Soft Comput (2009) 13:1073–1080 DOI 10.1007/s00500-008-0380-2
ORIGINAL PAPER
Design principles of adaptive cellular immunity for artificial immune systems Hugo Antonius van den Berg
Published online: 31 October 2008 Ó Springer-Verlag 2008
Abstract Artificial immune systems (AISs) have been proposed as a new computing paradigm. This paper reviews design principles of adaptive cellular immunity, based on the immunological literature rather than the simplified mathematical models which have thus far dominated the development of framework for design, interpretation, and application of AISs. Keywords Artificial immune systems Adaptive immunity Design principles
1 Introduction An understanding of the apparent design principles of control and regulation in the cellular immune system has gradually been emerging in the past decade (for reviews of various aspects, see Refs. van den Berg and Rand 2007; Bongrand and Malissen 1998; Davis et al. 2003; Goldrath and Bevan 1999; Goldstein et al. 2004; Lanzavecchia and Sallusto 2000; Mu¨ller and Bonhoeffer 2003; Roncarolo and Levings 2000; Sansom 2000; Stevanovic´ and Schild 1999). Computer scientists who want to apply these insights to software engineering problems, building so-called artificial immune systems (AISs), find themselves confronted with a
bewildering plethora of molecules and cell types, which makes it difficult to glean the essential mechanisms from the accidental details amidst the dense thickets of molecular and cellular biology. The purpose of this paper is to describe and motivate the main design principles of adaptive cellular immunity and its control as found in mammalian immune systems. The aim is to provide a tutorial overview of the abstract structure of this arm of the immune system, as revealed by recent immunological research. This paper does not consider the very important question of whether this structure can operate in a feasible presentday computational architecture. The design principles set forth here may only be applicable to systems that are in some appropriate sense sufficiently large (say [106 autonomous agents). Nevertheless, it seems worthwhile to sketch AIS design in a system that is ‘‘large enough,’’ anticipating real-world systems of this magnitude.
2 Immunity in a computational system Immunitas means ‘freedom from undesirable influences’. Thus the term immune system refers to any system that safeguards such freedom, even in a non-biological setting. 2.1 The protected system
An earlier version of this work was presented as a position paper at the ARTIST Network for Artificial Immune Systems meeting held on 8th–9th November 2004. H. A. van den Berg (&) Warwick Systems Biology Centre, University of Warwick, Coventry CV4 7AL, UK e-mail:
[email protected]
Let S denote the system that is to be protected by the immune system. For the present purposes, it suffices to think of S as a dynamic, richly structured data environment that is vast and open. Interactions with the outside world lead to the creation of data structures in S that may at times interfere with its proper operation, in ways that cannot in general be anticipated. The task of the immune system is to
123
1074
neutralize data structures that are associated with operational difficulties. These troublesome data structures may be data aggregates per se, such as viruses, or they may be runnables which hamper normal operation. Where S is embodied (Stepney 2007), the data structures may be created in S as representations of difficulties experienced in the physical world (cf. Timmis et al. 2008). The essential point is that the data structures that are to be targeted by the immune system are associated with harm to the mother system S. The property of being associated with harm can to some extent be pre-specified: some features of data structures will invariably be indicative of an attack or disruption. This corresponds to innate immunity. The focus of this paper is on adaptive immunity, which protects against attacks whose hallmarks cannot be fully specified in advance. Immunity presupposes an ability to detect the presence of operational difficulties arising within S (and to classify their nature); this ability is assumed to reside within S itself. The immune system is aroused and directed by the ‘distress’ signals emitted by components of S that experience difficulties. The decision whether there is an emergency affecting operations is a task that naturally falls to the affected components within S itself (Matzinger 2001). Anomalies (new data structures) may merely reflect S’s appropriate response to changing circumstances; for this reason only S itself can decide whether an immune response is needed. Design principle I (Self diagnosis) System S is equipped with self-diagnostic capability that transmits a global signal (i.e. a signal accessible by all components of the immune system). This diagnostic capability decides whether an anomaly is illicit by detecting operational difficulties. This hinges on feedback signals arising from S’s interactions with the real world; thus embodiment of S is a crucial prerequisite (Stepney 2007; Stepney et al. 2004). While not strictly describing a property of the immune system itself, design principle I is nonetheless fundamental to adaptive immunity. 2.2 Detectors and data fragments The agents of immunity track down harm-associated datastructures, and to this end they need to have detectors, which could be based on any suitable classifier algorithm (see Lanzi 2008, for a recent review), the only prerequisite being that the detector algorithm returns, for every input pattern, a scalar recognition strength value (which will be typically very small for all but very few input patterns). For definiteness, the detectors will be taken to be TankHopfield type networks, each characterized by a particular weight matrix (Rojas 1996).
