Brief introduction to
Artificial Immune Systems.

Sławomir T. Wierzchoń

The aim of this section is brief presentation of the main mechanisms used by the immune system. A reader interested in a deeper review of natural immune system from a mathematical perspective is referred to [9] while computer-oriented treatment of the problem can be found in [2] or [13].

The basic building blocks of the immune system are white blood cells, or lymphocytes. There are two major classes of lymphocytes: B-cells, produced in the bone marrow in the course of so-called clonal selection (described later), and T-cells, processed in the thymus. B-lymphocytes are related to humoral immunity: they secrete antibodies. Among the B-cells are "memory" cells. Living relatively long and "remembering" foreign proteins they constantly re-stimulate the immune response of the organism. On the other hand, T-cells are concerned with cellular immunity: they function by interacting with other cells. T-cells divide into CD4 lymphocytes or helper T-cells, and CD8 lymphocytes, called cytoxic or killer T-cells, that eliminate intracellular pathogens. Generally, helper T-cells activate B-cells promoting their growth and differentiation into an antibody-secreting state. Activated B-cells cut protein antigens into smaller parts (peptides) and present them to killer T-cells. These last cells are responsible for killing virally infected cells and cells that appear abnormal. B and T cells see different features of pathogens, and as described, behave quite differently. Thus simulating B or T cells we obtain computer systems solving different tasks. Computer systems simulating T cells are used to broadly understood anomaly detection, e.g. computer viruses detection, while the systems simulating B cells behavior are oriented towards pattern recognition problems or optimization problems. Perhaps the first paper devoted to pattern recognition application of artificial immune system was [6]. The ideas presented in this paper was developed independently by de Castro and von Zuben, [3], and Timmis, [10]. All these approaches are concerned with unsupervised learning. A specialization of these approaches to the supervised learning was proposed in [12]. In the sequel we will focus on the principles which control B-cells population only. Some computer results are briefly sketched here.

B-cells synthesize and carry on their surfaces molecules called antibodies which act like detectors. The specialized portion of the antibody molecule used for identifying other molecules is called paratope. Being a 3-D structure with uneven surface the paratope have a unique shape which is referred to as the specificity. The regions on any molecule that the paratopes can attach to are called epitopes. If the two colliding molecules have complementary specificities, they bind to each other and the strength of the bond (called affinity) depends on the degree of complementarity. A molecule bound by an antibody is referred to as the antigen. A crucial role of the immune system is the binding of antibodies with antigens which serves to tag them for destruction by other cells. This process is termed antigen recognition. To treat formally the recognition problem, Perelson [8] introduced the notion of the shape space. Namely, if there are m features influencing the interaction between the molecules (i.e. the length, height, width, charge distribution, etc.) and D_i, i = 1, ..., m is the domain of i-th feature then a point in m-dimensional space S being the Cartesian product of D_i's is the generalized shape of a molecule. Typically S is a subset of m-dimensional Hamming space, or m-dimensional Euclidean space.

When a B-cell recognizes an antigen, it may be stimulated to clone (i.e. producing identical copies of itself) as well as to secrete free antibodies. This process of amplifying only those cells that produce a useful antibody type is called clonal selection. The number of clones produced by a lymphocyte is proportional to its stimulation level. Clones are not perfect, but they are subjected to somatic mutation (characterized by high mutation rate) that result with children having slightly different antibodies than the parent. These new B-cells can also bind to antigens and if they have a high affinity to the antigens they in turn will be activated and cloned. The rate of cloning a cell is proportional to its "fitness" to the problem: fittest cells replicate the most. The somatic mutation guarantees sufficient variation of the set of clones, while selection is provided by competition for antigens. The whole process of (in fact Darwinian) selection and differentiation of B-cell receptors leading to the evolution of B-cell populations better adapted to recognize specific epitopes is said to be affinity maturation.

Besides somatic mutation the immune system uses a number of other mechanism to maintain sufficient diversity and plasticity. Particularly about five percent of the B-cells are replaced every day by new lymphocytes generated in the bone marrow. This process is termed apoptosis.

