I work mostly on DNA genomic sequences, and have limited experience with and knowledge of protein structure and homology modeling. The purpose of this message to the forum is to request feedback and advice regarding my proposed use of I-TASSER as a solution to my research problem - from both practical and theoretical points of view. Thank you!
A. In one "test" species, I have ~800 domains predicted by HMMER3, to a profile HMM from Pfam.
But across ~ 50 species, the total number of predicted domain sequences is ~ 17K. (length range is ~17aa - 103aa, with majority in 35-45aa range).
This is my full dataset.
B. Prediction domain sequences show big variations in score, length and sequence composition
- consequently, my domain-based phylogenetic tree has too many branches with zero bootstrap support.
I need some independent way to verify sequence-based predictions.
C. In RCSB PDB database, there are 7 PDBs that contain my domain of interest. I trimmed these PDB file down to domain only coordinates. Then using UCSF-CHIMERA, I could see that even though pairwise sequence identity is as low as 17%, but 3D overlap seems good.
D. Therefore, I want to use protein structure to segregate my sequence-based predictions into 4 categories - full-length, partial-length, degraded and false positives, or something on those lines.
From the I-TASSER tutorial (https://www.youtube.com/watch?v=quF4dqLGKFM), my understanding is that the longest / most computationally intensive step is structure assembly via Monte Carlo simulation.
With that as background info, these are my questions for your feedback, please:
1. Since my ~ 17K sequences are domain-only sequences, rather than full-length protein sequences, will it make homology modeling quicker and more practically feasible than multi-day runs for each query, i.e. domain sequence?
2. Would restricting the search space to only 7 trimmed, template PDBs, containing info for only my domain of interest, be a valid way to drastically reduce time and computational cost? If yes, by what factor will this reduction be? These trimmed PDB templates are 40-50aa long.
3. If yes to 1 and 2 above, is there a way to 'batch' submit this comparison for all 17K domain sequences vs. 7 trimmed PDBs to COMET @ XSEDE with your help?
4. If just a reduced PDB search space (7 PDBs) is scanned for each of the 17K domain sequences, would it make sense to even attempt the COACH/COFACTOR step for functional annotation of the domain sequences? Because my goal is not to annotate domain sequences, only to verify if they are F-box sequences or not.
5. Since I am unsure about domain boundary delineation using sequence-based methods, I plan to extend each of my ~ 17K domain predictions by up to 10aa on each side, and then input them. Does that violate any logic of I-TASSER use?
6. Can the I-TASSER results be parsed to classify my domain predictions into full-length, partial-length, degraded and false positives?That is my ultimate goal.
Thanks, in advance, for your advice and suggestions regarding both practical constraints and theoretical considerations.