The computer system PROSPECT for the protein fold recognition using the threading method is described and evaluated in this article. For a given target protein sequence and a template structure, PROSPECT guarantees to find a globally optimal threading alignment between the two. The scoring function for a threading alignment employed in PROSPECT consists of four additive terms: i) a mutation term, ii) a singleton fitness term, iii) a pairwise-contact potential term, and iv) alignment gap penalties. The current version of PROSPECT considers pair contacts only between core (alpha-helix or beta-strand) residues and alignment gaps only in loop regions. PROSPECT finds a globally optimal threading efficiently when pairwise contacts are considered only between residues that are spatially close (7 A or less between the C(beta) atoms in the current implementation). On a test set consisting of 137 pairs of target-template proteins, each pair being from the same superfamily and having sequence identity </= 30%, PROSPECT recognizes 69% of the templates correctly and aligns 66% of the structurally alignable residues correctly. These numbers may be compared with the 55% fold recognition and 64% alignment accuracy for the same test set using only scoring terms i), ii), and (iv), indicating the significant contribution from the contact term. The fold recognition and alignment accuracy are further improved to 72% and 74%, respectively, when the secondary structure information predicted by the PHD program is used in scoring. PROSPECT also allows a user to incorporate constraints about a target protein, e.g., disulfide bonds, active sites, and NOE distance restraints, into the threading process. The system rigorously finds a globally optimal threading under the specified constraints. Test results have shown that the constraints can further improve the performance of PROSPECT. Proteins 2000;40:343-354.