Task 2: Alignments
Sequence alignments are a good start to analyse a new protein sequence. Alignment search tools such as Blast, Psiblast or HHblits can be used to find related protein sequences.
Contents
Sequence Searches
The HFE protein sequence was used to conduct sequence searches with the three different search tools Blast, Psiblast and HHblits. CATH and COPS, two protein structure classification databases, were used to evaluate and compare the results of the different tools.
CATH is a database that hierarchically classifies all domain structures from PDB. This hierarchy consists of four major levels: 1. Class 2. Architecture 3. Topology 4. Homologous superfamily
We compared the first level of the CATH hierarchy (Class) to check whether the hits have the same fold class as the query. The class of a domain is determined by its secondary structure composition (mostly alpha, mostly beta, mixed alpha/beta or few secondary structures). Each domain of a protein is assigned a CATH class, this means, that one query protein can consist of several fold classes, one for each domain. The HFE protein consists of two domains that belong to the CATH fold class 2 and 3.
Another database for protein structure clustering is COPS. It consists of different structural groups that contain all proteins with a structural similarity above a defined threshold. For example, all proteins in the L30 group have a structural similarity of at least 30% percent. This means that if two proteins are not in the same L30 group, they have less than 30% structural similarity. The L30 groups of the hits were determined to check if they are in the same group as the HFE protein or not.
Blast
The Blast search in the big_80 database yielded 1504 hits.
<figure id="blast distribution">
</figure>
<xr id="blast distribution"/> shows the e-value and identity distributions of all unfiltered Blast search hits. There are only a few hits with a very low e-value of 1e-160, but most significant hits have an e-value between 1e-40 and 1e-50 (left small peak in a)). Most results have a worse e-value above 1e-20. The histogram of the sequence identity b) shows that nearly all hits have an identity between 20% to 40% to the query.
The Blast results only contain 13 PDB IDs. Nevertheless, the number of shared CATH classes between the reference sequence HFE and the hits were computed as described in the lab journal.
<css> table.colBasic2 { margin-left: auto; margin-right: auto; border: 2px solid black; border-collapse:collapse; width: 40%; } .colBasic2 th,td { padding: 3px; border: 2px solid black; } .colBasic2 td { text-align:left; } .colBasic2 tr th { background-color:#efefef; color: black;} .colBasic2 tr:first-child th { background-color:#adceff; color:black;} </css>
<figtable id="blast cath">
| number of same fold classes | frequency |
|---|---|
| 0 | 4 |
| 1 | 6 |
| 2 | 3 |
</figtable>
<xr id="blast cath"/> shows, that only 3 of the 13 PDB hits share all domain fold classes with HFE, 6 of the hits have one domain fold class in commom with the reference and 4 even none. Besides, only 6 of the 13 pdb hits are in the same COPS L30 group as the HFE protein. Those numbers indicate, that some of the Blast hits are not really related to HFE and thus, false positives.
Psiblast
Four different Psiblast runs against big_80 were executed using the following four parameter combinations:
- 2 iterations and default e-value cutoff 0.002 (j=2, h=2)
- 2 iterations and e-value cutoff 10E-10 (j=2, h=10E-10)
- 10 iterations and default e-value cutoff 0.002 (j=10, h=2)
- 10 iterations and e-value cutoff 10E-10 (j=10, h=10E-10)
The e-value and sequence identity distributions of the hits from the four runs were compared to identify the best parameter combination.