123
H. A. van den Berg
Design principle II (Detector repertoire) The immune system is based on a repertoire of different agents, each with its own particular weight matrix. Thus, each repertoire agent corresponds to one particular (random) fixed choice of the weight matrix, which is fixed for all time (thus, the weight matrix constitutes the identity of the agent). The number of detectors (and hence agents) in the repertoire is the repertoire size, generally a large number ([*106). The ‘detector repertoire’ is the totality of recognition ability distributed over these many agents. To seed the repertoire, the requisite number of weight matrices is generated at random. Although the detectors’ weight matrices are static, their activation properties are dynamic, as described below. The rationale behind design principle II is that each agent (with its associated, unique detector) recognizes some particular feature of a data fragment (or more precisely a certain collection of such features); the union of these features over the detector repertoire can be regarded as an essentially blind (random) attempt to cover the space of contingencies. It is this very ‘blindness’ that imparts robustness against subversion. The contrast between neural networks and AISs is essentially a role reversal of time and space: in a neural network the system evolves dynamically through a sequence of states, whereas in an AIS such states coexist all at once as a repertoire of static detectors. The learning is done by managing the activity of the agents, rather than adapting the parameters of the individual detectors. In terms of a familiar computational trade-off, this amounts to an exchange of processing power for storage capacity. A disadvantage is the need to sustain a large number of detectors, whereas an advantage is the ability to represent past experience without smoothing it out, as happens with an individual, dynamic neural network. The agents survey data structures in S. Accordingly, these data structures must be ‘cut up’ into fragments corresponding to the detector input size. This is the task of samplers. Design principle III (Local sampling) There are special data presentation agents called samplers which process data structures into data fragments of appropriate size for the detectors. Samplers are assigned to local subsystems (‘neighbourhoods’) of S, each sampler acting on data structures arising within its assigned neighbourhood. Again, the blindness of random sampling protects against subversion. Design principle IV (Selective data fragment presentation) A sampler discards most of the data fragments it generates, presenting only a small fraction to the repertoire agents.
