.

27
Estimates of K on artificial data
28
Estimates of k using real data 3454 Rat genes, 5
chips (alcohol sensitive vs. not)
29
Discrimination Regression

• Clusters of co-expressed genes are interesting,
  but just a first step. Really want predictive
  models, gene networks, etc.
• Biology tells us that the predictors are likely
  to involve interactions more than linear effects.
• Traditional statistics is not strong in
  non-linear models, high order interactions, large
  datasets.

30
Linear models vs. MMCs SLAM
31
Some Advice

• A little bioinformatics is good for you!
• Know how to use web data resources
• Know the kinds of analyses that are possible
• Sequence and structure computations are
  widespread and (fairly) easy. Finding exons,
  remote homologies, structural domains, fold
  families, etc. are routine.
• Generic clustering, discrimination/regression and
  density estimation tools exist (neural
  networks...)
• Collaboration with bioinformaticians is no worse
  than with statisticians.... -)