Systems Biology and Bioinformatics & Biobased Materials @UM

Deep learning-based prediction of molecular subtypes​ in human ovarian cancer
Project by: Georgia Liapi 

Ovarian cancer (OV), a disease clinically characterized by complex histological and molecular features, is often difficult to fully diagnose. Early diagnosis is of great importance, as 70 % of OV patients are diagnosed late, with an advanced disease. During diagnosis, pathologists microscopically inspect Haematoxylin and Eosin stained images from a patient’s tumor tissue, to decide about the OV histological subtype, while in parallel complex molecular tests are performed in tumor samples of this patient. 

Pathologists are able to recognise the five most common OV histological subtypes in histology images, but visually distinguishing the underlying molecular subtype from a given histology image of an OV patient is a very demanding task. Nowadays, OV histology image datasets are publicly available, with studies providing a label, either an OV histological or molecular subtype for each of these images. Previous studies have used Deep convolutional neural networks (CNNs) in cancer-focused and image-based classification tasks, with promising results.

In the current study, six different pre-trained CNNs were retrained to classify OV histology images into five main OV histological subtypes. The CNNs with the best classification performances were trained again to classify images of a certain OV histological subtype (High - grade Serous Ovarian Carcinoma) into four main molecular subtypes. ShuffleNet and Xception, achieved high Areas Under Receiver Operating Curves (AUROCs) , per subtype, in the histological subtyping, but encountered difficulties in the molecular subtyping, reflected by low AUROCs and by the number of misclassified patients . Deep learning - based OV molecular subtyping remains a challenge, demanding further image preprocessing, before re - training a CNN, to use it for substituting traditional molecular tests . 

Assessing Outcomes and Risk Factors Of Atrial Fibrillation Through A Lifetime Population-Level Markov Model.
Project by: Cristian Barrios Espinosa 

Atrial fibrillation (AF) is characterized by the abnormal electrical activity of the atria in the heart. It presents a major global health burden because it is one of the most common heart diseases that carries great economic and human cost. The development of new kinds of research in AF is very relevant since the treatments are still very inefficient. Computational and systems biology approaches are increasingly used to study AF.

These approaches include mechanistic studies (computational and experimental), usually at the molecular and cellular level, and clinical studies dealing with long-term patient outcomes. The former allows for understanding the pathophysiology of AF in a few seconds or minutes without focusing on the long-term effects on the patient. The latter unravels important clinical consequences of AF (e.g. increase of the risk of mortality and stroke) without considering the mechanisms that lead to these outcomes. My thesis aimed to bridge the gap between these two kinds of knowledge by developing a new type of patient-level Markov model, representing the transitions between different clinical states.

The model simulates a virtual population of patients over a 100-year lifetime period assessing clinical outcomes, risk factors, and mechanisms together. In the figure below there is a schematic representation of the model. In this model, the risk factors for AF that are considered are: age, sex, and sleep-disordered breathing (SDB); the only clinical outcome included is mortality;  and the model simulates changes in electrical characteristics and structural properties (fibrosis) that play a central role in the mechanisms of AF.

The model was calibrated to reproduce real-world data with respect to the epidemiology of AF and SDB, and mortality in the general population and AF patients. Finally, the calibrated model was used to analyze the potential therapeutic effects of different types of risk-factor management for SDB. This was done by conducting ‘virtual clinical trials’ that would be challenging in the real world for economical or ethical reasons. Conclusion: We successfully developed a model that connects mechanistic and epidemiological research in AF and show how this model can be used to perform a ‘virtual clinical trial’ to generate new hypotheses that can be used to guide new clinical research and ultimately lead to better AF management.

Research Internship 
Project: Raphael Stolpe 

During my internship, I studied the behavioural aspect of biological systems. In computational neuroscience, a subfield of systems biology, neural phenomena underlying behaviour are studied. Neural phenomena represent computations necessary for a biological system to produce behaviour. Yet, computational modeling of neural phenomena does not account for functionality, i.e. meaningful interaction with an environment. Recent advances in deep reinforcement learning however opened up the possibility to generate complex, functional models capable of interacting fluently with an environment.

