CHIMeRA datasets

We have seen how we can extract useful information also from unstructured databases using a web-scraping pipeline. The on-line doctor web pages could be very useful for a toy model application like the SymptomsNet one but if we want produce scientific relevant results we have to take care about the validity of data. Since the English datasets availability is easier than the Italian one we moved to more “robust” databases.

As told in the previous sections, there are a lot of studies performed on the disease associations to other biological compounds and in many cases the resulting datasets are publicly available on Internet. This is the case of the DisGeNET or DrugBank datasets which contain the relationship between a large number of diseases with genes/variants (SNPs) and drugs (and other information), respectively. The DisGenet allows to download the datasets already stored into a well structured network format (sparse adjacency matrix, with 210498 associations between 117337 SNPs, 10358 diseases and 17549 genes) while the DrugBank poses more issues to the treatment of data: the DrugBank database was designed to provide a large set of information related to each drug using its own website and thus it needs a huge pre-processing of the JSON dataset structure to highlight all the possible network associations (14812 drugs, 649 metabolite pathways, 3256 gene targets, 40 SNPs targets, and 532 food interactions). Using the DisGenet we can connect the diseases to their related genes and variants. From the reviewed format of the DrugBank, instead, we can link each disease to the associated drugs. Associated to each drug we have also a list of gene and SNP targets which can be merged to the information provided by DisGenet. Moreover, we can insert also food interactions, metabolite pathways and drug interactions (synergy or not) extracted from DrugBank. We would stress that, despite the trivial overlaps between the same data sources (genes, diseases and SNPs up to now), just using the rearrangement of these pair of datasets into a network structure, we can already provide a possible extrapolation of the underlying information using the paths between nodes. Starting from a disease inside the DisGenet, using a single-database approach we can study the “causality” relationships with the connected genes and SNPs. Using a multiple-databases (or a network-of-networks structure) approach we can map that disease to other kinds of information like drugs, foods and metabolite pathways. The aim of such a network-of-networks structure is to unveil relationships hidden by the underwhelming overlap between single-type information across different databases. The set of different information merged can thus be exploited for applications such as wide-scale drug effect evaluation and design, addressing general diagnostic questions for systems medicine and diseases etiology expansion. In other words, a network-of-networks structure allows the inference of missing connections using node contraction. A full list of the information collected by our web-scraping and rearrangement pipelines is shown in the Table.

To enlarge our disease information we looked for other on-line data sources. A very interesting database is given by HMDB (Human Metabolite Data Bank)which comprises a vast amount of metabolites and metabolite-pathways with the associated drugs and diseases (114003 metabolite entries with chemical taxonomies and $\sim$25000 human metabolic and disease pathways ¹. The interconnections with the previous discussed datasets are straightforward but in this case the data are not publicly available and thus we had to apply a web-scraping algorithm to get its information. An analogous procedure was applied to extract the data included into the RXList database. RXList is an on-line website very similar to the previous discussed auto-diagnosis tools in which we can find associations between diseases and drugs and other several pathogenic associations. In this case we have a further distinction between diseases: we have diseases related to drugs and diseases connected to other caused-diseases. We can take care of this kind association using directional links ². We remark that each web-scraping pipeline is customized according to a precise website, so for each analyzed case a different code was developed to address the data extraction.

All these information can enrich our database and the description of a given disease but we have to face the problem of data merging. As previously discussed we do not have a unique nomenclature for diseases and thus we can find analogous names (periphrases or synonyms) which identify the same concept (disease). A useful tool to overcome these issues could be given by a synonym dictionary: a powerful example is given by the CTD (Comparative Toxicogenomics Database) (7212 diseases with mapped synonyms and 4340 diseases with related phenotypes). Using the CTD jointly with the SNAP (Stanford Large Network Dataset Collection, 8803 disease terms with related synonyms) database we could enlarge the number of synonyms associated to each disease name.

	disease	drug	food	gene	metabolite	phenotype	SNP	metabolic pathway
disease	CTD + RXList + SNAP	RXList	x	DisGeNET	HMDB	CTD	DisGeNET	HMDB
drug	RXList	DrugBank	DrugBank	x	x	x	x	DrugBank
food	x	DrugBank	x	x	x	x	x	x
gene	DisGeNET	x	x	x	x	x	x	x
metabolite	HMDB	x	x	x	x	x	x	HMDB
phenotype	CTD	x	x	x	x	x	x	x
SNP	DisGeNET	x	x	x	x	x	x	x
metabolic pathway	HMDB	DrugBank	x	x	HMDB	x	x	x
# nodes	63974	35161	532	18799	114100	13214	117337	1329

A full list of the information collected by our web-scraping and rearrangement pipelines is shown in Fig. 1.

We remember that the crucial point of our merging procedure is given by the disease nodes since they are the node type shared along (almost) all the databases. The help given by synonym dictionaries certainly increases the overlap between the mined datasets but we chose to maximize it using a pre-processing NLP pipeline. So, we started our pipeline using a word standardization, i.e converting all the words into their lower case formats and replacing all the punctuation characters with a unique one ³. Then, we noticed that a not negligible part of words involved into the disease names was useless for the description: words like “syndrome”, “disease”, “disorder”, “deficiency”, … are not descriptive and so we can filter them. Now, we could split the disease name into a series of token according to the list of words which compose it (tokenization) and sort them.

