APPLICATION OF NEURAL NETWORKS TO AUTOMATED ASSIGNMENT OF NMR-SPECTRAOF PROTEINS

Simulated neural networks are described which aid the assignment of pr otein NMR spectra. A network trained to recognize amino acid type from TOCSY data was trained on 148 assigned spin systems from E. coli acyl carrier proteins (ACPs) and tested on spin systems from spinach ACP, which has a 37% sequence homology with E. coli ACP and a similar secon dary structure. The output unit corresponding to the correct amino aci d is one of the four most activated units in 83% of the spin systems t ested. The utility of this information is illustrated by a second netw ork which uses a constraint satisfaction algorithm to find the best fi t of the spin systems to the amino acid sequence. Application to a str etch of 20 amino acids in spinach ACP results in 75% correct sequentia l assignment. Since the output of the amino acid type identification n etwork can be coupled with a variety of sequential assignment strategi es, the approach offers substantial potential for expediting assignmen t of protein NMR spectra.