Divulgação - Defesa Nº 249

Aluna: Laila Barros Campos

Título: “Modelos de estimativa de afinidade de proteínas para projeto inteligente de fármacos”

Orientador: Wellington Pinheiro dos Santos - (PPGEC)

Examinador Externo: Giselle Machado M. Moreno - (USP)

Examinador Interno: Sidney Marlon Lopes de Lima - (PPGEC)

Data-hora: 22/Junho/2022 (14:00h)
Local: Formato Remoto (https://meet.google.com/kyz-ioyc-jcb)

---

Resumo:

As proteínas são objetos de estudo importantíssimos no âmbito das pesquisas biomédicas, uma vez que podem ter papel principal na descoberta de medicamentos e em diagnósticos de doenças. Esses compostos dificilmente atuam isoladamente enquanto desempenham suas funções, sendo assim muito comum formarem compostos entre si e os estudos das afinidades entre essas proteínas são bastante influentes nas descobertas e produções de novos fármacos antivirais e de vacinas. O objetivo principal desse projeto consiste em contribuir nesses estudos de afinidades entre proteínas propondo uma arquitetura profunda híbrida baseada em rede pseudo-convolucional para descrição de complexos de proteínas em imagens, rede neural convolucional para extração de características e regressores para estimar o grau de afinidade entre proteínas em um complexo. A etapa de pseudo-convolução extrai as sequências de RNA das proteínas de cada complexo, gerando duas matrizes de co-ocorrência que são posteriormente concatenadas. Essas matrizes então passam pelo processo de extração de atributos através de uma rede neural convolucional VGG19. Esses atributos passam a ser os dados de entrada para os treinamentos dos modelos de regressão que irão exercer a função de predição de afinidade entre as proteínas dos complexos. Dentre os quatro regressores treinados sendo estes o de Regressão Linear, Random Forest, Máquina de Vetor de Suporte (SVM - Support Vector Machine) e Multilayer Perceptron (MLP), os melhores resultados foram os de Random Forest e SVM. A melhor configuração do Random Forest foi a de 300 árvores, obtendo as melhores médias dos coeficientes de correlação Spearman (0,7067) e Kendall (0,5216) entre os dados preditos e os reais. A configuração do SVM com função kernel RBF, C = 0,1, g = 0,01 apresentou a melhor média do coeficiente de Pearson (0,6645) e a configuração desse mesmo regressor com função kernel RBF, C = 1,0 e g = 0,01 obteve o melhor RMSE (2,1383). É importante observar também a consistência dos resultados desses regressores, pois apresentaram baixos desvios padrão desses coeficientes.

Divulgação - Defesa Nº 248

Aluno: Thiego Buenos Aires de Carvalho

Título: “MapView - Exploring Datasets via Unsupervised View Recommendation”

Orientador: Fernando Buarque de Lima Neto - (PPGEC)

Co-orientador: Denis Mayr Lima Martins

Examinador Externo: Ana Carolina B. Salgado - (CESAR School)

Examinador Interno: João Fausto Lorenzato de Oliveira - (PPGEC)

Data-hora: 13/Junho/2022 (9:00h) - AM
Local: Formato Remoto (http://meet.google.com/czp-vucp-xcq)

---

Resumo:

Data are valuable assets to industries, government agencies, and research institutes. All these entities have a growing need of analyzing large data volumes that are generated from a variety of sources for helping users to communicate or to support their decision-making. Exploring even a simple database is not a trivial task, inasmuch as it requires technical knowledge which many new and non-technical data users do not have. This task includes writing SQL queries to retrieve a data set from larges database to uncover insights, patterns, and points of interest among them. Furthermore, in large volumes of data, finding valuable data that matches a certain user’s purpose requirement is challenging, especially under restrictive budget/time constraints. However, this task is typically manual, ad-hoc, and time consuming. To address these challenges, researches have proposed tools to support data exploration tasks, especially by means of View Recommendation. Under this research stream, a view can be seen as a visual representation of query’s result on database. Instead of showing a set of results produced by a query over a database, as a table like SQL represents, the result-set is then plotted using histogram or bar chart. Systems that use this approach start by creating all possible views, filter out non-informative candidates and recommend the most interesting views according to some objective functions. The goal of those solutions is to improve data exploration by guiding the user, showing the next best view to be explored, enabling users to quickly understand the data and find insights. View Recommendation is especially challenging in the context of Data Marketplaces since every data interaction incurs monetary cost. Due to this, instead of an iterative process of querying and analyzing unrelated views, each of which the user must pay for, a more suitable approach would consider a recommendation of bundles of related views. In this work, we propose and implement a new approach for View Recommendation called MapView, which is based on Self-Organizing Maps (SOM) and helps non-technical users with both technical expertise and time limited, in data exploratory tasks. Our proposed approach employs SOM as a clustering mechanism to group and recommend exploratory data views to users. This recommendation process can also be personalized to help meeting user’s intention in an interactive manner. To address View Recommendation in Data Marketplaces, we introduce the problem of recommending view bundles. In particular, we focus in cases where the data consumer’s budget to interact with the marketplace is limited. We investigate data exploration tasks that require several iterations to uncover valuable insights in the data, where view bundle recommendation allow for a multi-perspective view of the target data without overflow the user’s budget. We also investigate how SOM and Genetic Algorithm could be combined to recommend near-optimal view bundles while took a specified cost limit into account. The experimental results show that MapView is effective in recommending valuable views, hence, being of aid in data exploration tasks. Complementary views are recommended according to the user’s interest. This even within a tight budget.

Divulgação - Defesa Nº 247

Discente: Henrique José Amorim de Andrade

Título: “Synthesis of Satellite Images from the Past”

Orientador: Prof. Bruno José Torres Fernandes - (PPGEC)

Examinador Externo: Martijn Zoet - (Zuyd University of Applied Sciences)

Examinador Interno: Carmelo José A. Bastos Filho - (PPGEC)

Data-hora: 25/Maio/2022 (8:30h) - AM
Local: Formato Remoto (https://meet.google.com/tes-ytrh-wsa)

---

Resumo:

Historical maps are the primary source for spatial information. These documents play an important role in the conservation and understanding of history. However, a broad access to these documents is limited, firstly because only a few copies were digitized, and secondly because these documents often have a dated language and visuals. In a world used to satellite images and GPS navigation in mobile phones, we hypothesize if an architecture using neural network can reinterpret a historical map, presenting its information as a satellite image from the past. Although previous works have been handling information from historical maps, as well as creating reinterpretations of such documents, this work presents an architecture of generative adversarial networks for transferring style between satellite images and historical maps. The literature demonstrates the use of these networks in the creation of convincing artificial media for a common observer, popularized by the DeepFake trend in the last years. This work also proposes a creation of a satellite imagery collection - for training the networks - whose aspects would be transferred to the final synthesized output. We present an architecture to synthesize satellite images from the past, and discuss its steps. Finally, we present and discuss synthesized satellite images of Recife, Brazil, from the past and how this technique can be a tool for other areas of study and the public debate.