=Paper=
{{Paper
|id=Vol-1795/paper2
|storemode=property
|title=Studying the Cohesion Evolution of Genes Related to Chronic Lymphocytic Leukemia Using Semantic Similarity in Gene Ontology and Self-Organizing Maps
|pdfUrl=https://ceur-ws.org/Vol-1795/paper2.pdf
|volume=Vol-1795
|authors=Efstratios Kontopoulos,Theodoros Moysiadis,Maria Tsagiopoulou,Sándor Darányi,Peter Wittek,Nikos Papakonstantinou,Stavroula Ntoufa,Georgios Meditskos,Yiannis Kompatsiaris,Kostas Stamatopoulos
|dblpUrl=https://dblp.org/rec/conf/swat4ls/KontopoulosMTDW16
}}
==Studying the Cohesion Evolution of Genes Related to Chronic Lymphocytic Leukemia Using Semantic Similarity in Gene Ontology and Self-Organizing Maps==
https://ceur-ws.org/Vol-1795/paper2.pdf
Studying the Cohesion Evolution of Genes Related to
Chronic Lymphocytic Leukemia Using Semantic
Similarity in Gene Ontology and Self-Organizing Maps
Efstratios Kontopoulos1, Theodoros Moysiadis2, Maria Tsagiopoulou2, Sándor Da-
rányi3, Peter Wittek3,4, Nikos Papakonstantinou2, Stavroula Ntoufa2, Georgios Med-
itskos1, Kostas Stamatopoulos2, Ioannis Kompatsiaris1
1 Information Technologies Institute, Thessaloniki, Greece
{skontopo, gmeditsk, ikom}@iti.gr
2 Institute of Applied Biosciences, Thessaloniki, Greece
{moysiadis.theodoros, m.tsagiopoulou, npapakonstantinou,
sntoufa, kostas.stamatopoulos}@certh.gr
3 University of Borås, Sweden
Sandor.Daranyi@hb.se
4 ICFO – The Institute of Photonic Sciences, Spain
Abstract. A significant body of work on biomedical text mining is aimed at un-
covering meaningful associations between biological entities, including genes.
This has the potential to offer new insights for research, uncovering hidden
links between genes involved in critical pathways and processes. Recently,
high-throughput studies have started to unravel the genetic landscape of chronic
lymphocytic leukemia (CLL), the most common adult leukemia. CLL displays
remarkable clinical heterogeneity, likely reflecting its underlying biological
heterogeneity which, despite all progress, still remains insufficiently character-
ized and understood. This paper deploys an ontology-based semantic similarity
combined with self-organizing maps for studying the temporal evolution of co-
hesion among CLL-related genes and the extracted information. Three consecu-
tive time periods are considered and groups of genes are derived therein. Our
preliminary results indicated that our proposed gene groupings are meaningful
and that the temporal dimension indeed impacted the gene cohesion, leaving a
lot of room for further promising investigations.
Keywords: Chronic Lymphocytic Leukemia, Gene Ontology, Semantic Simi-
larity, Semantic Drift, Self-Organizing Maps.
1 Introduction
A significant and continuously growing body of work on biomedical text mining is
aimed at uncovering meaningful associations between biological entities, including
genes (e.g. [1]-[3]). This line of research promises to deliver new scientific insights,
uncovering hidden links between genes involved in critical pathways and processes.
Chronic lymphocytic leukemia (CLL), the most common adult leukemia in the
West, is a chronic, incurable malignancy of mature B lymphocytes. CLL displays
remarkable clinical heterogeneity linked to the underlying biological complexity that,
despite recent progress, remains insufficiently characterized and understood [4].
Moreover, CLL has not been subjected to analysis with text mining tools, though such
approach holds great promise for uncovering hidden knowledge within the wealth of
published biomolecular information from CLL patient profiling studies.
Towards addressing this vital challenge, in this paper we propose a novel method-
ology for studying genes and their potential relationship to CLL. Based on a set of
CLL-related PubMed abstracts, we are deploying a state-of-the-art software tool for
generating groups of genes belonging to three consecutive time periods. After meas-
uring the semantic similarity of genes within the groups, we investigate the types of
explicit and implicit information that can be derived from each group. This infor-
mation could be valuable in characterising groups of interest. Furthermore, and since
the groups are expected to evolve over time, we are also investigating the impact in-
flicted on the groupings by the temporal dimension, mainly focusing on the temporal
evolution of the cohesion among CLL-related genes and the relevant extracted infor-
mation. Temporal refers to the comparison of the results among the three periods. All
our implementations are based on established open-source software tools.