The immune system possesses two types of response: primary and secondary. The primary response occurs when the immune system encounters the antigen for the first time and reacts against it. To learn the structure of the antigen epitopes, affinity maturation is used. The primary response can take some time (usually about 3 weeks) to destroy the antigen. If the body is reinfected with a previously encountered antigen, it will have an adapted subpopulation of B-cells to provide a very specific and rapid secondary response. Usually it is very fast and efficient. From a computer science perspective the primary response corresponds to the identification of clusters in the training data, while the secondary response -- to the pattern recognition problem, i.e. the assignment of a new data into one of existing clusters. Interestingly, the secondary response is not only triggered by the re-introduction of the same antigens, but also by infection with new antigens that are similar to previously seen antigens. That is why we say that the immune memory is associative.

The final immune system principle that plays a useful role in designing artificial immune system is that of immune network theory formulated by Jerne, [7], and further developed by Perelson, [8]. According to this theory (called also Jerne’s hypothesis) the immune response is based not only on the interaction of B-cells and antigens but also on the interactions of B-cells with other B-cells. These cells provide both a stimulation and suppression effect on one another and it is partially through this interaction that the memory is retained in the immune system.

The immune system is in permanent flux. The whole network is subjected structural perturbations through appearance and disappearance of some cell species. The introduction of new species is caused by somatic mutation, apoptosis, or combinatorial diversity (e.g. genetic operations). A crucial issue is the fact that the network as such, and not the environment, exerts the greatest pressure in the selection of the new species to be integrated in the network. Thus, the immune network is self-organizing, since it determines the survival of newly created clones, and it determines its own size. This is referred to as the meta-dynamics of the system, [11].

The model of immune memory proposed by Jerne resembles the models of hypercycles or autocatalytic sets considered in the context of prebiotic chemical evolution, cf. [1] or [4]. It seems that careful examination of these models may be of value in constructing effective data analysis algorithms.

References

[1] Bagley, R.J. and Farmer, J.D. Spontaneous emergence of a metabolism. In: C.G. Langton, C. Taylor, J.D. Farmer, S. Rasmussen, eds., Proc. of the Workshop on Artificial Life (ALIFE’90), vol. 5 of Santa Fe Institute Studies in the Sciences of Complexity, Addison-Weseley 1992, 93-140
[2] Dasgupta, D. (ed.) Artificial Immune Systems and Their Applications. Springer-Verlag Berlin-Heidelberg 1999
[3] De Castro L.N., von Zuben, F.J. aiNet: An artificial immune network for data analysis. In: H.A. Abbas, R.A. Sarker, Ch.S. Newton, eds., Data Mining: A Heuristic Approach, Idea Group Publishing, USA, 2000
[4] Eigen, M. Selforganization of matter and the evolution of biological macromolecules. Naturwissenschaften 58 (1971)465-523
[5] Farmer, J.D., Packard, N.H., Perelson, A.S. The immune system, adaptation, and machine learning. Physica D 22(1986)187-204
[6] Hunt, J.E., Cooke, D.E. Learning using an artificial immune system. J. of network and Computer Applications, 19(1996)189-212
[7] Jerne, N.K. Towards a network theory of the immune system. Ann. Immunol (Inst. Pasteur) 125C(1974)373-389
[8] Perelson, A.S. Immune network theory. Immunological Reviews, 110(1989)5-33
[9] Perelson, A.S., Weisbuch, G. Immunology for physicists. Reviews of Modern Physics, 69(1977)1219-1265
[10] Timmis, J.I. Artificial immune systems: A novel data analysis technique inspired by the immune network theory. Ph. D. Thesis. Department of Computer Science, University of Wales, Aberystwyth, September 2000
[11] Varela, F.J., Coutinho, A. second generation immune networks. Immunology Today, 12(1991)159-166
[12] Watkins, A.B. AIRS: A resource limited artificial immune classifier. M. Sc. Thesis. Department of Computer Science, Mississippi State University, December 2001
[13] Wierzchoń, S.T. Artificial Immune Systems. Theory and Applications (in Polish). Akademicka Oficyna Wydawnicza EXIT, Warszawa 2001. ISBN 83-87674-30-3, 282+vii pp.