Abstract: To solve the bottleneck of memory and running time problem in protein structure predicting with large-scale data set, a neural networks' distributed learning algorithm is studied. Based on the analysis of the previous methods of distributed learning algorithm in a distributed computing environment with a large number of processors, a new distributed neural network based on multi-agents, and a learning algorithm based on chips to work on the distributed neural networks are created, which approach the global optimum by competition from a group of agents that have the different training sample chips to process. It is proved that the new distributed learning algorithm is better on balancing the parallel computing efficiency and the success rate of learning compared with the previous algorithms. The experiment results demonstrate that this method is feasible and effective in practical applications when used in protein secondary structure predicting with large size of amino-acid sequence.