1. Estimation of human endogenous retrovirus activities from expressed sequence databases
Merja Oja, Jaakko Peltonen, Sami Kaski
University of Helsinki and
Helsinki University of Technology

2. Human endogenous retroviruses
Human endogenous retroviruses (HERVs) are DNA sequences
A retrovirus becomes endogenous when it infects a germ-line cell
HERVs are remains from ancient infections
8% of the human genome is retroviral
Human endogenous retroviruses or HERVs for short
retrovirus (for example HIV is a retrovirus)
and is inherited by the children of the host
ancient infections = happened tens of millions of years ago

3. Proliferation of endogenous retroviruses
HERVs belong to transposable elements, in other words,
The present day HERVs are mutated and mainly unable to transpose
Some terminology: Elements with a common ancestor are grouped to families
HERVs have moved independently in our genome
several mutated versions of the originals

4. Why are HERVs interesting?
can express viral genes in human tissues
is this connected to disease?
some HERV derived normal human genes
may regulate close by human genes
offer recombination sites for genome reorganizations
Their presence in the genome may affect the function of nearby human genes
--- promoters, --- extra exons
There are A few known instances of HERV derived normal human genes
this means that some HERVs have adopted functions beneficial to us

5. Why measure HERV activities?
HERVs are expressed in disease tissues
what is their role?
HERVs are expressed in healthy tissues
What controls HERV activation?
we need genomic neighborhood and activity level
The connection between HERVs and disease is unknown.
Are there more HERV derived genes?
Is there a difference between healthy and diseased cases?
we need the DNA surrounding the HERV
(we can search for transcription factor binding sites)
We want to measure the activities of all HERVs to get a large data set for further analysis.

6. Current methods and problems
HERV expression can be measured with
RT-PCR (real time polymerase chain reaction)
retrovirus microarray
Problem: all these methods give activities for families
hundreds of sequences in a family
not know which of the are active
We want activities for individual HERVs
RT-PCR: accurate technique, the activity of one thing at a time
retrovirus chip: expensive, noisy, several HERVs at the same time.
individual HERVs, this is needed in order to analyze the control mechanisms and to find potential disease associated HERVs.
it is difficult to design probes that target only individual sequences within the family.

7. Our approach
We introduce a model for estimating HERV activities from sequence databases
Expressed sequence tags (ESTs)
ESTs are a snapshot of gene activity in a cell
dbEST contains million of human ESTs
Principle: activity based on the number of ESTs from the HERV
ESTs are short and noisy versions of the gene transcripts present in cells
(a snapshot of all mRNA)

8. Cross-talk problem
Counting the ESTs coming from one particular HERV is not easy
ESTs are noisy and HERVs are very similar to each other
EST could be coming from a number of HERVs
Cross-talk between HERVs
We model the uncertainty in the matching statistically
the noise level in ESTs can be larger than the sequence differences within a family
it may be hard to determine exactly from which HERV an EST stems from
Our model will give relative activity levels for the HERVs based on EST data.

9. Generative model for ESTs
We introduce a generative mixture model for the ESTs
The model mimics the actual biological generation process
each mixture component is a hidden Markov model

10. Hidden Markov Model Mixture
corresponds to one large HMM where
chooses one of the N HERV-specific sub-HMMs
low-quality end part.
We use the Baum-Welch algorithm to learn the whole mixture.
are estimates of the HERV activities.
HERV specific sub-HMMs
Probabilities of first transition = HERV activities

11. HMM mixture We constrain the model complexity by sharing parameters.
We use two heuristics to reduce computation time
days instead of years!
Sharing parameters
- all match states share parameters (emit correct nucleotide)
blocks
reduce the time by the factor of 1000.
To limit to only a relevant portion of the HERV sequence (HERV is much longer than the EST)
Use only some probable pairs and assume that all other EST-HERV pairs have zero probability
These heuristics don’t seem to have an effect on the results.
The heuristics reduce the computation time from years to days per iteration. Without it would take 45 years to compute one iteration.

12. Experiments with generated data
Can we reconstruct the activity profile from generated data?
Data:
small set of 91 real HERVs
from these using our HMM model.
We selected HMM-parameters and activity distribution (and thus the level of cross-talk) to be similar to those in real data.

13. No cross-talk approach
A simple heuristic BLAST-based approach assumes no cross-talk
BLAST is used to retrieve ESTs from dbEST
HERVs as query sequences
The BLAST-approach:
each EST is simply assigned to best matching HERV
HERV activity: count of ESTs
Our method is designed to tackle the cross-talk problem
BLAST is a sequence similarity algorithm, which is generally used to make database queries

14. Results with generated data
Our method is able to learn the activities fairly well.
What is more important the inactive HERVs are not mistaken as non-active and vice versa.
Our method and the BLAST approach perform about equally.

15. Results from generated data experiment
Both our method and the heuristic BLAST based approach work well
The BLAST approach works surprisingly well
The data contains less cross-talk than anticipated
The fast BLAST approach can be used in large experiments
Our method and the BLAST approach perform about equally.
this suggests it can be used for large tests where HMM training would be computationally too costly.

16. Experiment with real data
What are the real activities of HERVs?
from three HERV families.
HERVs obtained by RetroTector
dbEST by BLAST
(RetroTector is a program designed for finding retroviral elements from genomes.
It was implemented by are collaborators in Sweden. It contains their expert knowledge on retroviruses).

17. Results with real data
The method finds relative activities of the HERVs.
Each family has only a few active HERVs and many that are only slightly active.
The activity of most of them was earlier unknown.
Results seem reliable.
method.
were reoptimized for each bootstrap replicate; other parameters were kept fixed.

18. Biological results The activity levels of the
HERVs were previously unknown
The age of the element doesn’t explain activity
The intactness doesn’t explain activity
Model sequences are among the active elements, but not the most active
The activity levels of most of the HERVs were previously uknown.
This is truly new information.

19. Conclusions
We have introduced a generative model for estimating the activity levels of individual HERVs.
The model is able to learn true activities
A heuristic BLAST approach works surprisingly well
Our method is general; it can be applied to study activities in different conditions, etc.
Results with generated data prove that we can reliably learn the activities
The method is general; it can be used
to compare different conditions,
to study HERVs in other organisms
elements.