A neural network approach for the prediction of mitochondrial transit peptides (mTPs) from the malaria-causing parasite Plasmodium falciparum is presented. Nuclear-encoded mitochondrial protein precursors of P. falciparum were analyzed by statistical methods, principal component analysis and supervised neural networks, and were compared to those of other eukaryotes. A distinct amino acid usage pattern has been found in protein encoding regions of P. falciparum: glycine, alanine, tryptophan and arginine are under-represented, whereas isoleucine, tyrosine, asparagine and lysine are over-represented compared to the SwissProt average. Similar patterns were observed in mTPs of P. falciparum. Using principal component analysis (PCA), mTPs from P. falciparum were shown to differ considerably from those of other organisms. A neural network system (PlasMit) for prediction of mTPs in P. falciparum sequences was developed, based on the relative amino acid frequency in the first 24 N-terminal amino acids, yielding a Matthews correlation coefficient of 0.74 (90% correct prediction) in a 20-fold cross-validation study. This system predicted 1177 (22%) mitochondrial genes, based on 5334 annotated genes in the P. falciparum genome. A second network with the same topology was trained to give more conservative estimate. This more stringent network yielded a Matthews correlation coefficient of 0.51 (84% correct prediction) in a 10-fold cross-validation study. It predicted 381 (7.1%) mitochondrial genes, based on 5334 annotated genes in the P. falciparum genome.