s. Identifying the correct gene improved from 77% to 91% when considering the mutations.
To demonstrate practical relevance, we set up a mutation screening for five membrane proteins from the family of G proteincoupled receptors to evaluate a solvation energy based model for the prediction of stabilising regions in membrane proteins. We identified 35 mutations in text. 25 out of 35 mutation phenotypes reported in literature were in compliance with the prediction of the energy model, which supports a relation between mutations and stability issues in membrane proteins.
Proteins carry out most cellular functions as they are
acting as building blocks for structures, enzymes,
gene regulators, and are involved in cell mobility
and communication
It is envisaged to provide an automated system for the interpretation of structure-function relations in the context of genetic variability data. Despite the availability of numerous biomedical data collections, valuable information about mutationphenotype associations is still hidden in nonstructured text in the biomedical literature. Thus text mining methods are implemented to automatically retrieve these data from the 18 millions of literature references in PubMed. The extracted knowledge will be stored in one homogeneous data store and integrated with already available data from suitable databases. On the basis of all these combined data, new hypotheses can be formulated, like the prediction of phenotypic effects induced by mutations. At the moment, we are populating a database with organism specific protein-mutation associations which we envisage to apply on diverse biological problems, such as the detection of mutation centred genedisease associations in human. 2
Genomic variation data has already been collected for many years. Single nucleotide polymorphisms (SNPs), which make up about 90% of all human genetic variation and occur every 100 to 300 bases along the 3-billion-base human genome, are available as large collections. Single amino acid polymorphisms (SAPs) are often manually extracted from literature and curated into databases, originating from wet lab experiments. Additionally, some structures of such mutations may be revealed in crystallography experiments and might eventually end up as distinct structures in the Protein Database PDB. Of particular interest is the identification of mutations which have a strong influence on the stability of proteins. Therefore, the biomedical literature can be systematically searched for information about mutation-phenotype associations by text mining, which may lead to new insights beyond information in existing databases. For the text mined data it is additionally possible to weight or prioritise information according to their publication date, the involved authors and the journal. Considering these meta data can be relevant if for instance an already published assumption has been proven wrong in a more recent publication, or for determining whether a protein is a hot topic or if the information is already available for years. Furthermore, it is possible to receive a more detailed view on a protein’s characteristics, e.g. if a certain interaction only takes place under specific conditions, or if an interaction is prevented by the conformational change of a protein domain triggered by a point mutation. 2.1
Data on mutations have been collected for years, for
numerous species and by different organisations for
diverse purposes. There are many efforts to cope
with the data, which is being made available in a
growing number of databases. The Human Genome
Variation society
Besides these central repositories, there are small
specialised databases, such as the infevers
autoinflammatory mutation online registry
In contrast, unpublished SNPs normally make their way into large locus specific data repositories. Since August 2006, there is a wiki based approach SNPedia in contrast to classical databases collecting information on variations in human DNA. 2.2
Despite the availability of numerous biomedical data
collections, valuable information about
mutationphenotype associations is still hidden in
nonstructured text in the biomedical literature. Thus
text mining methods are implemented to
automatically retrieve these data from the 18 millions of
referenced articles in PubMed. Text mining aims to
automatically extract and combine information spread
in several natural language texts and by this
generating new hypotheses. One of the key prerequisites for
finding new facts (e.g. interactions or mutations) is
the named entity recognition (NER) in text, the
assignment of a class to an entity (e.g. protein), as well
as a preferred term or identifier, in case an entry in
a database, such as UniProt, or a controlled
vocabulary like the Gene Ontology (GO)
Up to now, there have already been diverse
examples for the successful application of text mining to
the mutation retrieval task. Early examples are the
automatic extraction of mutations from Medline and
cross-validation with OMIM
More and more groups are working on
mutations in proteins and their involvement in
diseases.
