Threading and fold recognition

The twilight zone problem

Homology modeling needs a template you can actually find by sequence, and that assumption breaks once pairwise identity drops toward the twilight zone below roughly 25 to 30 percent. When PSI-BLAST or an HHblits search returns no template above that band, the alignment is unreliable and a comparative model built on it will be wrong. At that point you switch from finding a homolog to recognizing a fold.

Fold recognition works because tertiary structure is far more conserved than sequence, so two proteins can share the same fold at under 15 percent identity. Threading exploits this by placing the query sequence onto every backbone in a library of known folds and scoring how well that sequence tolerates each structural environment. The scoring uses knowledge-based statistical potentials over residue contacts, burial, and secondary-structure compatibility rather than sequence similarity alone.

Profile-profile vs threading

HHpred and its engine HHsearch are profile-profile methods: they build a hidden Markov model from the query multiple-sequence alignment and align it against HMMs of known structures, which is dramatically more sensitive than sequence-to-sequence search. Phyre2 is built directly on this HHsearch backend plus predicted secondary structure. Pure threading, as in the LOMETS meta-server behind I-TASSER, goes further by also evaluating whether each query residue fits the physical environment of the template position it lands on.

Reach for these methods precisely when a confident homolog is absent: no significant BLAST or HHblits hit, or hits only into the twilight and midnight zones. Use straightforward homology modeling instead when you already have a clear template above about 30 percent identity, because it is faster and its alignment is trustworthy. I-TASSER then goes beyond template detection by reassembling threading fragments with replica-exchange Monte Carlo simulations and clustering the decoys with SPICKER.

Building and searching a profile

# HHsuite3: build a query HMM, add predicted secondary structure, search PDB70
hhblits -i target.fasta -d $HHDB/UniRef30_2023_02/UniRef30_2023_02 \
        -oa3m target.a3m -n 3 -e 1e-3 -cpu 4

addss.pl target.a3m target.a3m -a3m         # append PSIPRED secondary structure

hhsearch -i target.a3m -d $HHDB/pdb70/pdb70 \
         -o target.hhr -cpu 4 -p 20 -Z 250 -B 250

Query         target
Match_columns 168
No_of_seqs    94 out of 1873
Neff          7.4
Searched_HMMs 86792

 No Hit                              Prob E-value P-value  Score  SS Cols Query HMM Template HMM
  1 6f3a_A HAD-like hydrolase        98.7 3.1E-16 8.9E-21 118.4 14.2  152    6-166     2-159 (161)
  2 4wpk_A phosphatase, HAD superfam 97.9 5.6E-14 1.6E-18 104.1 12.8  148    9-165     4-158 (160)
  3 2h1t_B beta-phosphoglucomutase   96.2 2.2E-11 6.3E-16  92.7 11.0  141   12-160     7-152 (155)
  4 1zjj_A nucleotidase, DDDD motif  71.4 1.8E-02 5.1E-07  41.9  6.3  118   15-149    10-141 (144)

Read the Probability column first, not the E-value: it is HHpred's calibrated estimate that the template shares a true homologous fold with the query. Above 95 percent the fold assignment is essentially certain even at negligible sequence identity, 50 to 95 percent is a candidate that needs corroboration, and below 50 percent is speculative. Here hits 1 through 3 all report the same HAD hydrolase fold at over 96 percent, so the topology is confident while hit 4 at 71 percent is only suggestive.