BIOINF 3360 Computational Immunomics

Transcription

1 BIOINF 3360 Computational Immunomics Oliver Kohlbacher Summer Summary

2 TYOTK Things you ought to know! Caveat: These are just examples, not the full list of potential exam questions!

3 TYOTK: Immunology Basics Describe Jenner s experiment Describe Pasteur s first experiment Name the key organs of the immune system What is the role of lymphatic system? What is hematopoiesis? Explain the differences between innate and adaptive immunity What types of T cells are there and what is their role?

4 Jenner s Experiment In 1788, it was folk lore that milkmaids having contracted cowpox did NOT get smallpox Edward Jenner infected an 8 year old boy with cowpox Subsequent exposure to smallpox did not result in an infection Vaccine was invented (latin vacca means cow ) ) Protection from infection can be achieved by vaccination

5 Pasteur s First Experiment In 1881, Louis Pasteur had managed to culture cholera bacteria and had proven that injection of the bacteria into chickens causes disease Tried the experiment with an old culture (after his summer holiday) Chickens became ill but recovered Tried a fresh culture on the same chickens: they survived again ) Attenuated bacteria can be used for vaccination

6 Organs of the Immune System Schünke et al., Prometheus LernAtlas der Anatomie, Thieme, 2004, S. 50

7 Lymphatic System Lymphatic vessels connect the organs of the immune system They form a network similar to blood vessels in the whole body Lymph nodes are small structures containing specialized structures where immune cells are concentrated and encounter antigens transported through the lymphatic vessels

8 Hematopoiesis All blood cells arise from hematopoietic stem cells (HSCs) HSCs are pluripotent stem cells HSCs reside in bone marrow and generate leukocytes of the immune system as well as platelets and erythrocytes

9 Innate and Adaptive Immunity Immune System Innate (non-specific) immunity Anatomic barriers (Skin, mucous membranes) Physiological barriers (temperature, ph) Phagocytic barriers (cells eating invaders) Inflammatory barriers (redness, swelling, heat and pain) Adaptive (specific) immunity Antigen specificity Diversity Immunological memory Self/nonself recognition

10 T Cells There are two well defined subtypes of T cells T helper cells (T H ), also known as CD4+ cells Recognize antigens presented by other cells Recognition leads to activation Activation leads to secretion of cytokines then activating T C and B cells Cytotoxic T cells (T C ), also known as CD8+ cells Activated T C recognize antigens presented by other cells Activated T C proliferates and differentiates to cytotoxic T lymphocyte (CTL) CTLs can attack and lyse infected cells (CD = cluster of differentiation, a membrane glycoprotein used to differentiate cell types)

11 TYOTK: Immunology Basics Describe the structure of an antibody What is the proteasome? What does MHC class I look like? What are the key steps in MHC class I antigen processing? What are the key steps in MHC class II antigen processing? Differences in the structures of MHC class I and II? Differences between active and passive immunization? How does active immunization achieve lasting effects? What is the major histocompatibility complex?

12 Antibody Structure

13 The Proteasome A protease complex consisting of four rings of protein subunits with a central channel

14 MHC Class I Structure Extracellular space Cytoplasm

15 Major Class I Processing Steps

16 MHC Class II Antigen Processing

17 MHC Structure

18 Immunization and Vaccines Immunity can be achieved by either active or passive immunization Passive immunization is in many cases done by injection of preformed antibodies Treatment of, e.g., black widow spider bites, diphtheria, and snake bite Mostly no long-lasting effect Active immunization aims to elicit long-lasting immunity by activating lymphocytes Can be achieved by the administration of vaccines 18

19 19 Active Immunization - Vaccines Long lasting effects by activating lymphocytes (generation of memory cells)

20 20 MHC Alleles HLA region is highly polymorphic Every human has two copies of each of the three class I and the three class II genes Hence, every individual possesses between three and six distinct alleles for both MHC classes, which correspond to an equal number of different MHC molecules with distinct binding specificities

21 TYOTK: Machine Learning What is the epitope prediction problem? What does the SYFPEITHI database contain? How does one determine MHC-peptide IC50 values experimentally? What is a position-specific scoring matrix and how does one predict epitopes with it? Define sensitivity and specificity Define Matthews Correlation Coefficient What is a separating hyperplane and how can it be described mathematically Explain the largest margin idea

22 Epitope Prediction Problem Given an Ag sequence and a set of target alleles, which subsequences are epitopes? PSSVQTEKKKKSDGKIKKDEDRYKTRDLWNNFSYF PEITHIVIKESTVSINKQDNKKMELKLSSHDEALSF ASLIDGYFRLTADAHHYLCTDVAPPLIEHNIKNGCH GPICTEYAINRLRQEG 22