Through the combination of different data from literature and databases it is possible to derive new facts, e.g. novel gene-disease associations or the influence of mutations on protein-protein interactions. The approach is designed in such a way, that it can in principle be applied to any kind of genetic data for answering disease centred questions. For the moment, we concentrate on collecting available high quality data on protein point mutations from curated databases and from peer-reviewed literature. For the latter we will present a flexible approach for both the specific and high-throughput retrieval of mutations. In detail, the following tasks have to be performed: (1) Identify genes/ proteins in abstracts. (2) From this subset consider only these which additionally contain information about mutations. (3) Propose potential protein - mutation pairs. (4) Filter proposed pairs by sequence compliance. (5) Utilise this information for the refinement of the original gene/protein identifier. 3.1
Gene normalisation This module allows for the
automated named entity recognition of genes and
proteins. Our approach performs gene name
disambiguation by using background knowledge to
match a gene with its context against the text as a
whole
In the process of recognising mutations in text, the normalisation, i.e. the direct association to specific proteins, remains a challenge. This is due to the fact that the abstracts of relevant publications typically mention more than only one single mutation and protein. Thus, a mutation-protein association purely based on their co-occurrence in one abstract is not sufficient, as it would result in a permutation with a huge number of false positive predictions. The problem becomes even more evident, when considering that both gene and mutation tagging are imperfect, achieving a precision of 80 to 90% each.
A method is desired, that both disambiguates the
relations of candidate mutations and proteins, and
filters out false positives from the underlying
individual mutation and protein recognition tasks. There
are approaches which apply a word distance
metric for assigning a mutation to its nearest occurring
protein term, which is error prone, as matching
mutation and protein do not necessarily have to occur
close to each other in the abstract or even in the
same sentence. The statistical approach GraB is an
excellent tool for the automatic extraction of
Protein Point Mutations using a Graph Bigram
association
Sequence Checks Mutations are commonly
described as the substitution of a wild-type by a
mutant amino acid at a given position. Our method
compares the wild-type residue as described in a
mutation mention with the UniProt/Swiss-Prot and
PDB protein sequences for all candidate proteins.
It is important to incorporate sequences from both
repositories, as the sequence numbering can differ
and it is not always evident from a publication’s
abstract, which numbering the mutation notation refers
to. To map UniProt IDs to PDB and vice versa, we
used PDB cross-references in
UniProtKB/SwissProt from http://beta.uniprot.org/docs/pdbtosp
and the residue specific comparison between
PDB and SwissProt sequences as provided by
http://www.bioinf.org.uk/pdbsws/
The developed mutation retrieval pipeline can be accessed through two different interfaces (see Figure 1), which offer dependent on the annotation task, either a systematic or quick and flexible solution. The following approaches have been implemented: • Organism-centred approach (database) All available mutations for a given organism will be retrieved in one single literature screening and stored in the Mutation database. This approach relies on the large-scale identification of gene mentions in PubMed abstracts, which have to be compiled for organisms of interest prior to a mutation screening. As of now, gene mention data is available for human, mouse, and yeast. However, data for additional relevant organisms will be added on a regular basis in the near future. • Protein-centred approach (on-the-fly)
It is possible to retrieve relevant data for a single gene or a list of genes/ proteins for any organism. For this purpose, the gene identification part performed by the gene normaliser is replaced by a direct full text search in the PubMed library using the Entrez Programming Utilities. Again, the result is a set of abstracts, which is subsequently processed by the MutationTagger.
As described above, we defined the input set of documents for the organism-centred mutation mining approach by scanning the whole PubMed database for abstracts mentioning at least one gene or protein of a pre-defined species. For this filtering step, we relied on the gene normalisation techniques of our gene normaliser, which was applied to all PubMed abstracts in advance and has shown 85% F-measure for human genes and slightly lower for other species. However, the gene normalisation proposes by default only one single identifier per gene mention, even if a set of different candidate identifiers was computed. According to internal ranking mechanisms, only the top scoring candidate is considered. This leads to a possible scenario, where in some cases the correct identifier is ranked lower and would be neglected for any subsequent data procession. In case of our mutation mining algorithm, we assume that some mutations cannot be associated to the correct protein, because the gene tagging task already failed.
