You might want to check out this cautionary paper, which claims
simple sequence-based features contain insufficient information to be useful for predicting protein-protein interactions, but that protein domain-based features have some predictive value.
Yes, there are quite a lot of sequence-based methods to predict PPI. Their accuracy varies widely, depending on the method used and the quality of training data. Did you have a particular protein family in mind?
I'd start at PubMed with a search for "protein-protein interaction prediction sequence" - currently 289 results (162 freely-available and 26 reviews).
Depending on what exactly you mean by "predict from sequence alone", you might want to try STRING. It allows you to query with a sequence (or more sequences if you like) in order to retrieve a protein interaction network. Just starting from sequences (and no IDs) it will thus allow you get a predicted interaction network. However, the underlying evidence is based on numerous other things than just the sequences that you queried with, and I am thus not sure if it is what you are looking for.
This is one of the starting points of predicting protein protein interaction : Server: http://csbg.cnb.csic.es/mtserver/ References: http://pdg.cnb.uam.es/pazos/mirrortree/ I like the way these people have start from a simple idea in 2000 and are still working on it. the last paper is 2010 talks about Context Mirror Tree.
There are so many different ways to predict PPI using only sequence information. the new methods have integrated Classification methods, Bayesian framework into their work. have a look at this paper for example:http://www.ncbi.nlm.nih.gov/pubmed/17360525
also look at this for specific organism: http://cic.scu.edu.cn/bioinformatics/predict_ppi/default.html
There are also works on predicting PPI combining 3D structure information with sequence.