Information Technology Reference
In-Depth Information
4 Gene Libraries for Avoiding Self
In the above analysis, it would seem that the AIS has optimised for creating a large
number of antibodies; clearly it is effective at avoiding self. Could gene libraries
provide a bias to assist negative selection; that is, make the creation process cheaper?
Certainly, if we change the fitness function to be purely the avoidance of self (ie the
success rate of antibody creation) then gene libraries indeed have a profound effect on
the cost of negative selection [1].
However, it is possible that this reduction in the cost of negative selection comes at
the cost of other desirable features. In order to investigate this hypothesis we used a
similar GA setup with simple AIS gene library individuals (3 libraries, 16 bits
(5+6+5), 6 bit r-contiguous matching) and measured both the efficiency of producing
detectors, and also the diversity of different detectors produced. As can be seen from
Figure 3 this 'pure' measure has the effect of reducing genome diversity: in other
words, one gets a high proportion of 'safe' (non self reactive) antibodies - but also a
large number of duplicates. Clearly there is a trade-off between coverage and cost of
creation.
Fig. 3.
Effect of using avoidance of self as a fitness function (self), as opposed to coverage
(antigen), combined (both) or simply using a random creation strategy. The left figure shows
that AIS individuals can evolve gene libraries with a far higher (36%) chance of producing
valid antibodies than one whose fitness function measures only coverage (antigens; 13%) and
far above random creation (5%). All differences are statistically significant (wilcoxon). In the
right figure, the 'self' AIS individuals have roughly half the diversity of the others (unique
number of antibodies; 470 cf 950 antigen, 983 random). All differences significant except
antigen/random.
5 Mapping Antigen
It is well known that many real proteins fall into “families” with similar
configurations, and that in general both the sets of self proteins and possible antigens
will come from a non-uniform probability distribution across the space of possible
conformations. The same general consideration is true for many real world datasets;
