I would like ensure that my reasoning is correct. Assuming that I know the aminoacids sequence of the protein of interest. I can't say anything about the structure looking only at the aminoacids sequence of this protein. But if I know this protein from another organism and the structure of this protein is known, then I can compare both of the sequences and conclude something, right? what I mean, is that there is no specific sequence corresponding to, for example, helix-two- turn-helix motif, and that I can take this sequence, check if my protein has it and say that there is helix-two- turn-helix motif or not. I can do this only by comparison to the protein which structure is already known, right?
It seems to me that you're asking about homology modelling. In that case, yes you need to compare your protein of interest to a protein (or proteins) of known structure. Homology modelling in a nutshell includes three (four?) steps: template identification/template alignment, modelling, quality assessment.
You start with finding a template for your modelling. This is usually done by sequence alignment, for instance BLASTing. Preferably you use multiple sequence alignment which more sensitively aligns conserved regions. You then want a template with as high sequence identity as possible (above 50 % usually produces models with about 1 Å RMSE  in main chain atoms. Avoid less than 30 % where modelling errors rapidly increase.)
There are then a number of different modelling strategies (wikipedia). But basically, they all aim to predict the structure of the conserved protein core as well as possible (which usually is what you're really interested in). Peripheral amino acids are more dynamic and more prone to evolution and are therefore more difficult to predict. Then, most importantly, you assess the quality of your model. This can be done by calculating violations of statistical potentials or physics based conformational energies (or using more advanced methods like multivariate regression methods). As in all modelling, this really is a most crucial step because prediction with a poor model is misleading and utterly useless.
If you don't find any template you could resort to the exciting field of De novo protein structure prediction, where the aim is to predict the structure from the amino acid sequence alone. I am not very familiar with their methods, but de novo prediction is hard (!). I don't remember any exact numbers but the number of conformations in a normal sized protein is astronomically large, which leads to great algorithmic and computational challenges. Additionally, without any reference sequence the model assumptions are greater than those of homology modelling. Although, I have heard that the field has been making great progress the last few years.
Edit: It struck me that you might be asking about protein fold recognition as well. There exists a large number of different tools and methods for recognizing and locating protein domains using the amino acid sequence as input. Many of them are available as web servers. For instance phyre which uses the amino acid profile and predicted secondary structures to search structure libraries. Threading based methods like MUSTER. A number based on Hidden Markov Models (HMMs) also exist. For instance FISH which uses structure anchored HMMs.