23 The SYFPEITHI Database SYFPEITHI contains only naturally processed ligands and T- cell epitopes First listing was published in 1995 and included a couple of hundred peptides The database currently contains over 4500 peptides known to bind MHC molecules (most human and mouse) SYFPEITHI is both a database and a prediction method The best source for naturally processed T-cell epitopes 23

24 IC 50 from Radiolabeled Peptides * +! 2 MHC +!! * 3 4 MHC 5! 1. Isolate MHC 2. Bind radiolabeled reference peptide to it 3. Add unlabeled peptide at varying concentrations 4. Separate bound/unbound peptides by ultrafiltration 5. Determine amount of radioactivity bound to MHC Sette et al., Mol. Immunol. (1994), 31,

25 Position-Specific Scoring Matrices PSSMs can be generated from aligned sequences from a certain MHC allele. A probability can be assigned to each amino acid a in each position p of the peptide: The probability p(a) of observing the peptide a 1 a 2 a p can be calculated as: However, we also need to consider the background amino acid probability. Lund et al., p

26 26 Sensitivity and Specificity Finding the balance between these four classes is often difficult In most cases, one can increase TP or TN at the cost of higher FN or FP This balance is expressed by the terms sensitivity and specificity Sensitivity expresses the fraction of positives correctly identified (not accounting for FP): increase it to pick up everything, along with piles of crap SE = TP/(TP + FN) Specificity in turn expresses the fraction of correct identification only: increase it to get correct positive classifications at the cost of missing some SP = TN/(FP + TN)

27 Matthews Correlation Coefficient As sensitivity and specificity are interdependent, one often uses Matthews Correlation Coefficient (MCC), which distills SE and SP into a single number MCC is in the range of It yields values of +1 for ideal classifiers 0 for random classifiers -1 for perfectly incorrect classifiers 27

28 28 Linear Separating Hyperplanes Suppose we have a hyperplane that separates positive and negative points. The points x on this hyperplane satisfy: Let d + == d - be the shortest distance from the separating hyperplane to the closest points of each class. The margin of the hyperplane is described as d + + d -. An SVM simply searches for the hyperplane with the largest margin! d + d -

29 TYOTK: Machine Learning Give the Lagrangian formulation of a linear support vector machine Explain the kernel trick Give three examples for commonly used kernel functions How can epitopes be encoded for machine learning using SVMS?

30 The Lagrangian Formulation We have a convex optimization problem (quadratic criterion with linear inequality constraints) An effective way to solve such problems is to introduce Lagrange multipliers. The constraints will be replaced by constraints on the Lagrange multipliers themselves, which are easier to handle. The training data will only occur in the form of dot products between vectors. This will finally give the following function to maximize: Some examples of Lagrange multipliers can be found at: 30

31 31 Example Kernel Transformation 1 x 2 x 2 φ =-1 =+1 x 1 =-1 =+1 x 1 φ(x 1,x 2 ) = (x 1,x 22 )

32 Kernel Functions A kernel function must fulfill certain properties, e.g. it has to be symmetric. Some examples of commonly used kernels: The polynomial kernel: The Gaussian radial basis function: A kind of two-layer sigmoidal neural network: 32

33 33 SVMHC Data Encoding The data was encoded using sparse encoding A = C = D = The peptide AC would be encoded AC = A 9mer is encoded by a vector containing 180 elements

34 TYOTK: MHC Class II What is the typical length of a class I epitope and a class II epitope? What is a binding core and how can it be determined? Sketch the overall idea of the SDA algorithm What are virtual matrices? What is the key idea behind leveraging? What is contained in the feature vector if you apply leveraging? What is the meaning of slack variables in support vector regression? Why is leveraging not useful on HLA-A2? How do you combine antigen processing with MHC binding predictions?

35 Epitope Length Distribution Endosomal cleavage produces longer peptides than proteasomal cleavage Class II epitopes are thus generally longer (13-18) than class I epitopes (8-11) Number of known class II epitopes length Data for all HLA-D* epitopes. Data taken from FIMM.