On the other hand, it should be possible to improve the performance of both entity recognition techniques for genes and mutations by combining the results. The idea is to run both approaches with low precision thus receiving a high recall, permutate all elements of both sets, and then consider the intersection of all combinations that fit. Mutation and gene product are considered to be a valid pair, if the wild-type residues at the mutated position in the protein sequence and in the reported mutation match (as described in section 3.1). For all proposed gene identifiers, protein sequences are obtained and checked for compliance with the reported wild type amino acid. The score of identifiers that show a match are increased, which might lead to a re-ranking of the identifiers for one gene entity. This could further improve the original gene normalisation approach for candidate entities which are reported to show a mutation.
Example As shown in Figure 2 our gene normaliser identified CCP (human crystallin, gamma D; EntrezGene ID 1421) as the top candidate gene name for abstract PMID 8142383. The mutation tagger identified a replacement of tryptophan with glycine at position 191 as the only mutation mentioned in the paper. None of the protein sequences retrieved for human CCP showed a tryptophan residue at position 191, which means that this gene identifier was not supported by mutation information. However, besides human crystallin, there was also cytochromec peroxidase in yeast (EntrezGene ID 853940) proposed as an alternative identifier, which received a lower score. As the product of this gene showed a tryptophan residue at postion 191 (according to PDB sequencing) the score was increased making it the new top candidate. Indeed, manual curation of the corresponding literature confirmed, that the only gene mentioned in the abstract is cytochrome-c peroxidase in yeast. The same positive re-ranking finding the correct gene identifier through mutation information was shown for human TP53 in paper 11254385, and human amylase alpha in paper 15182367. 4
Mutation database In order to establish a mutation database, which will eventually store all protein point mutations mentioned in PubMed abstracts for all organisms of interest, a first platform has been realised, comprising a MySQL database, which can be accessed by a web-interface.
To populate the database, in a first step the PubMed corpus is filtered for abstracts mentioning at least one gene or protein using the named entity recognition algorithm as described in Section 3.1, which is currently working for the three organisms human, mouse, and yeast. This led to a set of set of 3,443,566 abstracts proposing more than 10 millions of potential protein candidates. In a second step, the mutation extraction algorithm is applied on this corpus and the retrieved information is transferred into the database. In total, 258,511 mutations were found in 78,968 abstracts. Subsequently, for all candidate genes found in these abstracts, the corresponding sequences are obtained and checked for compliance with the wild type amino acid at the position of the mentioned mutation, which led to a number of 877,183 potential protein - mutation pairs. Out of these, 127,384 are supported by sequence (74,722 if multiple mentions of the same mutation in one abstract are counted as one) in contrast to 131,127 (77,643) mutations which have not passed the sequence filter. In summary, from all mutations identified by the plain algorithm, about 49% could be supported by gene associations based on sequence check. These data were retrieved from 41,384 (52%) abstracts in total.
Evaluation We evaluated our approach on two
different tasks: pure identification of a mutation in
a text, and the identification of correct
mutationprotein pairs. An evaluation of our method on
the test data from MutationFinder
To assess the refinement possibilities for falsely top ranked gene names, from the 182 abstracts we took the subset of those, the gene normaliser identified genes from one of the 10 supported species: human, mouse, yeast, rat, fruit fly, H. pylori, S. Pombe, C. Elegans, A. Thaliana, and D. Rerio. This led to a subset of 22 abstracts. In the initial run, the gene name identifier identified in 17 of 22 abstracts (77%) the correct gene as the top ranked candidate. However, after the gene tagging refinement by applying the sequence filter to all candidate genes, the genes of 3 more papers were identified correctly replacing the original and false top candidate. This led to the correct protein normalisation for 20 out of 22 (91%) publications. For the remaining 2 publication, the correct genes could not be identified, as they were from species, the gene identifier does not yet support. The suggested genes from mouse were first falsely predicted, which were then not supported by the sequence checks. By this the proposed identifiers were brought below the threshold, resulting in no gene identification at all for these 2 abstracts and turning the 2 “false positives” to “false negatives”. On-the-fly vs. database approach We evaluated the results of the two access approaches (database and on-the-fly) for human Aquaporin-1, as part of the stability analysis of protein membranes (see Section 5). The precision of the on-the-fly approach is expected to be lower, as the first step is more general due to relying on full text searches instead of entity recognition. Indeed, in comparison to the unique 20 mutations found by the organism-centred approach, 9 additional mutations were found, of which all were false positives, actually appearing in Aquaporin-2 or 4. This supports the good precision of the named entity approach for the gene normalisation. 5
Integral membrane proteins play an important role
in all organisms, especially as transporters. Due to
their striking importance, mutations in membrane
proteins are known to be the cause of many
hereditary diseases, such as cystic fibrosis, or retinitis
pigmentosa. The reason are often conformational
changes in proteins, which may lead to malfunction
of a whole protein complex. Unfortunately,
identified structures for membrane proteins are still rare.
