Fold　recognition　is　an　important　issue　in　protein　structure　research.　The　Rossmann-fold　protein　that　has　typical　structure　is　a　common　kind　of　alpha/beta　protein.　The　training　set,　selected　from　22　families,　is　constituted　of　79　Rossmann-fold　proteins　which　have　less　than　25%　sequence　identity　with　each　other.　The　hierarchical　clustering　method　according　to　RMSD　is　applied　and　a　profile-HMM　based　on　structure　alignment　is　built　for　each　cluster.　Testing　on　9505　proteins　with　less　than　95%　sequence　identity　from　Astrall.65,　the　sensitivity,　specificity　and　MCC　are　93.9%,　82.1%　and　0.876　respectively.　The　result　shows　that　building　profile-HMMs　after　classification　could　reach　precise　fold　recognition　while　a　unified　one　cannot　he　built　due　to　there　are　too　many　members　in　training　set.