Design principles of adaptive cellular immunity for artificial immune systems
The rationale behind design principle IV is discussed below in section 2.5. Design principle V (Global control of data fragment presentation) The activity of the samplers is governed by the global self-diagnostic signal. This last principle allows the samplers to attract the attention of the repertoire to problem areas. 2.3 Agent interactions and activation A basic event is the exchange of information between a sampler and a repertoire agent. Repertoire agents establish random contacts with samplers. Whenever a contact is established, the sampler presents the data fragments. The agent’s detector translates each data fragment into recognition strength value that depends on both the data fragment and the detector’s weight matrix. Design principle VI (Aggregate stimulus) The repertoire agent collects the recognition strengths for all data fragments presented by the sampler, and adds these values together. This sum is the stimulus registered for the data exchange. The repertoire agent responds when the stimulus exceeds an activation threshold value stored in the agent. Thus the agent does not respond to individual data fragment signals, but to an aggregate stimulus derived from all data fragments presented by a sampler during a given data exchange contact. The threshold is a private agentspecific value. It is not immutable, but adapts to stimulation (design principle VII). 2.4 Agent responses When the threshold is exceeded by the stimulus, the agent responds; a stimulus exceeding the threshold is said to be superliminal. The global self-diagnostic signal governs the nature of the response. Design principle VII (Threshold adaptation) In the absence of global signals that indicate operational difficulties, a repertoire agent responds to a superliminal signal by slightly incrementing its threshold value. The rationale is that repertoire agents adapt to data fragments associated with the normal mode of operation of S, so that stimuli derived from such data fragments do not evoke aggressive immune responses. At the same time, the small size of the adjustment ensures maximal sensitivity to salient stimuli associated with illegitimate data (van den Berg and Rand 2004a). The above process may result in an agent having a very low or very high threshold. In either case, the agent is unlikely to respond efficiently to salient stimuli: in the
1075
former case because its detector hardly recognizes anything, in the latter because the agent has become very insensitive. Deletion of such useless agents improves the overall efficiency of the repertoire (in terms of protection gained per computing power expended), as is discussed further in Sect. 3.1.2 below. The core task of the immune system is to hunt down harm-associated data structures. Design principle VIII (Immune response activation) In the presence of global signals that indicate operational difficulties, (i) a repertoire agent responds to a superliminal signal by requisitioning a fixed amount computational resources to scan S; and (ii) a sampler that delivers a superliminal stimulus increases its ability to form further contacts. Responding agents may be of various types, depending on S. Examples of agent types that are useful in many systems include killer agents that induce the destruction of data structures processed by the sampler by tagging them for destruction by some appropriate dedicated mechanism, and reporter agents that record the role of the tagged data structures within S’s operations. Superliminal signals received under a global ‘SOS’ (Gallucci and Matzinger 2001) signal endow agents with computational resources. The finiteness of the resources allocated means that the agents need subsequent activating contacts with samplers to proceed; this constitutes a quality control mechanism. There is also a mechanism that keeps the activation of samplers in check: Design principle IX (Regulatory agents) An agent that frequently makes threshold adjustments as defined in (VII) undergoes a type change and becomes a regulatory agent, which behaves as follows. In the presence of global signals that indicate operational difficulties, a regulatory agent responds to a superliminal signal received from a given sampler by decreasing that sampler’s ability to form further contacts. The regulatory agents effectively specialise in recognizing normal operations, promoting allocation of resources to agents that recognize salient (harm-associated) data fragments. The rationale is that a sampler which activates a regulatory agent is likely to emit a false positive stimulus. If this sampler were to activate a killer agent, inappropriate deletion of legitimate (useful, vital) data structures could ensue. The action of the regulatory agent thus reduces the likelihood of auto-immunity, harmful immune responses directed against legitimate data structures. Self diagnosis ultimately directs which agents among the initial repertoire undergo the conversion to the regulatory type. Insofar as the evolution of S is open-ended, there can be no a priori mechanism deciding which agents are to be regulatory when the repertoire is first generated (the weight matrix is immutable and assigned at random for
123
1076
each detector). This is another instance of blindness imparting robustness. 2.5 Additional design principles that improve immune efficacy Two further design principles allow the samplers to exert control over the immune response. The first allows the sampler to emphasize certain data fragments, and thereby certain sources of data fragments. Design principle X (Data fragment weighting) The sampler assigns a weighting coefficient to each data fragment, and transmits the coefficient together with the data fragment. The repertoire agent or regulatory agent multiplies each data fragment signal with the associated coefficient before adding all together. Data fragment weighting corresponds to the molecular concept of avidity (van den Berg and Rand 2004b). Design principle XI (Second signal) Besides the data fragments, the sampler transmits a second signal to the repertoire or regulatory agent which modifies the latter’s activation threshold. The second signal represents a pathway whereby the global self-diagnostic information is conveyed to the repertoire agents. Consider the number and variety of agents responding to a given attack. There may be a wide range of data fragments recognized by the detectors among the responder agents, or, at the other extreme, the response may be narrowly focussed on a single data fragment: responses can be ‘wide spectrum’ or ‘narrow spectrum.’ A detailed mathematical analysis, not pursued here, shows that the width of the responder spectrum can be regulated by modulating the degree of correlation between data fragment weighting coefficients and the second signal. This allows the samplers to modulate the immune response. Design principle IV emphasises the selectiveness of data fragment presentation to repertoire agents. The key quantity is the typical difference between a stimulus derived from a ‘normal operations’ background and a stimulus derived from aberrant operations. The larger this difference, on average, the greater the likelihood of an effective response (van den Berg et al. 2001). This difference increases when an agent is presented with fewer data fragments per data structure, i.e. when data fragment presentation is highly selective (van den Berg and Rand 2003). Unfortunately, high presentation selectivity entails a significant risk that an agent will be presented with not even a single data fragment from a given data structure. This immunovisibility problem is solved through parallelism. Design principle XII (Restriction) Samplers are divided into several classes, such that each repertoire agent or regulatory agent interacts only with samplers from a single class.