Our preliminary results indicate that, compared to random group generation, the
groupings we considered within each time period are indeed meaningful. Naturally, in
some cases this difference is more emphatically reflected and statistically validated.
Furthermore, the temporal dimension indeed impacted the gene groupings. In particu-
lar, derived groups of genes with biological importance were not consistent through-
out the three time periods considered, and in some cases the changes are due to ex-
plicit biological explanation.
2 Related Work
A significant body of work exists on the area of biomedical text mining, namely the
application of information extraction and text mining approaches on large amounts of
biomedical literature with the goal of uncovering novel associations between biologi-
cal entities of interest such as genes. For instance, in [1] the authors carried out auto-
mated extraction of explicit and implicit biomedical knowledge from publicly availa-
ble gene and text databases, in order to create a gene-to-gene co-citation network for
13,712 named human genes by automated analysis of titles and abstracts in over 10
million MEDLINE1 records.
In [2] the authors presented a method for creating a network of gene co-
occurrences from the literature (article abstracts) and for partitioning it into communi-
ties of related genes, expected to be related by function. They noted that the derived
gene communities are not meant to perfectly reproducing biological reality.
A couple of other related approaches are [3] and [5]: The former explored the utili-
ty of Latent Semantic Indexing (LSI) to automatically identify conceptual gene rela-
1 https://www.nlm.nih.gov/pubs/factsheets/medline.html
tionships from titles and abstracts in MEDLINE citations, while the latter presented
GenCLiP, a tool that searches gene lists to identify functional clusters of genes based
on up-to-date literature profiling.
LSI is based on globally optimal dimensionality reduction, namely, singular value
decomposition. If we reduce the dimensions to two, we gain visual insight on the
layout of the data, but a globally optimal embedding disregards the local topologies of
the data points. Complete isometric feature mapping (Isomap [6]), t-Distributed Sto-
chastic Neighbor Embedding (t-SNE [7]), and self-organizing maps (SOM [8]),
among many other techniques, provide a two-dimensional embedding that preserves
the local topology, at the expense of distorting the global alignment of data points.
SOM is particularly useful in online updates.
The above approaches aim to facilitate researchers to swiftly search for known in-
teractions and to provide insight into unexplored connections. This is also one of our
aims in this paper. However, we strongly believe that it is even more important to
detect potential semantic drifts within the groups of genes, and, if such drifts occur, to
evaluate them. Thus, we are not solely interested in gene groupings, but rather in re-
vealing implicit information and assessing occurring drifts. To the best of our
knowledge, this novel problem has not been tackled yet in biomedical sciences.
3 Semantic Similarity and Applications in Biomedicine
Semantic similarity is a measure for calculating the likeness of the meaning or seman-
tic content of a set of terms or documents. One can use dictionaries, semantic net-
works, thesauri and, more recently, ontologies as the underlying resource for measur-
ing semantic similarity between terms.
The various metrics for calculating semantic similarity are grouped in four catego-
ries: (a) Path-based metrics, where the similarity between two terms depends on their
relative position in the underlying taxonomy, as well as on the length of the path link-
ing the concepts (e.g. [9, 10]); (b) Content-based metrics are based on the information
content available for each concept; the more common information two concepts share,
the more similar they are (e.g. [11, 12]); (c) Feature-based metrics are based on the
properties of the underlying ontology for obtaining a similarity value. The more
common characteristics two concepts have, the more similar they are (e.g. [13]); (d)
Hybrid metrics combine the ideas presented above (e.g. [14]).
Especially in the biomedical domain, measuring semantic similarity constitutes a
core process towards information extraction and information retrieval. Due to its vast
popularity, the Gene Ontology (GO) [15] typically plays the role of the underlying
ontology model [16, 17]. Some of the most popular tools for calculating the semantic
similarity in GO terms include GOssTo [18], GOSemSim [19] and G-SESAME [20].
The latter is a very popular set of online tools and was deployed in this work as the
tool for measuring semantic similarities of genes in GO.
4 Methodology
Based on previous preliminary work of ours in a different domain [21], we are pro-
posing a novel methodology for studying the temporal evolution of cohesion among
CLL-related genes, which is composed of the following steps (see Fig. 1):
1. A set of input documents relevant to CLL are fed to Somoclu [22], a massively
parallel open-source implementation of SOMs.
