Antibodies - key molecules of the immune system - are increasingly used as ther- apeutic drugs. The variable domain of the antibody is responsible for binding to an antigen and contains polypeptide loops, each referred to as a 'Complemen- tarity Determining Region (CDR)'. Five of these loops can be modelled with acceptable accuracy but so far, the sixth (CDR-H3) remains more difficult for all but the shortest loops. Various methods have been attempted to model this CDR, and while accu- racy has

more »

... improved, more improvement is still needed. One such method is to apply machine learning in an attempt to discern which features of a sequence lead to what final conformation; such an approach was tested in 1995. Since that time, the number of structures available for learning has increased dra- matically, as have methods and technologies available for building and training neural networks. This work revisits the neural network approach to modelling and scoring CDR-H3 in light of these advances. We conclude that several common neural network architectures cannot accurately model CDR-H3 from sequence alone, but show reasonable performance in selecting a good candidate from a large set.