ARS A Sequence alignments
We compare different methods to search a database of non-redundant proteins (for details see protocol).
- This simple search finds 3763 sequence matches with an e-value better (smaller) than the default 10.
- Of these, 3120 have an e-value matching the default criterion for inclusion in the iterative BLAST.
- Some of the sequence matches occur twice with different alignments. The number of unique sequence matches is: 3513.
- The distributions of percent sequence identity (see <xr id="fig:ARSA_blastHistogram_id"/>) and e-Values (see <xr id="fig:ARSA_blastHistogram_eVal"/>) shows that there are many sequence matches between 20 and 40 percent sequence identity and that the majority of e-Values is around 10^-6.
- The blast search in big_80 finds only one matching pdb entry. However, this is partly due to the way of clustering used for big_80 (based on CD_hit), where long sequences are preferred over shorter ones.
|Runtime [s]||j = 2||j = 10|
|h = 0.002||280||2111|
|h = 10E-10||280||2040|
- Running iterative blasts takes a while (see <xr id="tab:ARSA_psiBlast_runtimes"/>). The more iterations, the longer the run-time. However, decreasing the inclusion threshold speeds up the process since fewer sequences are added to the profile.
|# matches||j = 2||j = 10|
|h = 0.002||6421||7554|
|h = 10E-10||7115||8366|
- The number of unique matches increases with more iterations (see <xr id="tab:ARSA_psiBlast_hits"/>). Notably, the number of unique matches increases with decreased inclusion threshold. Probably the PSSM built with stricter inclusion is more specific and therefore produces more significant hits.
- <xr id="fig:ARSA_psiBlastHistogram_eVal"/> shows the distribution of e-values for the different PSI-BLAST runs.
- The distribution shifts to higher (less meaningful) e-values with more iterations (compare orange vs. green and blue vs. violet distributions).
- It shifts to lower (more meaningful) e-values with more stringent inclusion cutoff (compare orange vs. blue and green vs. violet distributions).
- Likewise, <xr id="fig:ARSA_psiBlastHistogram_id"/> shows the distribution of sequence identities for the different PSI-BLAST runs.
- Changing the e-value cutoff (orange vs. blue, green vs. violet) hardly changes the distribution.
- Increasing the number of iterations moves the distribution to less sequence identity.
- These observations are as expected: More iterations can move the profile to become less significant for the protein family. A more stringent inclusion cutoff makes the profile more sensitive for the family.
- The sensitivity (ratio of the found true positives to all possible hits) and accuracy (ratio of true positive among all found hits) of the search is defined here by the hits among pdb sequences.
- All pdb codes in the same COPS (for more info see http://www.came.sbg.ac.at/resources.php) L30 or L40 group as ARS A (structural representative 1nl2) count as "true positives".
- For "sensitivity" all pdb codes in the same COPS L40 group as ARS A (29 structures) count as possible hits.
- All pdb hits together define the numerator for "accuracy".
|Sensitivity [%] in L40 / L30||j = 2||j = 10|
|h = 0.002||86 / 64||72 / 57|
|h = 10E-10||72 / 57||72 / 57|
- <xr id="tab:ARSA_psiBlast_sensitivity"/> summarises how the sensitivity changes with different parameter settings.
- The sensitivity is identical for almost all settings. Only the profile produced with a low number of iterations and default inclusion parameters retrieves more true positives.
|Accuracy [%] in L40 / L30||j = 2||j = 10|
|h = 0.002||35 / 61||47 / 82|
|h = 10E-10||40 / 62||24 / 38|
- <xr id="tab:ARSA_psiBlast_accuracy"/> summarises how the accuracy changes with different parameter settings.
- The accuracy is highest using the profile with default inclusion parameters and 10 iterations. This is because the fewest total number of pdb hits was found with this search, while still finding the true positives found by the other searches (no significant drop in sensitivity).
- This is not what I would have expected. I thought the accuracy would be highest in the searches with lower inclusion threshold. I also expected the accuracy to go down with increasing the number of iterations and to decrease less with a tighter inclusion threshold. -- This might be a result of the bad statistics.
|j = 2||j = 10|
|h = 0.002|
|h = 10E-10|
- However, looking at the results in detail (<xr id="tab:ARSA_psiBlast_pdb"/>) reveals that more stringent profiles (lower e-value inclusion cutoff) lead to more separation between the structural similarity groups. This also separates the true and false positives better. Increasing the number of iterations has a similar effect.