2. With the help of an accompanying input set of gene terms, Somoclu generates a
two-dimensional (2D) map of gene clusters, based on the terms’ co-occurrence in
the set of input documents.
3. The semantic similarities of the gene terms residing in the groups are measured us-
ing the G-SESAME set of online tools (see previous subsection). G-SESAME was
chosen due to its novel Aggregate Information Content (AIC) semantic similarity
metric that relies on the information content of all ancestor terms of a GO term
[23]. AIC implicitly reflects the GO term’s location in the GO graph and effective-
ly represents how humans use this term and its ancestor terms to annotate genes.
4. A set of semantically meaningful gene groupings is derived, along with potentially
useful implicit information.
The above series of steps is clearly not restricted to CLL, and can be easily applied for
input documents relevant to any malignancy of interest.
Fig. 1. Proposed methodology steps.
We build a vector space model by random indexing of the evolving text collection
and use Somoclu to train a self-organizing map (step 2) that is able to closely track
the changes. Random indexing captures term co-occurrence patterns similarly to an
ordinary term-document matrix, but the resulting vector space is inherently lower
dimensional – we used two hundred dimensions. The SOM training procedure used
Euclidean distances to map points to the 2D grid. The input corpus is split into time
periods (see next subsection) and a vector space is built for each period. We use
Emergent Self Organizing Maps (ESOMs) [24] to project the vector space to a 2D
surface where clusters and shifts are more apparent. Because ESOMs reproduce the
local but not the global topology of the high-dimensional space [8], the derived clus-
ters are locally meaningful and consistent on a neighbourhood level only.
5 Experiment Design
In order to study the temporal evolution of the semantic cohesion among CLL-related
genes, we used a dataset of PubMed abstracts on CLL (downloaded in April’16),
which we divided in three periods, after selecting two milestones with significant
impact to the bibliography that followed them. The milestones are: (a) The end of the
Human Genome Project (HGP) [25], which signalled the complete sequencing of the
human DNA for the first time and revealed a big part of today’s known genes; (b) The
first appearance of Next-Generation Sequencing (NGS) [26] technologies in CLL,
which allowed an entire human genome to be sequenced within a single day. Thus,
the resulting three periods are: Period 1 – From 3/1949 (first CLL publication) to
5/2004 (the end of HGP): 4390 abstracts; Period 2 – From 6/2004 (first publication
after HGP) to 7/2010 (appearance of NGS methods): 2567 abstracts; Period 3 – From
8/2010 (first NGS publication for CLL) to 4/2016: 3456 abstracts.
The list of all known human genes, which served as the underlying vocabulary for
Somoclu, was downloaded from the HUGO Gene Nomenclature Committee
(HGNC)2; besides the “official” gene symbols, the latter also includes a list of synon-
ymous symbols referring to each gene, which were also considered in our analysis.
5.1 CLL Gene Groupings
Based on the input corpus and list of gene terms, Somoclu generates 2D maps of the
derived clustering of genes for the three periods (see Fig. 2 – term locations are indi-
cated with white dots – term labels are omitted here). As already mentioned, clusters
are locally meaningful and consistent on a neighbourhood level only.
Fig. 2. 2D maps of gene groupings per period, visualized by Somoclu. First period is on the
left, second period in the middle and third period is on the right.
Drift indications are also highlighted in the visualization: content splitting tenden-
cies are indicated by the ridge wall width and height around the clusters, so that the
method yields an overlay of two aligned contour maps in change, i.e. content structure
vs. tension structure. Thus, in the figure, the lighter coloured areas (blue to green)
stand for mostly inactive content: points assigned to these areas cluster nicely, their
distance in the original high-dimensional space was low. The darker areas (yellow to
red) indicate tensions: the two-dimensional embedding is distorted, adjacent points
refer to distant locations in the original high-dimensional space. Content can still be
2 http://www.genenames.org/
mapped in the latter, but it’s more likely to drift than not. Overall, Fig. 2 indicates a
lot of turbulence in all periods, in the sense that the embedding undergoes rapid and
abrupt changes, reflecting similar tendencies in the distances in the original space.
Genes grouped within the same group (or mapped onto the same neuron in the grid
used by the algorithm) indicate an inherent interrelatedness, e.g. they are related to the
same functionality, but also potentially hold additional important information regard-
ing unexpected relationships between genes, which was not previously determined.
