Motif recognition is a powerful homology based sequence analysis tool for clustering new protein sequences into different families based on characteristic motifs. Compared to BLAST, these approaches typically have lower false positive rates and can reveal more remotely related family members. However, the current motif databases do not cover all the sequences in protein sequence databases. One of the major reasons for the low coverage of motif databases is that there is only a small set of known member sequences available for constructing protein motifs for many gene families. I have designed a new algorithm, "mFISHER", to detect protein motifs from only 2-5 known member sequences by artificial evolution of given sequences based on a position specific PAM evolution model. Based on my test results on 160 motif families, the overall average recall rate or sensitivity (true/(true + false negative)) and specificity (true/(true + false positive)) are 88% and 95%, respectively. Compared with MEME (Multiple EM for Motif Extraction), mFISHER is better based on the recall rate, especially when only 2 or 3 members are available. Both approaches have the similar sensitivity. MFISHER is promising for constructing protein motifs when only a few known members.