![](bgb.png)
ment of structural properties of pathways. Such an
assessment is performed by the DNN model. It is dif-
ficult to imagine how the same assessment could be
done through an algorithm on graphs since the struc-
tural properties to be assessed are not known in ad-
vance, but inferred.
In order to consolidate the results we obtained, as
future work we plan to extend the investigation of ro-
bustness to a dataset in which also very large graphs
(> 40 nodes) are included. On one hand, this will be
challenging from a computational point of view, be-
cause some biological networks included in the orig-
inal dataset comprise a number of nodes in the order
of hundreds and thousands. On the other, having a
large dataset will probably be beneficial to our model,
since the effectiveness of Deep Neural Networks is
generally proportional to the number of training ex-
amples. Another line of research to pursue concerns
model explainability. Indeed, our DNN has thousands
of parameters, which make explaining the “why” be-
hind their predictions (i.e. which parts of the pathway
contributed to the prediction, and to what extent) a
hard task. Motivated by this challenge, we plan to de-
velop generative models of pathway networks to work
towards the goal of making these models explainable.
Furthermore, we will consider enriching the
dataset with information we have omitted in the
present study. In particular, we may include arc la-
bels (multiplicities of reactants/products) in order to
evaluate their significance. Moreover, we may in-
clude something about kinetic formulas, such as their
parameters (properly normalized). The latter addition
could, in principle, improve the accuracy of the model
on small subgraphs, but its effect on the accuracy of
big ones has to be carefully evaluated.
Lastly, we plan to apply the approach to the as-
sessment of other dynamical properties such as other
notions of robustness as well as, for example, mono-
tonicity, oscillatory and bistability properties.
ACKNOWLEDGEMENTS
This work has been supported by the project
“Metodologie informatiche avanzate per l’analisi di
dati biomedici” funded by the University of Pisa
(PRA 2017 44).
AUTHORS’ CONTRIBUTION
The four authors contributed equally to this work.
REFERENCES
Fages, F. and Soliman, S. (2008). From reaction models to
influence graphs and back: A theorem. In Fisher, J.,
editor, Formal Methods in Systems Biology, pages 90–
102, Berlin, Heidelberg. Springer Berlin Heidelberg.
Fey, M. and Lenssen, J. E. (2019). Fast graph represen-
tation learning with PyTorch Geometric. In ICLR
Workshop on Representation Learning on Graphs and
Manifolds.
Gilbert, D., Heiner, M., and Lehrack, S. (2007). A unifying
framework for modelling and analysing biochemical
pathways using petri nets. In Calder, M. and Gilmore,
S., editors, Computational Methods in Systems Bi-
ology, pages 200–216, Berlin, Heidelberg. Springer
Berlin Heidelberg.
Gillespie, D. T. (1977). Exact stochastic simulation of
coupled chemical reactions. The journal of physical
chemistry, 81(25):2340–2361.
Glorot, X., Bordes, A., and Bengio, Y. (2011). Deep sparse
rectifier neural networks. In Proceedings of the Four-
teenth International Conference on Artificial Intelli-
gence and Statistics, pages 315–323.
Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep
Learning. The MIT Press.
Gori., R., Milazzo., P., and Nasti., L. (2019). Towards an
efficient verification method for monotonicity proper-
ties of chemical reaction networks. In Proceedings of
the 12th International Joint Conference on Biomedi-
cal Engineering Systems and Technologies - Volume
3: BIOINFORMATICS,, pages 250–257. INSTICC,
SciTePress.
Hucka, M., Bergmann, F. T., Dr
¨
ager, A., Hoops, S., Keat-
ing, S. M., Le Nov
`
ere, N., Myers, C. J., Olivier, B. G.,
Sahle, S., Schaff, J. C., et al. (2018). The systems
biology markup language (sbml): language specifica-
tion for level 3 version 2 core. Journal of integrative
bioinformatics, 15(1).
Iooss, B. and Lema
ˆ
ıtre, P. (2015). A review on global sen-
sitivity analysis methods. In Uncertainty management
in simulation-optimization of complex systems, pages
101–122. Springer.
Janocha, K. and Czarnecki, W. (2017). On loss functions
for deep neural networks in classification. Schedae
Informaticae, 25.
Karp, P. D. and Paley, S. M. (1994). Representations of
metabolic knowledge: pathways. In Ismb, volume 2,
pages 203–211.
Kingma, D. P. and Ba, J. (2015). Adam: A method for
stochastic optimization. In Proceedings of the 3rd In-
ternational Conference on Learning Representations,
ICLR.
Kipf, T. N. and Welling, M. (2017). Semi-supervised classi-
fication with graph convolutional networks. In 5th In-
ternational Conference on Learning Representations,
ICLR.
Kitano, H. (2004). Biological robustness. Nature Reviews
Genetics, 5(11):826.
Kitano, H. (2007). Towards a theory of biological robust-
ness. Molecular systems biology, 3(1).
BIOINFORMATICS 2020 - 11th International Conference on Bioinformatics Models, Methods and Algorithms
42