I am trying to perform de novo structure prediction of a 1135 long protein sequence; can DMPfold and I-Tasser work well for de novo structure prediction of 1135 amino acid long sequence?
For DMPfold you might like to look at Figure 5B of the paper, which implies that performance starts to drop off after ~600 residues. Only proteins up to 500 residues were used in the training set, for example.
A protein of that length is almost certainly multi-domain. I would consider doing domain prediction; we have a tool DomPred, but many others are available. Then run the sequences of the domains individually and recombine them later.
I don't know about I-Tasser's ability to do this, perhaps see if they mention it in their papers.
Generally, but not always the I-Tasser algorithm is not in the same league as a covariant mutation-based analysis (psicov) and don't think many people would disagree with this. I don't do I-Tasser but I'm aware it is the algorithm of choice for rough and ready predictions.
Psicov based models are essentially de novo predictions, based on amino acid mutation patterns, which I assume are refined against known homologous structure in DMFold (inaccuracies can creep through). I suspect the starting point of I-Tasser is a known homologue and it's associated known structure.
Having said that you need alot of sequence data to make psicov covariance work (>8000 aligned sequences) and (strictly in my opinion) it will also depend on the speed of mutation. Higher eukaryote gene alignments could require much higher numbers of taxa in the alignment. The covariance value will give you the strength of the signal and values >0.7 should be used. I personally consider>0.5 to 0.7 borderline, which is fair enough because those are the generic thresholds in statistics. I don't know if you can access the underlying covariance values in DMFold, or moreover set the threshold.
So I-Tasser could outperform DMFold in given situations, but the underlying output of covariance based approaches should make it clear how well the approach has performed. Obviously if it performs well I'd go with that.
I would follow@jgreener 's advice on the follow up program.