For instance, two characteristic neurons in our experiment were: (1) Neuron [ccr7,
ccr2, cxcr2, cxcr1] in the first period, referring to two families (CCR and CXCR) of
chemokine receptors; (2) Neuron [tlr10, tlr6, tlr2, tlr1] in the third period that in-
cludes genes belonging to the same family (TLR), which showed significant associa-
tion with B cell physiology and CLL pathogenesis.
Furthermore, and even more importantly, the spatial behaviour of genes from peri-
od to period is of great significance. For instance, imagine two genes that are close to
each other in the first two periods but drift apart in period 3. This could possibly im-
ply that some underlying event taking place between periods 2 and 3 resulted to the
genes moving away from each other and that further investigation would be needed.
5.2 Statistical Evaluation
In order to assess the groupings derived from Somoclu, we conducted a statistical
evaluation, based on the groupings derived in each period. The aim was to assess
whether proximity of genes in the toroid plane indicates a strong underlying semantic
similarity. We focused on neurons containing more than one genes. To evaluate the
consistency of our approach in each period, we have checked whether the assignment
of genes within a randomly selected neuron was better than random assignment.
The key idea was to simulate random assignment by the empirical probability dis-
tribution of the average semantic similarities corresponding to randomly generated
neurons based on our data. Thus, for each neuron, we computed the average semantic
similarity with G-SESAME, and determined their empirical probability distribution.
For the actual groups that emerged from Somoclu’s analysis, we calculated the
AIC semantic similarities among all terms within each neuron and computed their
average. Then, we compared the empirical probability distributions of the actual and
random groups both graphically and based on statistical tests. In particular, the non-
parametric Mann-Whitney test was selected, since the sample sizes where not large
enough in every case to consider a parametric t-test. In addition, we performed a 1-
sided test (since the actual mean was higher than the random in all 6 cases) to evalu-
ate the alternative hypothesis that actual groups will exhibit higher values of average
similarities compared to the random groups. In case the null hypothesis of equality
was rejected in most cases, our expectation that the actual assignment is better than a
random assignment would have been validated.
Towards the same direction, we computed the 90%, 95% and 99% percentiles,
based on the probability distribution of the random groups, and computed the respec-
tive percentage of actual group averages that took value beyond these thresholds. Our
expectation was validated in each case where the percentage was higher than the
complementary probability (10%, 5% and 1%, respectively). To further consolidate
the results, these comparisons were performed, based on the exact binomial test. The
confidence level in all tests was set to 5%.
5.3 Discussion and Assessment of Results
Regarding the grouping of genes, we observe in Table 1 that in all cases the mean of
averages is higher in actual groups compared to random groups. In half the cases this
observation is emphatically validated through the Mann-Whitney test (p-value<0.05).
In two other cases the non-rejection of the null hypothesis was rather marginal and in
only one case (F3) the p-value was large (0.526). This validates our expectations re-
garding higher values of average AIC semantic similarity within the actual groups and
shows that the groupings considered are meaningful.
Table 1. The number of the groups and the corresponding mean of the average semantic simi-
larities are displayed per state, as well as the derived p-value of the distribution comparison
according to the Mann-Whitney test. The upper panel refers to C and the bottom panel to F.
GO Size Mean P-value
Period
category Random Actual Random Actual (1-sided)
1 93 23 0.380 0.395 0.174
C 2 151 45 0.401 0.424 0.011
3 220 93 0.399 0.421 0.023
1 93 23 0.402 0.437 0.202
F 2 151 45 0.382 0.423 0.018
3 220 99 0.389 0.392 0.526
Table 2. The 90%, 95% and 99% percentiles, based on the probability distribution of the ran-
dom groups, together with the corresponding p-values of the exact binomial test that compares
the percentage of actual group averages to the complementary probability 10%, 5% and 1%,
respectively. The upper panel refers to C and the bottom panel to F.
GO category Period 90% P-value 95% P-value 99% P-value
1 0.477 0.408 0.512 0.105 0.536 1.000
C 2 0.527 0.947 0.551 0.665 0.611 0.363
3 0.506 0.021 0.537 0.002 0.627 0.014
1 0.542 0.684 0.597 0.320 0.716 0.206
F 2 0.509 0.012 0.541 0.006 0.667 1.000
3 0.526 0.667 0.576 0.555 0.650 0.630
Table 2 further validates the results of Table 1, since statistically significant differ-
ences in the distributions are mostly reflected in significant results in Table 2, howev-
er, this is not always the case (C2). This could be due to many reasons, e.g. note that
the number of actual groups in C2 is small in comparison to the respective number of
random groups. All the above are also reflected in a smooth representation of the
empirical probability distributions for the actual and random case in Table 1.