123
H. A. van den Berg
This restriction partitions the agent repertoire into as many subrepertoires as there are classes of samplers. The simplest way to implement design principle XII is as follows: a number of classes n [ 1 is chosen and each sampler is assigned arbitrarily, initially and immutably, to one of these n classes. Similarly, each agent is assigned to one of these n classes when the repertoire is first generated. Subsequently, samplers and agents are only allowed to exchange information when they belong to the same class (e.g. checking whether they have matching ‘‘class tags’’). Every immune response must wind down eventually. Lingering killing agents not only hog precious computational resources, they also pose a risk of collateral damage of legitimate data structures. On the other hand, if the activity of the responding agents decreases too rapidly, the disturbance may ‘flare up’ again. Thus, the resolution of an immune response must balance the risk of flare-ups against the investment of resources in an inappropriately on-going response. Design principle XIII (Resolution of the immune response) Killer agents not engaged in tagging data structures attack other killer agents with a rate governed by the global self diagnostic signal. It can be shown (van den Berg and Kiselev 2004) that this principle optimizes the ratio of killer agents to illegitimate data in S. Since this ratio depends on the severity of the disturbance (van den Berg and Kiselev 2004), the global self diagnostic signal should modulate the rate at which ‘idle’ killer agents attack one another. A final design principle ensures acquired immunity: Design principle XIV (Memory) Agents that have been actively involved in previous responses are less dependent on the second signal (principle XI). Such veteran ‘memory’ agents have become more autonomous, less dependent on global self-diagnostic evidence of operational difficulties. This enables the memory agents to respond before a disruption has had the chance to manifest itself extensively. In other words, if a particular problem recurs, memory agents will start routing the source of the trouble well before it starts to adversely affect S; this is the essence of acquired immunity. 2.6 Design constraints The various design parameters of the immune system are interrelated as follows. The cost of the working repertoire is proportional to the number of data fragments recognized, which scales as the detector capacity times the number of repertoire agents. The requirements of sensitivity and specificity tend to favour a larger repertoire of low-capacity detectors over a small repertoire with high-capacity
Design principles of adaptive cellular immunity for artificial immune systems
detectors. However, data fragment size imposes a lower bound on detector capacity. In its turn data fragment size is bounded below by the need to uniquely identify data structures (Burroughs et al. 2004). These various bounds thus impose a minimal (optimized) working repertoire size (Perelson and Oster 1979), and S clearly must be large enough to sustain this minimum. Optimizing immune efficiency amounts to maximizing the immune protection gained per unit of computational/memory resources allocated to the immune system. Increasing the computational investment above the minimum size (readjusting for an optimal trade-off between detector capacity and number of agents) will generally increase efficiency and reliability, but with diminishing returns as a coverage of ‘data fragment space’ gradually saturates. The extraction of data fragments from data structures poses a very hard encoding problem, which presents the same general difficulties as those encountered in genetic algorithms (e.g. Goldberg 1988). In a sense, the possibility of a general optimal solution is precluded by the openendedness of the problem. As regards design principle XII, it may be seen that restriction is neutral with respect to computational effort: that is to say, if the average number of data fragments per data structure decreases n-fold, the number of parallel subrepertoires must increase n-fold, giving the same amount of CPU time spent on data fragment scanning.
