1. Telling self from non-self: Learning the language of the Immune System
Rose Hoberman and Roni Rosenfeld
BioLM Workshop May 2003

2. Understanding the Immune System
The Goals:
characterize the differences between the languages of self vs. non-self
explain (and predict) which self proteins (or regions of proteins) are auto-reactive, which proteins are highly allergenic, ...
create better predictors of immunogenicity
Possible applications:
vaccine development
treating auto-immune diseases
co-opt the immune system for cancer therapy

3. Focus on T cells Essential component of the adaptive immune system
kill virus-infected cells
stimulate B cells to produce antibodies
coordinate entire adaptive response
Amenable to sequence-based analysis
T cell’s recognize short amino acid chains

4. Specificity of T cells Through a process of DNA splicing each T cell
has a unique surface molecule called a T cell receptor (TCR)
recognizes a unique pattern
A T cell epitope
the region of an antigen capable of eliciting a T cell response
short peptide (amino acid chain) derived from a protein antigen.

5. Predicting Epitopes Many proteins are not immunogens
Even an immunogenic protein might have only one or a few T cell epitopes
We have millions of T cells, each of which recognizes only a few patterns
How can we predict epitopes?

6. Two Possible Constraints
Machinery for generating and displaying peptides
Many peptides will never even be presented to T cells
Process of maintaining self-tolerance
T cells should not attack cells displaying only peptides derived from self proteins

8. TCR-MHC-Peptide Binding

9. Modelling the Peptide Pipeline
Binding and cleavage databases
over 10,000 synthetic and pathogen-derived peptides
~400 MHC I and II alleles
Prediction methods
position specific probability matrices
neural networks
peptide threading
Large amount of data and body of research

10. Two Possible Constraints
Machinery for turning proteins into peptides
Many peptides will never even be presented to T cells
Self-tolerance
T cells should not attack cells displaying only self proteins

11. Self Tolerance
T cells originate in the bone marrow then migrate to the Thymus where they mature
Selection of T cells through binding to self MHC-peptides in thymus
Strong binders are killed (clonal deletion)
Remaining T cells are (usually) no longer self-reactive

12. Finding Immunogenic Regions of Proteins
Method 1:
learn to predict which peptides will be generated, transported, and bound with MHC molecules
Method 2:
learn to discriminate self from non-self and use these models to classify each possible peptide

13. Related Work
Compositional bias and mimicry toward the nonself proteome in immunodominant T cell epitopes of self and nonself antigens
Ristori G, Salvetti M, Pesole G, Attimonelli M, Buttinelli C, Martin R, Riccio P. FASEB J. 14, (2000).

14. Self-Reactive Protein
Multiple Sclerosis (MS) is caused by the destruction of the Myelin sheets which surround nerve cells
T cells erroneously attack the Myelin Basic Protein (MBP) on the surface of the Myelin cells
Well-studied protein; known which regions are immunogenic

15. Unigram Models Ristori et al created two sets (self/non-self)...
Human genomes
Microbial genomes (Bacteria/Viruses)
We created three sets...
Human
Pathogenic bacteria
Non-pathogenic bacteria

16. A Simple Self/Non-Self Predictor
For each window of size ~7-15
Calculate the probability that the subsequence was generated by each unigram distribution (running average)
The ratio of the two probabilities gives a prediction of the degree of expected immune response
Similar to Betty’s segmentation by ratio of short-range/long-range models

17. Prediction of IP Values (Ristori et. al.)

18. Pathogenic vs. Non-Pathogenic

19. A Simple Extension
Do amino acid physical and chemical properties have any predictive power?
bulkiness and hydrophobicity measures result in better predictions on MBP than self/non-self
but Ristori et al claim that their predictions are better than any previous work
Question:
which existing prediction model works best?

20. Where to Go From Here?
Understand relative performance and strengths/weaknesses
self/non-self modelling
more traditional epitope prediction methods
how to combine these methods
what is the right evaluation function?

21. Future Work features data modelling higher level n-grams
expression level of genes
exploit the differences between pathogen/non-pathogen as well as self/non-self
data
auto-immune proteins
epitopes of known pathogens, ...
modelling
more powerful than simple ratio of probabilities