In order to biologically evaluate the drift between the three periods, the set of
genes in the map was analysed per period, based on KEGG pathway analysis, a com-
mon methodology for discovering the enriched molecular pathways that correspond to
a set of genes, the latter given as an input. The results revealed that 53 out of 80
pathways were common in all three periods, most of them with a critical role in CLL,
such as the B cell receptor, NF-kappaB, p53, PI3K-Akt and signalling pathways [27].
This indicates the similarities between the three periods, but also highlights their dif-
ferences, since a number of 27 pathways emerged in only one or two of the periods.
For instance, MAPK and Fc epsilon RI signalling pathway are key pathways in CLL
but appeared in only two of the three periods
Fig. 3. Smoothed empirical probability distributions for the actual and random groups for C, F
for all three periods. Pink represents actual and light blue represents random distribution.
The next step was to focus on a neuron level and explore the differences drifting
between periods. Neurons with high values of average semantic similarity were of
particular interest. One characteristic example is neuron [trim8, zbtb4, tp53inp1, e2f5]
that exhibited the highest average semantic similarity in the third period, when the
cellular component category was considered. None of these genes appeared in the first
two periods, but all of them are transcription factors and show common regulation.
Another example is neuron [wee1, cllu1, per2, per1, clock], which also exhibited
high similarity in the third period under the cellular component category as well. The
wee1, per2, per1 and clock genes are regulatory elements of the circadian clock, while
the cllu1 gene is overexpressed in CLL and serves as a prognostic marker. In the first
period only wee1 appeared, while in the second period cllu1, per2 and per1 were
found in the same neuron. This observation could be possibly explained by the fact
that the circadian clock in cancer was well documented from 2009 and onwards,
which almost coincides with the end of the second period. This coexistence could
suggest an association of circadian clock and the overexpression of cllu1 in CLL.
Finally, we note that no neurons were unchangeable (same set of genes) in all three
periods. Thus, it would be of interest to expand the groups considered, in order to
include the neighbourhoods of neurons (e.g. at a radius of 2 or 3 cells). Another justi-
fication is the fact that the three periods did not overlap. Thus, well studied genes in
one period may be missing from the literature in the next period, and even more spe-
cifically the interest for relation among genes in a specific period could have faded
out in the next one, or could have been inexistent in the previous.
6 Conclusions and Future Work
The paper proposed a novel methodology for studying genes and their potential rela-
tionship within the field of CLL, which can be easily transferred to other disease set-
tings as well. The methodology capitalizes on GO-based AIC semantic similarity and
a state-of-the-art ESOM implementation for generating groups of genes belonging to
consecutive time periods. Our experiments 3 verified the validity of the methodology
and demonstrated that the temporal dimension indeed impacted the gene groupings.
This is largely a preliminary investigation with room for further explorations. In
particular, instead of three consecutive periods (1, 2 and 3), it would be interesting to
consider three overlapping periods (1, 1+2, 1+2+3). Such a consideration would
maintain all the information derived from past periods, although obsolete information
would also be maintained. Another interesting parameter is the number of periods,
which could be increased, under the constraint that the number of abstracts per period
is not greatly decreased. This would enable to collect more information in terms of
drifting and to statistically evaluate it. Of a more immediate interest would be to study
characteristic grouping examples already derived, such as neuron [wee1, cllu1, per2,
per1, clock] (see Section 5.3), in terms of implied information, based on GO. A sub-
sequent expert biological assessment could reveal valuable novel knowledge.
Acknowledgments. This project has received funding from the European Union’s
Seventh Framework Programme for research, technological development and demon-
stration under Grant Agreement Number FP7-601138 PERICLES.
References
1. Jenssen, T. K. et al. (2001). A literature network of human genes for high-throughput anal-
ysis of gene expression. Nature Genetics, 28(1), pp. 21-28.
2. Wilkinson, D. M., Huberman, B. A. (2004). A method for finding communities of related
genes. National Academy of Sciences, 101, pp. 5241-5248.
3. Homayouni, R., Heinrich, K., Wei, L., Berry, M. W. (2005). Gene clustering by latent se-
mantic indexing of MEDLINE abstracts. Bioinformatics, 21(1), pp. 104-115.
4. Fabbri, G., Riccardo D.-F. (2016). The molecular pathogenesis of chronic lymphocytic
leukaemia. Nature Reviews Cancer, 16, pp. 145-162.