36 Binding Core Sette et al. identified the region required for binding of a peptide to murine I-A d Experiment Take full sequence Truncate C-terminus until binding ceases Truncate N-terminus accordingly Determine IC 50 for each peptide Sequence rel. affinity ISQAVHAAHAEINEAGR 1.0 ISQAVHAAHAEINE 1.1 ISQAVHAAHAEIN 1.0 ISQAVHAAHAE 0.6 ISQAVHAAHA 0.2 ISQAVHAAH <0.01 QAVHAAHAEINEAGR 1.1 AVHAAHAEINEAGR 0.5 VHAAHAEINEAGR 0.3 HAAHAEINEAGR AAHAEINEAGR <0.01 Sette et al., Nature (1987), 328, 395

37 SDA Algorithm Initialize binders, nonbinders Determine initial model Determine binders based on current model Determine optimal model for binders/nonbinders yes Model changed? no Done Mallios, Bioinformatics (1999), 15, 432

38 Virtual Matrices Each allele corresponds to a unique combination of independent pockets Pockets 4, 6, 7, and 9 are polymorphic Sturniolo et al., Nat. Biotechnol. (1999), 17, 555

39 Leveraging Classical MHC prediction Data from one allele Encode peptide as feature vector Train allele-specific model on this Leveraging Build multi-allele model Include features coding for MHC sequence Model learns interaction between peptide features and MHC sequence features Allows the construction of models for alleles without experimental data!

40 NetMHCpan HLA Pseudo-Sequence Contact map: which MHC residue is in contact with which peptide residue Nielsen et al., PLoS ONE (2007), 2(8): e796. doi: /journal.pone

41 Support Vector Regression (SVR) Sparse solution:

42 TYOTK: Vaccine Design What are promiscuous epitopes? Define the epitope selection problem Why are MHC allele frequencies important for vaccine design? What is the general form of a linear programming problem? Name an algorithm for solving LPs and give its core idea What is the objective function in vaccine design? Formulate the epitope selection problem as an ILP

43 Promiscuous Epitopes A promiscuous epitope is a peptide that can be presented by multiple MHC alleles They are interesting candidates for vaccine design: Promiscuous epitopes cover a broader range of the population than epitopes binding to just one of the alleles Identifying promiscuous epitopes is rather straight-forward: Predict binding for a range of different MHC molecules Find those peptides with a high score for more than one allele

44 Vaccine Design As discussed earlier, most vaccines are based on attenuated pathogens Recently, epitope-based vaccines have become an interesting alternative

45 MHC Allele Frequencies Different populations have very different allele frequencies! Data from

46 Linear Programming Many optimization problems can be formulated as linear programming (LP) problems Linear programming problems are optimization problems with linear objective function and linear inequalities as constraints They can thus be written in the form maximize c T x subject to Ax b Another common notation for this is z = max {c T x : Ax b, x i 0} Lee et al., Brief. Bioinformatics (2006), 7:140

47 Algorithms The best-known algorithm for solving LP problems is the simplex algorithm introduced by George Dantzig in 1947 The simplex algorithm is quite involved in detail, so we will only discuss the basic ideas: Start at some vertex of the polytope defined by the LP constraints Move along these facets to an adjacent vertex Move only to vertices with lower or equal value of the objective function Terminate if no such vertex can be found G.B. Dantzig. Linear Programming and Extensions. Princeton, NJ: Princeton University Press, 1963.

48 ILP-based Vaccine Optimization Rewrite epitope selection problem as ILP Maximize overall immunogenicity for the target population Key assumptions: Immunogenicities are additive, across alleles & across epitopes Probability directly affects allele s contribution to overall immunogenicity Constraints Maximum number of peptides/epitopes Cover a large number of alleles Cover a large number of antigens Advantage: finds an optimal solution no heuristic!

49 ILP Formulation Overall formulation: Toussaint, Dönnes, Kohlbacher, PLoS Comput. Biol., 2008, 4(12): e

50 TYOTK: Systems Immunology What does Systems Biology mean? What type of omics data is commonly used in systems biology studies? What is the difference between high-throughput and low-throughput (classical) biological data? What mathematical models can be used to model hostpathogen interactions? What is a peptidome?

51 Systems Biology Wikipedia Systems biology is a relatively new biological study field that focuses on the systematic study of complex interactions in biological systems, thus using a new perspective (integration instead of reduction) to study them. Particularly from year 2000 onwards, the term is used widely in the biosciences, and in a variety of contexts. Because the scientific method has been used primarily toward reductionism, one of the goals of systems biology is to discover new emergent properties that may arise from the systemic view used by this discipline in order to understand better the entirety of processes that happen in a biological system. (06/06/2008)

52 'omics' - Data High throughput techniques provide data for one specific type of relationship Genomics: DNA sequence data Transcriptomics: mrna concentration Proteomics: Protein concentration/sequence Metabolomics: metabolite concentrations Interactomics: protein-protein interaction data

53 Models of HIV Infection Wodarz presented simple models for modeling the dynamics of HIV and the immune system. The equations used are simple predator-prey models that have been used for a long time in biology. (v) Wodarz, in Immunoinformatics: Bioinformatic strategies for better understanding of immune function.

54 Mathematical Models of HIV Modeling the viral load and immune response during therapy. Strong treatment Weak treatment

55 christophriesinger.de/?p=851