To further increase the overlap we cut the inflected words to their root form using a stemming algorithm: the stemmer strength has to be tuned according to the desired result. A first processing was performed using a Lancaster stemmer (more aggressive). If the resulting output was too short to be compared with other names the starting token was processed by a Porter Snowball stemmer (less aggressive). The stemmer algorithm choice is a very crucial task for NLP because using it we drastically loose information (it is not reversible). Other processing steps were performed for critical cases encountered during the analyses: these steps constrain our pipeline and they tuned it for the underlying application.

The work-flow outputs include multiple false-positive matches: the pipeline performs a brute force processing and some information lost along the steps could be significants. In these cases we have multiple processed names belonging to different kind of diseases: an example is shown in Fig. 2. Considering the original name and the processed one (pipeline output) we merged two names using a score match. This can be achieved introducing the standard word metrics: a common distance between words can be evaluated using the Levenshtein Distance which follows the equation

$$d_{a, b}(i, j) = \left\{ \begin{array}{rc} \max(i, j) & \mbox{if } \min(i, j) = 0 \\ \mbox{min} \left\{ \begin{array}{r} d_{a, b}(i - 1, j) + 1 \\ d_{a, b}(i, j - 1) + 1 \\ d_{a, b}(i - 1, j - 1) + 1_{(a \neq b)} \\ \end{array} \right. & \mbox{otherwise} \\ \end{array} \right.$$

where a and b are two strings of length |a| and |b| respectively. The indicator function $1_{(a \neq b)}$ is equal to 0 when $a_i = b_j$ and 1 otherwise. In this way the Levenshtein distance between a and b can be computed evaluating the distance between the first i characters of a and the first j characters of b. Despite the apparently complex mathematic formulation of this function, the Levenshtein Distance is a particular case of the more general Edit Distance, i.e a way to quantifying how dissimilar two strings are to one by counting the minimum number of operations required to transform one string into the other. Also in this case an example could be explainer: given the two string “kitten” and “sitting” their Levenshtein distance is equal to 3, in fact

1. kitten -> sitten (substitute “s” for “k”)
2. sitten -> sittin (substitute “i” for “e”)
3. sittin -> sitting (insert “g” at the end)

Using the Levenshtein equation we evaluated the distance between the two original names and we associate the disease to the higher scorer. A summary scheme of our pipeline is shown in Fig. 2.

The described NLP pipeline further increases the overlap between databases (e.g CTD-SNAP 24.17%; DisGenet-RXList 19.78%). We manually supervised and checked the merging procedure taking care to reduce the false positive percentage. In some cases the overlap percentage remained low also after the application of our pipeline (e.g. RXList-HMDB 8.03%; SNAP-HMDB 0.39%). Different data sources could be focused on different types of information and it is therefore reasonable to assume that in some cases the overlap is low. We supervised these critical cases with a manual check and we demonstrated our hypothesis. This behavior supports our databases choice: they include complementary information which could improve the informative efficiency of our structure. At the same time this result also proved the efficiency of our pipeline and it confirmed that the union of multiple data sources could effectively enlarge our knowledge about biomedical compounds.

The output of our merging procedure allowed the realization of the CHIMeRA network, i.e a network with more than $3.6\times10^5$ nodes and more than $3.8\times10^7$ links (ref. Fig. 3). In our resulting structure we have 7 node types: disease (63974), drug (35161), gene (18799), SNP (117337), metabolite (114100), phenotype (13214), metabolite-pathway (1329) and food (532). The full network adjacency matrix is still a block matrix, i.e we have not all the combinations of information in our databases. An emblematic case is given by the food nodes: we have food information only into the DrugBank dataset and thus they would be connected only with drug types. On the other hand our network architecture could be easily improved adding new data sources: the same food nodes are pendant nodes that could be easily connected to other kind of data introducing novel node types or just filling the available blocks. CHIMeRA is still a work in progress project so we are still looking for improvements and new databases to add.

next »

1. The human metabolite-pathways can be divided into different types according to the informations stored in the HMDB dataset. The interactions between HMDB and DrugBank was already established through a vast series of hyper-links which connect them using metabolites and metabolite-pathways information. In this way we mapped also the information related to the metabolite-pathways types to the DrugBank dataset, obtaining a finer grain nomenclature and classification of these data. These informations can be used to improve our disease information. In the previous Table is shown only the aggregated data. ↩

2. For sake of clarity, we encountered the same issue also into the DrugBank dataset in which we had intra-drug connections. ↩

3. An unexpected issue arise in this step: different databases use different enumeration system. In some entries we found disease names associated to numbers which identify their multiple types. An example could be “Polyendocrine Autoimmune Syndrome type 1” but at the same time in a second database the same disease could be represented by “polyendocrine autoimmune TYPE I”. Despite the global differences between the two names, given in this case by upper- and lower-cases of some letters and the deletion of some words, a very critical odds is the enumeration style. The performances of our pipeline dramatically increased using a roman_number_converter algorithm. ↩