<figure id="psiblast boxplot">
 a) Boxplot of the log(e-value) of the four different Psiblast searches against big_80: 2 iterations with default e-value cutoff (0.002) and 10E-10 and 10 iterations with default e-value cutoff (0.002) and 10E-10. The thick black line inside the box is the median and the upper and lower end of the box the 25% and 75% quantils. The single points above the whiskers are outliers. |
 b) Boxplot of the identity of the results of the four different Psiblast searches against big_80: 2 iterations with default e-value cutoff (0.002) and 10E-10 and 10 iterations with default e-value cutoff (0.002) and 10E-10. The thick black line inside the box is the median and the upper and lower end of the box the 25% and 75% quantils. The single points above and below the whiskers are outliers. |
</figure>
<figtable id="psiblast summary">
| Psiblast parameters | |||||
|---|---|---|---|---|---|
| j=2, h=0.002 | j=2, h=10E-10 | j=10, h=0.002 | j=10, h=10E-10 | ||
| # hits | 2564 | 3163 | 2601 | 3343 | |
| # PDB hits | 66 | 72 | 56 | 66 | |
| CATH classes in common with HFE | 0 | 21 | 21 | 20 | 21 |
| 1 | 42 | 48 | 33 | 42 | |
| 2 | 3 | 3 | 3 | 3 | |
| # hits within same L30 group | 6 | 6 | 6 | 6 | |
</figtable>
<xr id="psiblast summary"/> summarises the results of the different Psiblast runs. The run with default parameter settings (j=2 and h=0.002) found the least hits. Nevertheless, the boxplot of the e-value distribution (<xr id="psiblast boxplot"/> a)) shows, that the default parameters can be considered the best: the median of the log(e-value) is smaller than the median of the other 3 psiblast runs. This indicates that the percentage of hits with a low and thus significant e-value is higher in the default psiblast run than in the other three. Besides, the default parameters also led to a higher percent sequence identity as indicated by the median of the cyan box in <xr id="psiblast boxplot"/> b).
The number of PDB hits varies among the different runs, but all found 3 proteins with the same two CATH classes as the query. There are 1095 proteins in the same COPS L30 group as the query HFE. Since the big_80 database does not contain all proteins from the PDB, the quality of the program cannot be assessed using the COPS L30 group of HFE. Nevertheless, from all found PDB hits, only 6 proteins belonged to the same L30 group.
PDB search
To get more PDB ids, we also did 4 Psiblast searches against the PDB seqres database using the profiles created with big_80. Those four result lists were compared as well.
<figure id="psiblast pdb">
 b) Boxplot of the log(e-value) of the four different Psiblast searches against PDB_seqres using the checkpoint files from the earlier runs. (2 iterations with default e-value cutoff (0.002) and 10E-10 and 10 iterations with default e-value cutoff(0.002) and 10E-10). The thick black line inside the box is the median and the upper and lower end of the box the 25% and 75% quantils. The single points below the whiskers are outliers. |
 Boxplot of the identity of the results of the four different Psiblast searches against pdb_seqres using the checkpoint files from the runs before. (2 iterations with default e-value cutoff (0.002) and 10E-10 and 10 iterations with default e-value cutoff(0.002) and 10E-10). The thick black line inside the box is the median and the upper and lower end of the box the 25% and 75% quantils. The single points above and below the whiskers are outliers. |
</figure>
<figtable id="psiblast summary">
| Psiblast parameters | |||||
|---|---|---|---|---|---|
| j=2, h=0.002 | j=2, h=10E-10 | j=10, h=0.002 | j=10, h=10E-10 | ||
| # hits | 4307 | 4814 | 3874 | 4058 | |
| CATH classes in common with HFE | 0 | 561 | 586 | 529 | 539 |
| 1 | 2798 | 3280 | 2397 | 2571 | |
| 2 | 948 | 948 | 948 | 948 | |
| # hits within same L30 group | 1047 | 1053 | 1045 | 1052 | |
</figtable>
The e-value and sequence identity distributions in <xr id="psiblast pdb"/> do not show a significant difference in the results, because the medians of all four e-value distributions are nearly the same. The last run with the parameter combination of j=10 and h=10E-10 has a higher percentage of low e-value hits, because the 25% quantil is below those of the other distributions. But the run with j=2 and h=10E-10 lead to more hits with a higher sequence identity, because the median slightly lies above the other medians. Apart from the query sequence, there were no PDB hits with a sequence identity above 40%.
