Bioinformatics Research Overview Li Liao Develop new algorithms and (statistical) learning methods that help solve biological problems > Capable of incorporating

Bioinformatics Research OverviewLi Liao

Develop new algorithms and (statistical) learning methods that help solve biological problems

> Capable of incorporating domain knowledge

> Effective, Expressive, Interpretable

1Li Liao, SIG NewGrad, 09/29/2008

Motivations• Understanding correlations between genotype

and phenotype• Predicting genotype <=> phenotype• Some Phenotype examples:

– Protein function– Drug/therapy response– Drug-drug interactions for expression– Drug mechanism– Interacting pathways of metabolism


Bioinformatics in a … cell


Credit:Kellis & Indyk4Li Liao, SIG NewGrad, 09/29/2008

Projects– Genome sequencing and assembly (funded by NSF)– Homology detection, protein family classification (funded by a DuPont S&E award)

Support Vector Machines Hidden Markov models Graph theoretic methods

– Probabilistic modeling for BioSequence (funded by NIH) HMMs, and beyond Motifs finding Secondary structure

– Systems BioinformaticsPrediction of Protein-Protein InteractionsInference of Gene Regulatory NetworksPrediction of other regulatory elements

Pattern analysis for RNAi (funded by UDRF)

– Comparative Genomics Identify genome features for diagnostic and therapeutic purposes (funded by an Army grant)


PeopleCurrent members:

- Dr. Wen-Zhong Wang (Postdoc Fellow)

- Roger Craig (PhD student)

- Alvaro Gonzalez (PhD student)

- Kevin McCormick (PhD student)

- Colin Kern (PhD student)

Past members:

- Robel Kahsay (Ph.D. currently at DuPont Central Research & Development)

- Kishore Narra (M.S. currently at VistaPrint, Inc.)

- Arpita Gandhi (M.S. currently at Colgate-Palmolive Company)

- Gaurav Jain (M.S. currently at Institute of Genomics, Univ. of Maryland)

- Shivakundan Singh Tej (M.S.)

- Tapan Patel (B.S. currently in MD/PhD program at U Penn)

- Laura Shankman (B.S., currently in PhD program at U Virginia)




Hybrid Hierarchical Assembly

• Three types of reads: Sanger (~1000bp), 454 (~100bp), and SBS (~30bp).

• Assembly of individual types using the best suited assemblers.– Phrap, TIGR, etc. for Sanger reads– Euler assembler and Newbler for 454 reads– Euler short, Shorty for SBS reads

• Hybrid and hierarchical – Use longer reads as scaffolding to resolve repeat regions

that are difficult for shorter reads– Use contigs from shorter reads (pyrosequencing) as

pseudoreads to bridge gaps (nonclonable and hard stops) with Sanger reads.


Major Findings

• Hybrid hierarchical assembly is proved to be an effective way for assembling short reads

• Incremental approach to selecting ABI reads is more effective than random approach in generating high coverage contigs

• Staged assembly using Phrap is an effective alternative to the proprietary Newbler assembler.

Publications:Gonzalez & Liao, BMC Bioinformatics 2008, 9:102.


Blue lines are contigs generated from hybrid assembly11Li Liao, SIG NewGrad, 09/29/2008

Detect remote homologues

Attributes:- Sequence similarity, Aggregate statistics (e.g., protein families),

Pattern/motif, and more attributes (presence at phylogenetic tree).

How to incorporate domain specific knowledge into the model so a classifier can be more accurate?

Results:

- Quasi-consensus based comparison of profile HMM for protein sequences (Kahsay et al, Bioinformatics 2005)

- Using extended phylogenetic profiles and support vector machines for protein family classification (Narra & Liao, SNPD04, Craig & Liao, ICMLA’05, Craig & Liao SAC’06, Craig & Liao, Int’l J. Bioinfo & DM 2007)

- Combining Pairwise Sequence Similarity and Support Vector Machines for Detecting Remote Protein Evolutionary and Structural Relationships (JCB 2003)


Non-linear mapping to a feature spaceΦ( )

xi Φ(xi)Φ(xj)xj

L() = i ½ i j yi yj Φ (xi )·Φ (xj )


1 1 1 1 0

1 1 1 1 1= 3 0.5

0 1 1 1 1 = 3 0.1x =

y =

z =

Ham

min

g di

stan

ce

Tre

e-ba

sed

dist

ance

Data: phylogenetic profiles - How to account for correlations among profile components?

profile extension (Narra & Liao, SNPD 04)Transductive learning (Craig & Liao, ICMLA’05, SAC’06, IJBDM, 2007)


1 1 0 1 0 0 0 1 1

10.33

0.67

0.34

0.5

0.75

0.55

1 0.33 0.67 0.34 0.5 0.75 0.55

Post-order traversal



Sequence Models (HMMs and beyond)

Motivations: What is responsible for the function?

– Patterns/motifs

– Secondary structure

To capture long range correlations of bio sequences

– Transporter proteins

– RNA secondary structure

Methods: generative versus discriminative

– Linear dependent processes

– Stochastic grammars

– Model equivalence


TMMOD: An improved hidden Markov model for predicting transmembrane topology (Kahsay, Gao & Liao. Bioinformatics 2005)


Mod. Reg.Dataset

Correcttopology

Correctlocation

Sens-itivity

Speci-ficity

TMMOD 1(a)(b)(c)

S-8365 (78.3%)51 (61.4%)64 (77.1%)

67 (80.7%)52 (62.7%)65 (78.3%)

97.4%71.3%97.1%

97.4%71.3%97.1%

TMMOD 2(a)(b)(c)

S-8361 (73.5%)54 (65.1%)54 (65.1%)

65 (78.3%)61 (73.5%)66 (79.5%)

99.4%93.8%99.7%

97.4%71.3%97.1%

TMMOD 3(a)(b)(c)

S-8370 (84.3%)64 (77.1%)74 (89.2%)

71 (85.5%)65 (78.3%)74 (89.2%)

98.2%95.3%99.1%

97.4%71.3%97.1%

TMHMM S-8364 (77.1%) 69 (83.1%) 96.2% 96.2%

PHDtm S-83(85.5%) (88.0%) 98.8% 95.2%

TMMOD 1(a)(b)(c)

S-160117 (73.1%)92 (57.5%)117 (73.1%)

128 (80.0%)103 (64.4%)126 (78.8%)

97.4%77.4%96.1%

97.0%80.8%96.7%

TMMOD 2(a)(b)(c)

S-160120 (75.0%)97 (60.6%)118 (73.8%)

132 (82.5%)121 (75.6%)135 (84.4%)

98.4%97.7%98.4%

97.2%95.6%97.2%

TMMOD 3(a)(b)(c)

S-160120 (75.0%)110 (68.8%)135 (84.4%)

133 (83.1%)124 (77.5%)143 (89.4%)

97.8%94.5%98.3%

97.6%98.1%98.1%

TMHMM S-160 123 (76.9%) 134 (83.8%) 97.1% 97.7%




Li Liao, SIG NewGrad, 09/29/2008 22

Inferring Regulatory Networks from Time Course Expression Data(Gandhi, Cogburn & Liao, 2008)

Expression Profile Clustering

Binary heat map

Boolean network algorithm

K-mean

Documents

Bioinformatics Research Overview Li Liao Develop new algorithms and (statistical) learning methods that help solve biological problems > Capable of incorporating