As a remaining caveat, deep learning models lag interpretability and do not account for neural phenomena. In my internship, I explored if dynamical systems theory, a mathematical field taught in the systems biology master, can be utilized to explain goal-oriented behaviour and pave the way to more biological realism. To that end, I applied the approach to a set of reinforcement learning problems including control of a simulated anthropomorphic hand. I found that point attractor networks and central pattern generators, special cases of dynamical systems previously described in the fields of dynamical neuroscience and robotics, are common behavioural strategies. The research during my internship was fascinating and sparked my interest to pursue a PhD in this exciting field.

Processing of tree-like structures via DLP using biocompatible-biodegradable resins for tissue regeneration
Project by: Marianna Arreguín Campos

Vascularization of scaffolds remains a critical challenge in the field of tissue engineering, preventing scientists to develop solutions for larger anatomical defects. Hydrogels, such as PEGDA or GelMA represent important materials for scaffolds due to their mechanical properties, biocompatibility, and biodegradability. Although DLP has been proposed as an efficient, low-cost and fast technique for manufacturing complex structures with very high resolution, tunable biocompatible resins are still required.

In this project, we combine different molecular weights of PEGDA to tailor on the one hand a variety of mechanical properties, which are an important requirement for materials that replace different types of native tissue, and on the other hand the swelling of the polymer which is another crucial parameter for vascular structures. This study demonstrates that it is possible to modify and potentially tune the mechanical properties of the hydrogels, as well as their swelling behavior, by combining varying ratios of different molecular weight PEGDA. The prepared PEGDA resin formulations proved to be easily printable with DLP at room temperature. The resin formulations were tested on a commercial printer and a custom-designed one, which proved to be the best choice for this research. Micrometer-sized structures of vascular scaffolds were obtained

Optimization of recombinant spider silk production in metabolically engineered tobacco
Project by: Nika Sokolova

An exceptional combination of strength and elasticity, together with intrinsic biocompatibility and biodegradability, make spider silk a coveted material in various biomedical and technical applications. The inability to farm spiders on a large scale has triggered two decades worth of research into heterologous expression of both native and synthetic spider silk proteins, which are notoriously hard to produce due to the repetitive nature of the gene and large, aggregation-prone amino acid sequence.

Out of all potential expression hosts, plants exhibit several inherent advantages such as stable gene expression, intracellular targeting, organ-specific accumulation; they are relatively cheap, safe and easy to maintain. Currently, large-scale molecular farming of spider silk is hampered by low yields. Here, a metabolic engineering approach was developed to optimize the production of spider silk proteins in tobacco. The project involved the use of techniques from bioinformatics, molecular biology, genetic engineering, plant biotechnology, DoE (design of experiments) and statistical analysis.

Hybrid composite electrode for supercapacitor
Project by: Bernal Garcia Lascurain

Supercapacitors have been acknowledged to be substitutes for Lithium-Ion batteries since they display improvements in stability, faster charge-discharge periods, and a lower environmental impact. In this research work, a new hybrid composite electrode for supercapacitor was developed by combining Alkali Lignin with a Selectively Oxidized Graphene (SOG) into a nanocomposite.

Here, it’s demonstrated an excellent synergic effect between these materials through electrochemical measurements giving an areal capacitance of 7.65 mF cm^-2 compared to 4.12 mF cm^-2 from pure SOG both at 5 mV s^-1. Then, a homemade device was fabricated using a free-standing self-healing hydrogel with Sodium Carboxymethyl Cellulose and SOG, showing a notable increase in frequency response due to a better ion-diffusion from its porous structure. Both devices showed outstanding stability after thousands of charge-discharge tests, which, in conjunction with the aforementioned results, make these materials a suitable green solution for energy storage.