Evaluating the CATH and SCOP annotations of the found hits revealed that all four searches found 948 hits with the same 2 CATH classes as reference. Besides, all runs could find nearly all 1095 proteins that are in the same COPS L30 group as HFE. Therefore, the decision which of the parameter settings leads to the best results cannot be based on the PDB hits, since the differences are minimal. Nevertheless, j=10, h=0.002 had the least hits, but the highest percentage of hits in the same L30 group as the query. It is thus best suited to find a high number of significant hits.
<figure id="psiblast venn">
</figure>
We also computed the hits overlap of the four different Psiblast searches in big_80, to see how comparable the found hits are. The overlap is illustrated in the Venn diagram in <xr id="psiblast venn"/>.
Additionally, we computed the overlap for the 100 best hits (based on e-value) of each of the runs (data not shown). They were exactly the same in all four runs. Thus, the parameter settings only matter if a high number of hits is needed, but not if onyl the best hits of a search are used in further analyses.
HHblits
We executed HHblits with default parameters and searched the uniprot20_02Sep11 database. HHblits is an alignment search tool that searches in a database of clustered sequences. A profile is generated from the query that is compared with the profile of each cluster. For each cluster, not only the representative, but also all other cluster members are reported in the results list. Thus, HHblits finds much more sequences than Blast and Psiblast. We found 2861 clusters and 37908 sequences in total. We used all sequences for the analysis, because one of HHblits' most important features is the high number of good results.
<figure id="hhblits distribution">
</figure>
<xr id ="hhblits distribution"/> a) shows the distribution of the log10 of the e-values of the HHblits hits. It is remarkable that HHblits finds so many hits with a very low e-value of 1E-130. One reason is, that the best hit belongs to a very large cluster. Nevertheless, in comparison to the Blast e-value distribution ( <xr id="blast distribution"/> a)), the percentage of hits with a high e-value is much lower in the HHblits results.
There are several hits with a very low sequence identity to the query (see <xr id="hhblits distribution"/> b)), but the distribution contains two peaks around 20% and 40% identity that correspond to distant and more closely related hits.
Comparison
<figure id="comparison">
</figure>
We used the results of the Psiblast run with 2 iterations and an e-value cutoff of 2e-3, because we also used default parameters for Blast and HHblits and because we also received the best results with this parameters. We compared the e-value and identity distributions. <xr id ="comparison"/> a) clearly shows that Blast has the lowest e-values, but the percentage of low e-value hits is higher in the HHblits results. The median sequence identity has the highest value among the Blast hits.
<figure id="comparison venn">
</figure>
The overlap between all hits and the top 100 hits are shown in <xr id ="comparison venn"/>.
<figtable id="best hits">
| search tool | rank of hit | |||||
|---|---|---|---|---|---|---|
| Psiblast | 1 | 2 | 3 | 4 | 5 | 6 |
| Blast | 1 | 3 | 2 | 52 | 109 | 81 |
| HHblits | 2 | not found | not found | not found | not found | not found |
</figtable>
We tried to identify the 6 best Psiblast hits with the lowest e-values in the results of the Blast and HHblits searches, see <xr id ="best hits"/>- The top 6 Psiblast hits are all present in the Blast results and the first 3 hits are the same. This confirms that Blast and Psiblast are very similar. HHblits, on the contrary, has definitely not the same results as Blast and Psiblast. From the top 6 Psiblast hits, only the first one is present in the second best cluster. This proves, that HHblits finds more distant related hits, which is also indicated by the low median sequence identity of the HHblits hits.
The best 5 hits from all three search tools belong to the MHC class I family and are thus correct hits.
