Dear Yong Jia,

The PSI-blast is used to find distantly related proteins, what you need is the results in the final iteration. The PSI-blast runs in the following five steps:

Hope my answer could help you to understand the use of PSI-blast.

Best wishes,

Luezhen

 @ Luezhen Thanks for the post. From my understanding, the candidate homologs identified in previous round should naturally be included and identified in subsequent iteration, could you explain the discrepancy between different iteration? And also, I suspect simply adopting the results from the final iteration may generate some false positive genes, especially in cases when the search is too exhaustive. Cautions on this have been mentioned in the PSI-BLAST tutorial on the NCBI web. However, I may be wrong, just trying to raise the question. Thanks. 

Yong

Dear Yong Jia,

Different iterations use different PSSMs. A PSSM is constructed using sequences found in the last iteration by default. So the results are different in these iterations. Why some previous sequences disappear in the latter iterations? The reasons might be the following:

Briefly speaking, sequences found in previous round is only used to construct PSSM.

As you mentioned, there are some pitfalls using the PSI-blast, and most of them are related to the issue of PSSM. A good PSSM helps you to find good distantly related sequences. A bad one will guide you to the wrong place.

The main source of false positives is the spurious amplification of sequences not related to the query. Once even a single spurious protein is included in a PSI-BLAST search above threshold, it will not go away.

In order to avoid these pitfalls, you couldn't just throw a query sequence to the PSI-blast, and let the program run many iterations until it converge. The attached image shows some solutions of these pitfalls. (This image comes from )

Best wishes,

Luezhen

I am very new to bioinformatics (so please correct me if my thoughts are wrong). However, this is what I thought concerning your first question:

Note that the PSSM is a POSITION-specific scoring matrix and therefor scores different positions differently. The first search uses a normal, non-position-specific scoring matrix (as far as I know). If the first search had found some sequences that are similar to your query sequence but not at all to the other results, this sequence has less impact on the constructed PSSM and might therefore be "outcompeted" in later iterations.

Imaging your first search returns five sequences. Four of these five sequences are very similar to your query sequence at positions, say A, B and E. However the fifth sequence is very similar to your query sequence at positions C and D. Now the first PSSM is computed and assigns higher scores to positions A, B and E (because they are better conserved through the results, than positions C and D) and lower scores to positions C and D. After several iterations, this might result in outcompeting the fifth sequence.

Did you ever try to do a PSI-blast on the NCBI database? It could be, that you would have to adjust your own database (to stay in the above example: you might want to add a few more sequences similar to the fifth sequence, so it is not outcompeted as easily).

Thea

 Thanks @Luozhen and @Kosche. So the take home message is to use the results from the final iteration, and be careful about possible sequence bias? Maybe even double check the structure of each sequence. Thanks.