3 All related resources are available at: https://github.com/skontopo/swat4ls2016
5. Huang, Z. X., Tian, H. Y., Hu, Z. F., Zhou, Y. B., Zhao, J., Yao, K. T. (2008). GenCLiP: a
software program for clustering gene lists by literature profiling and constructing gene co-
occurrence networks related to custom keywords. BMC Bioinformatics, 9(1), pp. 308.
6. Tenenbaum, J.; Silva, V. & Langford, J. (2000). A global geometric framework for nonlin-
ear dimensionality reduction. Science, 290(5500), pp. 2319-2323.
7. Van der Maaten, L. & Hinton, G. (2008). Visualizing data using t-SNE. Journal of Ma-
chine Learning Research, 9(85), pp. 2579-2605.
8. Kohonen, T. Self-Organizing Maps. Springer, 2001.
9. Leacock, C., Chodorow, M. (1998). Combining local context and WordNet similarity for
word sense identification. WordNet: An electronic lexical database, 49(2), pp. 265-283.
10. Wu, Z., Palmer, M. (1994). Verbs semantics and lexical selection. In Proceedings of the
32nd annual meeting on Association for Computational Linguistics (pp. 133-138). Associ-
ation for Computational Linguistics.
11. Jiang, J. J., Conrath, D. W. (1997). Semantic similarity based on corpus statistics and lexi-
cal taxonomy. arXiv preprint cmp-lg/9709008.
12. Resnik, P. (1995). Using information content to evaluate semantic similarity in a taxono-
my. arXiv preprint cmp-lg/9511007.
13. Sánchez, D., Batet, M., Isern, D., Valls, A. (2012). Ontology-based semantic similarity: A
new feature-based approach. Expert Systems with Applications, 39(9), pp. 7718-7728.
14. Dong, H., Hussain, F. K., Chang, E. (2009). A hybrid concept similarity measure model
for ontology environment. OTM Confederated Int. Conf., pp. 848-857, Springer.
15. Ashburner, M. et al. (2000). Gene Ontology: tool for the unification of biology. Nature
Genetics, 25(1), pp. 25-29.
16. Couto, F. M., Silva, M. J., Coutinho, P. M. (2007). Measuring semantic similarity between
Gene Ontology terms. Data & Knowledge Engineering, 61(1), pp. 137-152.
17. Lord, P. W., Stevens, R. D., Brass, A., Goble, C. A. (2003). Semantic similarity measures
as tools for exploring the gene ontology. Pacif. Symp. on Biocomputing 8(4), pp. 601-612.
18. Caniza, H. et al. (2014). GOssTo: a stand-alone application and a web tool for calculating
semantic similarities on the Gene Ontology. Bioinformatics, 30(15), pp. 2235-2236.
19. Yu, G. et al. (2010). GOSemSim: an R package for measuring semantic similarity among
GO terms and gene products. Bioinformatics, 26(7), pp. 976-978.
20. Wang, J. Z., Du, Z., Payattakool, R., Yu, P. S. Chen, C. F. (2007). A New Method to
Measure the Semantic Similarity of GO Terms. Bioinformatics, 23: 1274-1281.
21. Wittek, P. et al. (2015). Monitoring term drift based on semantic consistency in an evolv-
ing vector field. IJCNN-15, International Joint Conference on Neural Networks, pp. 1-8.
22. Wittek, P., Gao, S. C., Lim, I. S., Zhao, L. (2015). Somoclu: An efficient parallel library
for self-organizing maps. arXiv:1305.1422.
23. Song, X., Li, L., Srimani, P. K., Yu, P. S., Wang, J. Z. (2014). Measure the semantic simi-
larity of go terms using aggregate information content. IEEE/ACM Transactions on Com-
putational Biology and Bioinformatics (TCBB), 11(3), 468-476.
24. Wittek, P., Darányi, S., Liu, Y. H. (2014). A vector field approach to lexical semantics. In
Proceedings of QI-14, 8th International Symposium on Quantum Interaction, pp. 78-89.
25. International Human Genome Sequencing C. Finishing the euchromatic sequence of the
human genome. Nature. 2004;431(7011):931-45.
26. Schuster SC (January 2008). Next-generation sequencing transforms today’s biology. Nat.
Methods. 5.
27. Burger JA, Chiorazzi N. (2013). B cell receptor signaling in chronic lymphocytic leuke-
mia. Trends in Immunology, 34(12), pp. 592-601.