3 Discussion Living systems solve computational problems, employing design principles that may be transferrable to man-made computing. To borrow these principles, computer scientists must first master the underlying biology. Unfortunately, the AIS field exhibits a tendency to rely on simplified textbook accounts and to engage immunology only at ‘‘second remove’’ through mathematical models of some part of the immune system (de Castro and Timmis 2003; Timmis and Bentley 2002; Timmis et al. 2003; Nicosia et al. 2004; de Castro and von Zuben 2005). Moreover, as Stepney et al. (2004) argue, any emulation of a biological system is bound to lack the desired properties if it is built upon a sketchy, overly simplistic account of the biology. The aim of this paper was to present the design principles of cellular immunity in their own right, without explicit reference to the particular ‘wetware’ implementation of the immune system of our own body. The principles were abstracted from current immunological understanding with a view to applicability in systems engineering, and to delineating the simplest set of design principles that could still claim to reflect the essential workings of cellular immunology.
1077
3.1 Comparison to the paradigms of immunology Ideally, the immune system rapidly and reliably eliminates any spurious data structures without affecting the legitimate data structures. The usual concepts of sensitivity and specificity analysis can be applied to determine how far the actual implementation strays from this ideal. However, immunologists have traditionally discussed immune performance in terms of several paradigms; the following discussion establishes some continuity between the present design principles and these well-entrenched paradigms. 3.1.1 Self versus nonself The totality of legitimate data structures within S constitutes its ‘self.’ Harm-associated data then are ‘nonself’ or ‘foreign’. This self/nonself terminology will almost certainly become standard in the AIS field, given its prominence in the immunological literature. It is therefore important to realize that these terms are unfortunate and somewhat misleading: ‘self’ and ‘nonself’ unduly emphasise the provenance of the data structures at the expense of what really matters, which is the functional role of the data within S. The essential task of the immune system is to neutralize data structures that underlie aberrant behaviour of S. It is only of incidental importance whether these data were downloaded directly from outside (a virus) or arose within S in response to some disturbance (a tumour). Conversely, data structures of external origin may be quite benign or useful in S’s operations. Moreover, self/nonself poses a false dichotomy; it follows from design principles VII and X that (non)selfness is not absolute but a matter of degree (van den Berg and Rand 2004b), as expressed by Eq. (1) below. 3.1.2 Tolerance and negative selection Nonresponsiveness to self, known in immunology as tolerance, has been implemented here as design principle VII. Threshold adaptation is a dynamic never-ending process, and allows the repertoire to adapt itself to new, but legitimate, data structures, where the global self diagnostic capability decides upon the legitimacy of newly emerging data structures. This continual redefinition of self as viewed by the repertoire is congenial to the notion of ‘self assertion’ (Bersini 2003; Miller et al. 1998). As mentioned in Sect. 2.4 agents may end up having either very high or very low activation thresholds, and deletion of such agents improves efficiency by not wasting resources on essentially useless agents (deletion of the former is called positive selection in immunology; deletion of the latter negative selection; these terms refer to what the agent is being tested for, respectively: having a
123
1078
