Skip to content

SARC INTERNATIONAL CONFERENCE – Proceedings

Papers

  1. An Empirical Analysis of Different Protein Folding Algorithms on Cytokine Protein Structure

    Kazi Toufique Elahi, Swapnil Sharma Sarker, Raufun Talukder Raktim, Anika Tasnim Aurin, Shamim Akhter

    Paper ID SA-MLAI-DHKA-130123-200 · Pages 31–61

    Abstract

    Numerous applications, including identifying protein misfolding, protein docking, and comprehending the function of proteins, depend on accurate prediction of the three-dimensional structure of proteins. Cytokines, a type of protein involved in intercellular communication in the immune system and inflammation, are of special interest due to their medicinal significance. Accurate prediction of these protein types can be very resourceful in medical studies. Several deep learning-based algorithms for predicting protein structures from amino acid sequences, such as Alpha Fold 2, Omega Fold, and ESM Fold, have been developed in recent years. There has been significant research in this field, such as benchmark studies of nanobodies[6], comparative study of deep learning based model ESM2 and Homology modeling based Swiss Model[9], evaluating protein docking using Alpha Fold 2[7] and so on. But there has not been a study that analyzes the performance of these models on Cytokines. In this work, we assessed how well these models performed at predicting the Cytokine structures. Overall, these results highlight the importance of considering both accuracy and speed when selecting a protein structure prediction model. Keywords - Protein Folding, Cytokines, Structure Prediction, Alpha Fold, Omega Fold, ESM Fold.

    Download PDF
  2. Handwritten Word Recognition using Different Architectures of Convolutional Neural Network

    Bilash Chakraborti, Romana Rahman Ema, Sk. Shalauddin Kabir, Tajul Islam, Abu Rafe Md Jamil

    Paper ID SA-MLAI-DHKA-130123-3469 · Pages 36–71

    Abstract

    A language is mainly a combination of words. To recognize a particular means to recognize the words. Word recognition is very important for every language. The characters that are included in a character's image are identified by handwriting recognition software [1]. Handwritten word recognition (HWR) is not done accurately still. In previous times, the used technique was less accurate. To solve the problem, we proposed a technique in the proposed method. Five architectures of the Convolutional neural network (CNN) have been used in the proposed method. Those are ResNet-50, Alex Net, VGG-16, Efficient B2, and Efficient B3. Moreover, to train all models we use 2 datasets. First is Zilla (64), and Test (10). In the Zilla (64) dataset, there are 64 classes and almost 5000 input images. The 64 classes define 64 districts of Bangladesh. For Zilla (64) dataset, we get 92.12% validation accuracy from ResNet-50, 85.77% validation accuracy from Alex Net, 97. 69% validation accuracy from Efficient B2, 92.22% validation accuracy from Efficient B3, 77.31% validation accuracy from VGG-16. Almost every model gives a good accuracy rate compared to the existing work. A little bit of error was found because of the resemblance and lofty curvature nature of the data.

    Download PDF