Cytokines are an important class of proteins responsible for intercellular communication. The helical cytokines have a four-helix bundle fold, and they have remained largely intractable for sequence alignment methods due to their high evolutionary divergence. This paper presents a method that is specifically designed to recognize the helical cytokine fold in preprotein sequences such as full-length cDNA translations or transcripts predicted by gene finding methods. A protein fold is modeled by multiple sequence profiles, each representing a structurally conserved region. Nonstructural profiles are used to represent additional signals found in preprotein sequences. Profiles are connected by loop regions, each of a specified minimum and maximum length. A model for the helical cytokines is created by progressively improving a placement of four amphipathic helices onto training sequences. The sensitivity and specificity of the method are evaluated by a cross-validation procedure, demonstrating that cytokines with no intrafamily sequence similarity can be recognized. The method has been successfully used for the discovery of several new helical cytokines in the human genome.