Multiple sequence alignments
| < 30% sequence identity | > 60% sequence identity | whole range of sequence identity | |||||
| AC Number | Sequence Identity | AC Number | Sequence Identity | AC Number | Sequence Identity | AC Number | Sequence Identity |
| F6JYA9 | 0.29 | B4DDZ1 | 0.83 | B4DDZ1 | 0.83 | F6WCX4 | 0.34 |
| D9J389 | 0.28 | Q5EEZ1 | 0.76 | Q5EEZ1 | 0.76 | F6JYA9 | 0.29 |
| D5MSB3 | 0.27 | G3THV5 | 0.75 | G1MBW1 | 0.73 | D9J389 | 0.28 |
| B3FRK2 | 0.20 | G1MBW1 | 0.73 | H0VAR7 | 0.72 | D5MSB3 | 0.27 |
| 3ov6_A | 0.22 | H0VAR7 | 0.72 | F1PX48 | 0.71 | 2zok_E | 0.24 |
| 1p7k_L | 0.21 | F1PX48 | 0.71 | G1T7D7 | 0.70 | H0Y1D0 | 0.20 |
| H0Y1D0 | 0.20 | G1T7D7 | 0.70 | G5BQE5 | 0.67 | Q8SNJ4 | 0.22 |
| B3FRK3 | 0.18 | O35799 | 0.68 | Q95IT9 | 0.38 | B3FRK3 | 0.18 |
| Q8HWL2 | 0.17 | G5BQE5 | 0.67 | 2qrt_A | 0.38 | Q8HWL2 | 0.17 |
| Q8HX83 | 0.36 | ||||||
In order to assess the difference of multiple sequence alignments between close and related homologs, three different groups of sequences were selected. One containing only sequences with a sequence identity above 60% to human hfe, one with sequences below 30% identity and one with sequences covering the whole range of sequence identity. The selected sequences are listed in Table 5.
In order to ensure that also sequences with known structures are included in the alignments, the sequences from the following PDB structures were selected:
- 1p7k_L
- 3ov6_A
- 2qrt_A
- 2zok_E
ClustalW
<figtable id="clustalW">
</figtable>
<figtable id="clustalWSS">
</figtable>
MAFFT
<figtable id="mafft">
</figtable>
T-Coffee
<figtable id="tcoffee">
</figtable>
Expresso
<figtable id="tcoffee">
</figtable>
Discussion
In the above 60% sequence identity group, the residues and the gaps are well conserved, especially in the first third of the sequence. In the other two thirds, the sequence conservation drops slightly, but is still at a generally high level. In the below 30% identity group, the number of gaps inside the sequences is not much higher than in the above 60% group, but the overall residue conservation is significantly lower. Nevertheless, there are some very well conserved residues in this group that might be functionally important. In the mixed sequence identity group, the gaps are not as well conserved as in the other two groups and the conservation is even lower than in the below 30% identity group. But this effect is probably strengthened by the higher amount of sequences present. Another peculiarity is that the inclusion of 3D information with Expresso (figure 12), yields an alignment with a lot more gaps, but more conserved columns than the other programs. Although this could be a desirable property in an alignment, this behavior can be reproduced with the other programs by modifying the gap penalty parameters.
The different alignment programs yield comparable results, i.e. none of them is considerably better or worse than the other two. Nevertheless, some differences can be observed. The first notable one is that MAFFT yields a lot less conserved columns for the low sequence identity group than the other programs. Also, the positioning of consecutive gap columns varies between the programs.
Since there are only one to two bigger gaps in the alignments and the secondary structure segments are often not very long, it is hard to tell if the gaps lie between secondary structure segments or inside them (Figure 9). Based on the fact that the gaps do not disrupt a longer continuous secondary structure segment, one could argue that the gaps lie between ordered secondary structure segments. But given these specific alignments, there is not enough evidence to support this statement.
In our opinion, there are two distinct quality measures for multiple alignments. The first one is the number of highly conserved columns and the second one is the number of gaps in the alignment. There should be as many conserved columns and as less gaps as possible. These two measures should always be considered together and not isolated since the input parameters can be modified to maximize one of the two parameters. Therefore, a tradeoff between the two criteria has to be found in practice that always depends on the desired results. For example, if one searches for signal peptides, i.e. short and highly conserved sequence stretches, lower gap penalties should be chosen as when comparing the overall similarity of proteins.
Based on our experience, we would choose ClustalW for multiple alignments, because it yielded the best tradeoff between usability and result quality.
