Journal Club (March 24, 2008)

Pei J, Kim BH, Grishin NV. PROMALS3D: a tool for multiple protein sequence and structure alignments. Nucleic Acids Res. 2008 Apr;36(7):2295-300.

PROMALS3D generates accurate multiple sequence alignments using both sequence and structure information. PROMALS3D stands for "PROfile Multiple Alignment with predicted Local Structures and 3D constraints" and is available via web server: http://prodata.swmed.edu/promals3d

The earlier program PROMALS incorporates:

• database searching for homologs
• secondary structure prediction
• profile-profile alignment

PROMALS3D is PROMALS enhanced to use 3D structures. Other multiple sequence alignment programs have been enhanced with the ability to use structures, although the details differ:

• MAFFT server - can add in one or more pre-existing structural alignments in sequence form (aligned fasta) (max sequences >10,000!)
• T-Coffee server (Expresso/3DCoffee option) - the server will automatically find structures and align them for you (max seqences 50)

There are other web servers for these programs (at EBI, for example), but usually with fewer options exposed.

• initial clustering uses BLOSUM62 (limitation?)
  
  
• UniRef90 is UniProt filtered at 90% ID
  
  
• PSIPRED uses neural networks on profiles from PSI-BLAST
  
  
• PSI-BLAST search of structures is single iteration
  
  
• if structures mostly overlap, only the one with the better E-value is kept
  
  
• pairwise sequence profile-sequence profile alignments (incorporating the secondary structure predictions somehow) provide sequence constraints
  
  
• pairwise structure-structure or structure-sequence profile alignments provide structure constraints
  
  
• constraints integrated into consistency score (weighted structure constraints 1.5x relative to sequence constraints)

• benchmarks:
  • SABmark - pairwise reference alignments (consensus of SOFI and CE) within each group
    • twilight zone set - 209 groups, SCOP fold level
    • superfamilies set - 425 groups, SCOP superfamily level
  • PREFAB - 1682 multiple alignments, each containing 2 sequences with known structures (alignment consensus of FSSP and CE) and up to 24 homologous sequences; generally "easier" than SABmark
• scoring:
  • reference-dependent:
    • Q-score, number of correctly aligned residue pairs in test alignment divided by number of aligned residues in reference alignment
  • reference-independent:
    • GDT-TS (global distance test) fraction of residues that pair within some cutoff CA-CA distance; TS refers to average from cutoffs 1, 2, 4, 8 Å

SABmark-twi (209/10667) SABmark-sup (425/19092) (# mult alignments/# pairwise ref alignments) Method Q-score (max ≈ 0.71) GDT-TS Q-score (max ≈ 0.87) GDT-TS PROMALS3D (D + S) 0.602 0.264 0.805 0.417 PROMALS3D (F + S) 0.555 0.220 0.779 0.390 PROMALS3D (T + S) 0.540 0.249 0.766 0.412 PROMALS3D (D + F + S) 0.611 0.256 0.812 0.414 PROMALS3D (D + T + S) 0.603 0.264 0.805 0.421 PROMALS3D (F + T + S) 0.595 0.251 0.800 0.413 PROMALS3D (D + F + T + S) 0.616 0.260 0.812 0.420 3DCoffee (D + S) 0.574 0.252 0.802 0.421 3DCoffee (SAP + S) 0.553 0.222 0.786 0.390 Expresso webserver 0.508 0.206 – – PROMALS3D (D/2 + S) 0.475 0.198 0.716 0.364 3DCoffee (D/2 + S) 0.261 0.100 0.573 0.294 3DCoffee (D/2 + SAP) 0.255 0.095 0.572 0.289 --------- sequence-only methods --------- PROMALS 0.393 0.154 0.665 0.336 SPEM 0.326 0.124 0.628 0.318 MUMMALS 0.196 0.081 0.522 0.278 ProbCons 0.166 0.058 0.485 0.246 MAFFT-linsi 0.184 0.070 0.510 0.264 MUSCLE 0.136 0.056 0.433 0.233 T-Coffee 0.134 0.048 0.429 0.223 ClustalW 0.127 0.057 0.390 0.221 --------- structure-only methods --------- MUSTANG 0.550 0.230 0.779 0.404 PROMALS3D (D) 0.594 0.252 0.802 0.415

PROMALS and SPEM also use predicted secondary structure
The best scores in a column are bold
D: using DaliLite structural constraints
F: using FAST structural constraints
T: using TM-align structural constraints
S: using sequence information
SAP: using SAP structural alignments
D/2: using DaliLite alignments for half of the sequences
SAP/2: using SAP alignments for half of the sequences

• Q-score max < 1 because the benchmark consists of pairwise alignments that are not completely consistent within groups, whereas the output is a multiple alignment for the whole group
• unclear (to me) how multiple superposition methods are combined
• DaliLite generally gave the best structure constraints even though not used in constructing the reference alignments
• 3D-Coffee and Expresso default is SAP + S, unclear why Expresso webserver results are worse
• ½ structure alignment tests meant to better represent real-life data in which not all sequences have relatives of known structure; PROMALS3D outperforms 3DCoffee. Similarly, PROMALS outperforms T-Coffee. The authors reason that profile-profile alignments including secondary structure predictions (PROMALS) yield more useful constraints than pairwise sequence alignments (T-Coffee).
• structure-only much more successful than sequence-only (but of course, no good for sequences without structures)

• yes, for > 20% ID
• also creating a sequence profile for the 3D structure and then applying profile-profile alignment (darker shading) gives a better result

Method Set 1 (0.121/420) Set 2 (0.185/421) Set 3 (0.248/420) Set 4 (0.527/421) All (0.270/1682) PROMALS3D (D + S) 0.817 0.879 0.921 0.954 0.893 PROMALS3D (F + S) 0.745 0.850 0.896 0.947 0.859 PROMALS3D (T + S) 0.766 0.856 0.902 0.950 0.869 PROMALS3D (D + F + S) 0.818 0.886 0.919 0.952 0.894 PROMALS3D (D + T + S) 0.834 0.884 0.922 0.953 0.898 PROMALS3D (F + T + S) 0.794 0.875 0.909 0.952 0.883 PROMALS3D (D + F + T + S) 0.836 0.894 0.917 0.956 0.900 --------- sequence-only methods --------- PROMALS 0.570 0.771 0.875 0.946 0.790 SPEM 0.536 0.756 0.865 0.940 0.774 MUMMALS 0.457 0.693 0.834 0.939 0.731 ProbCons 0.428 0.672 0.826 0.936 0.716 MAFFT-linsi 0.443 0.681 0.826 0.938 0.722 MUSCLE 0.372 0.631 0.787 0.930 0.680 ClustalW 0.299 0.536 0.726 0.906 0.617

The total 1682 PREFAB alignments are divided into four semi-equal-sized sets according to sequence identity in the reference multiple alignment. The average sequence identity and number of alignments are in parentheses beneath the set names.

• conclusions much the same although the PREFAB tests are "easier"
• as %ID increases, adding structure information has less effect

More about the server:

• well documented
• many parameters are [optionally] adjustable
• I don't see anything about a maximum number of input sequences
• output includes intermediate results and the final alignment in three formats