functional detector and being non-responsive to self). It does not really matter if a high-threshold agent is deleted or not; its high threshold renders it anergic (non-functional) anyway (only wasting storage capacity). When the self of S is non-changing, it may be appropriate to implement negative selection once and for all: subject each newly generated detector to a battery of self data fragments and retain if it proves sufficiently nonresponsive (Ayara et al. 2002; Gonza´lez and Dasgupta 2002). However, this approach is perhaps best suited to a simplified concept of detectors (such as bit-matchers); for the more sophisticated and realistic detector concept used in the present paper, design principle VII is more appropriate. 3.1.3 Danger Regulation in the immune system is conveyed through the global self diagnostic signal, together with the samplers, which form a diffuse, non-centralized control system. The global signal has so far been treated as a binary signal (normal/abnormal operations). However, the nature of the operational difficulty can be specified in more detail by the global diagnostic signal, and this can be used to tailor the response to the trouble at hand. In the natural immune system, this task is carried out by the cytokine network, whose operation incidentally suggests a new unconventional computation paradigm of its own (Hone and van den Berg 2007). Control engineers may wonder why immunologists have historically been slow to appreciate design principle I. However, the historical emphasis on ‘recognition’ in immunology supported a concept of immune surveillance, with immune cells attacking anything that stimulates their receptors (Matzinger 2001). This naturally led to the self/ nonself concept and the conceptual difficulties that surround it. Polly Matzinger reinstated the primacy of what she calls the ‘danger’ signal (the term ‘danger’ is also used in artifical immune systems, e.g. Aickelin and Cayzer 2003). In Dunkin (1999), Matzinger compares the dendritic cells (the samplers) to a sleeping sheepdog, which only wakes up when the sheep (the components of S) start to panic in response to some outside source of distress or danger. 3.2 The scope for artificial adaptive cellular immunity AIS research has tended to seek inspiration in B cell (antibody-driven) rather than T cell immunity in a strict sense (see Nicosia et al. 2004; Timmis and Bentley 2002; Timmis et al. 2003 for an overview). Since T cell immunity is directed specifically against intracellular pathogens, it is not immediately clear that the design principles of its regulation have an application in man-made computational structures.
123
H. A. van den Berg
Consider an open-ended computational system that continually creates new data structures which serve both as records (data per se) and new code tailored to the system’s needs. Such a system naturally tends to be modular and nested (Simon 1996). Indeed, modular redundancy in S’s architecture would be essential, given the immune system’s strategy of deleting structures that it deems to be harmful; the cure should not be worse than the ailment. This description suggests a self-contained system that essentially needs to maintain its operational integrity by itself (that is, without intervention by human software engineers, or only communicating with humans through a relatively low-capacity channel) in an intricate environment which presents challenges that are impossible to anticipate in detail. A biologist would say that such a system needs to display homeostasis (Neal and Timmis 2005). It will almost inevitably develop redundancies in its code, which may at some point become malignant and perhaps even metastase, spreading through the system. In such a setting, an aggressive trouble-shooting system that hunts down and eliminates suspect pieces of code and data is vital. The core design principles (I–IX) ensure that the classification of data data fragments as legitimate or illegitimate emerges as a dynamic property of the agent repertoire. Classification adapts dynamically as the repertoire is presented with new data fragments that do not appear to be associated with trouble according to the global self-diagnostic signal. The repertoire’s detectors are generated at random and not intrinsically biased towards some pre-defined definition of harm-association. Nor are individual repertoire agents capable of becoming more attuned to recognition of