For this reason, we used a coarse grained model
presented by
The approach considers the solvation energy, which is based on the probability distribution for each amino acid within the integral part of a membrane protein to be facing the membrane or other proteins. The amino acid specific property inside or outside reflects the orientation of the amino acid side chains with respect to the centre of mass of the neighbouring residues. For a given mutation, the approach compares the solvation energies for wildtype and mutant residues. If the energies differ significantly, a destabilising effect is predicted, especially if the energies are changing from negative to positive or vice versa.
To quantify the ability of this model to predict the influence of mutations on the stability of membrane proteins, we compared already examined and published effects of mutations with the predictions of the sequence based model. For this purpose, we screened the literature for single point mutations reported for five membrane proteins from the family of G protein-coupled receptors (bacteriorhodopsin and halorhodopsin from Halobacterium salinarum, bovine rhodopsin, Na+/H+ antiporter from Escherichia coli, and human aquaporin-1). As described in Section 4, Protein-centred approach and Figure 1B, articles relevant for these proteins were identified by searching PubMed via the NCBI Entrez Programming Utilities. Abstracts for each protein were queried by the protein and gene name including the synonyms as derived from the corresponding PDB/UniProt entry.
The MutationTagger was applied on these five sets of abstracts for the extraction of mutation information. The application of sequence checks brought the results down to a reasonable number of proposed mutations, which were presented as HTML documents and subsequently manually curated. We only used the publications where a single point mutation was discussed in the context of stability or stability related function. Double or multiple mutations were not considered, as the determination of a direct relation between the reported effect and one of the mutations is not possible. If an appropriate mutation was found in the literature, we compared the solvation energies of both wild-type and mutant residues to decide, if the mutation was stabilising, slightly stabilising, slightly destabilising, or destabilising. Example Mutation T93P for bovine rhodopsin was reported to lead to a conformational change of the protein. Considering the two solvation energies of wild type Threonine (-0.66 a.u.) and mutant Proline (0.08 a.u.) a destabilising effect can be predicted, although both amino acids are actually classified as neutral. Without the change of sign from - to +, an only slightly destabilising effect would have been hypothesised.
Relevance We were able to show the ability of our mutation mining approach to retrieve publications containing mutation information for given proteins at a good precision. Due to the quick and precise retrieval of mutation data we were able to assess the soundness of the coarse grained model for the prediction of stabilising regions in membrane proteins. 25 out of 35 mutational effects reported in the literature for any of these five membrane proteins correlate with the predictions based on the solvation energy. These cases suggest a relation between mutations and stability issues in membrane proteins. Acknowledgement: We are grateful for financial support by the EU project Sealife and the BMBF Format Project CLSD and to Frank Dressel and Dirk Labudde for discussions on the application. 6
We developed a rule- and regular expression-based approach that allows for the retrieval of protein point mutations from the whole PubMed database specifically for any given protein. This flexibility makes it a powerful tool for immediately finding relevant data for follow-up studies, as we showed in the application on five membrane proteins. In addition, MutationTagger can be utilised for the species-wide identification of mutations in proteins mentioned in PubMed. We started to set up a mutation database which allows for systematically querying mutation related information, and finding relevant literature for subsequent studies. The sequence checks applied on identified mutations and candidate proteins have been proven to be an efficient, yet not sufficient filter for determing mutation-protein associations. The filter shows good sensitivity but improvable specificity, especially regarding the species level. Furthermore, we were able to show, that the mutation information from literature can even further improve the quality of the gene tagging algorithm we used, which already showed very good results.