data fragments associated with trouble: each agent’s detector is fixed. Yet the repertoire as a whole adapts. Let the index i range over the agents in the repertoire. Consider a data fragment x and let yi(x) C 0 denote the signal emitted by the detector of agent i when presented with x. Assume yi(x) to be bounded: y^i ¼ supx yi ðxÞ and form a function li ðxÞ ¼ yi ðxÞ=^ yi : This function li is the membership function for the fuzzy set of data fragments recognized by agent i. Assume that agent i responds to presentation with data fragment x if the membership value li(x) exceeds a threshold value wi(t); the threshold value is specific for agent i and can vary with time t. Data fragment x’s membership of the set of data fragments that evoke an immune response is then given as maxi{max{0,li(x) - wi(t)}}. This yields a fuzzy measure of a data fragment’s association with normal operations: llegitimate ðx; tÞ ¼ 1 maxi fmaxf0; li ðxÞ wi ðtÞgg:
ð1Þ
The dependencies on x, t, and i in this expression reflect the emergence of a ‘world view’ at the repertoire level. The
Design principles of adaptive cellular immunity for artificial immune systems
combination of global self-diagnostic capability and repertoire constitutes a cognitive system in the sense proposed by Hershberg and Efroni (2001): its perceptual sensitivities arise dynamically out of interaction with S’s environment, rather than being preordained. This dynamic ‘world view’ may come into its own in a relatively recent kind of application, involving adaptive profiles. Consider a human user of data resources, such as the internet, who (implicitly) provides diagnostic information on the saliency of data that pass through the system. The artificial immune system can estimate user-specific saliency for data fragments derived from data previously unseen by the user; Eq. (1) estimates membership of the adaptive profile (with ‘salient’ for ‘legitimate’). This could be used to set up a ‘‘soft firewall’’ based on user-specific ranking of saliency. Furthermore, web servers might be equipped with samplers, presenting data fragments from locally held files to web-crawling agents dispatched from users elsewhere on the globe. An agent encountering a highly salient data fragment would then report back with its location, much like a a scout bee. Applying this scouting to the user’s private archive of data accumulated in the past, one obtains a life-long memory management tool, reorganising the archive in the light of the user’s most recent interests (cf. Czerwinski et al. 2006). Acknowledgments The author thanks Andrew Hone, David Rand, Colin Johnson, Mark Neal, Jon Timmis, Susan Stepney, as well as two anonymous referees for stimulating discussions and suggestions. Support from the ARTIST network is gratefully acknowledged: http://www.artificial-immune-systems.org/artist.ht.
References Aickelin U, Cayzer S (2003) The danger theory and its application to artificial immune systems. In: Timmis et al. [41], pp 141–148 Ayara M, Timmis J, de Lemos R, de Castro LN, Duncan R (2002) Negative selection: How to generate detectors. In: Timmis and Bentley [40], pp 89–98 van den Berg HA, Kiselev YN (2004) Expansion and contraction of the cytotoxic T lymphocyte response—an optimal control approach. Bull Math Biol 66:1345–1369 van den Berg HA, Rand DA (2003) Antigen presentation on MHC molecules as a diversity filter that enhances immune efficacy. J Theor Biol 224:249–267 van den Berg HA, Rand DA (2004a) Dynamics of T cell activation threshold tuning. J Theor Biol 228:397–416 van den Berg HA, Rand DA (2004b) Foreigness as a matter of degree: the relative immunogenicity of peptide/MHC ligands. J Theor Biol 231:535–548 van den Berg HA, Rand DA (2007) Quantitative theories of T-cell responsiveness. Immunol Rev 216:81–92 van den Berg HA, Rand DA, Burroughs NJ (2001) A reliable and safe T cell repertoire based on low-affinity T cell receptors. J Theor Biol 209:465–486 Bersini H (2003) Self-assertion versus self-recognition: a tribute to Francisco Varela. In: Timmis et al. [41], pp 107–112
1079
Bongrand P, Malissen B (1998) Quantitative aspects of T-cell recognition: from within the antigen-presenting cell to within the T cell. BioEssays 20:412–422 Burroughs N, Kesmir C, de Boer R (2004) Discriminating self from nonself with short peptides from large proteomes. Immunogenetics 56:311–320 Czerwinski M, Cage D, Gemmell J, Catarci T, Marshall CC, PerezQuinones M, Skeels MM (2006) Digital memories in an era of ubiquitous computing and abundant storage. Commun ACM 49:44–50 Davis SJ, Ikemizu S, Evans EJ, Fugger L, Bakker TR, van der Merwe PA (2003) The nature of molecular recognition by T cells. Nat Immunol 4:1–8 de Castro LN, Timmis JI (2003) Artificial immune systems as a novel soft computing paradigm. Soft Comput 7(8):526–544 de Castro LN, von Zuben FJ (eds) (2005) Recent developments in biologically inspired computing. Idea Group, January 2005 Dunkin MA (1999) A maverick researcher bucks the establishment. In: Arthritis today magazine, March–April 1999 Gallucci S, Matzinger P (2001) Danger signals: SOS to the immune system. Curr Opin Immunol 13:114–119 Goldberg DE (1988) Genetic algorithms in search, optimization and machine learning. Addison Wesley, Reading Goldrath AW, Bevan MJ (1999) Selecting and maintaining a diverse T-cell repertoire. Nature 402:255–262 Goldstein B, Faeder JR, Hlavacek W (2004) Mathematical and computational models of immune-receptor signalling. Nat Rev Immunol 4:445–456 Gonza´lez F, Dasgupta D (2003) Neuro-immune and self-organizing map approaches to anomaly detection: a comparison. In: Timmis et al. [41], pp 203–211 Hershberg U, Efroni S (2001) The immune system and other cognitive systems. Complexity 6:14–21 Hone A, van den Berg HA (2007) Mathematical analysis of artificial immune system dynamics and performance. In: Flower D, Timmis J (eds) Silico immunology. Springer, Heidelberg Lanzavecchia A, Sallusto F (2000) Dynamics of T lymphocyte responses: intermediates, effectors, and memory cells. Science 290:92–97 Lanzi PL (2008) Learning classifier systems: then and now. Evol Intel 1:63–82 Matzinger P (2001) Essay 1: the danger model in its historical context. Scand J Immunol 54:4–9 Miller JFAP, Kurts C, Allison J, Kosaka H, Carbone F, Heath WR (1998) Induction of peripheral CD 8? T cell tolerance by crosspresentation of self antigens. Immunol Rev 165:267–277 Mu¨ller V, Bonhoeffer S (2003) Quantitative constraints on the scope of negative selection. TRENDS Immunol 24:132–135 Neal M, Timmis J (2005) Once more unto the breach: towards artifical homeostasis? In: de Castro, von Zuben [15], pp 340–365 Nicosia G, Cutello V, Bentley P, Timmis J (eds) (2004) Third international conference on artificial immune systems, September 2004. Lecture Notes in Computer Science, vol 3239. Springer, Heidelberg Perelson AS, Oster GF (1979) Theoretical studies of clonal selection: minimal antibody repertoire size and reliability of self-non-self discrimination. J Theor Biol 81:645–670 Rojas R (1996) Neural networks: a systematic introduction. Springer, Heidelberg Roncarolo M-G, Levings MK (2000) The role of different subsets of T regulatory cells in controlling autoimmunity. Curr Opin Immunol 12:676–683 Sansom DM (2000) CD28, CTLA-4 and their ligands: Who does what and to whom? Immunology 101:169–177 Simon HA (1996) The sciences of the artificial, 3rd edn. MIT Press, Cambridge
123
1080 Stepney S (2007) Embodiment. In: Flower D, Timmis J (eds) Silico immunology. Springer, Heidelberg, pp 265–288 Stepney S, Smith RE, Timmis J, Tyrell AM (2004) Towards a conceptual framework for artificial immune systems. In Nicosia et al. [30], pp 53–64 Stevanovic´ S, Schild H (1999) Quantitative aspects of T cell activation—peptide generation and editing by MHC class I molecule. Semin Immunol 11:375–384 Timmis J, Andrews P, Owens N, Clark E (2008) An interdisciplinary perspective on artificial immune systems. Evol Intel 1:5–26
123
H. A. van den Berg Timmis J, Bentley P (eds) (2002) 1st International conference on artificial immune systems. University of Kent at Canterbury, September 2002 Timmis J, Bentley P, Hart E (eds) (2003) Proceedings of the 2nd international conference on artificial immune systems, September 2003. Lecture Notes in Computer Science, vol 2787. Springer, Heidelberg
