Shaowen Zhu

Intro

My name is Shaowen Zhu and I am a 3rd-year PhD student from Dr. Yang Shen's Lab in Texas A&M University. I got my bachelor's degree in Electrical Engineering from Harbin Institute of Technology and a MS degree from my current Lab. My research interests are related to machine learning and bioinformatics, mainly covering computational protein design, graph representation and sequence modeling and generation. More details on my reseach topics can be found in my projects.

Experience

Research Assistant / Student Worker

Worked on projects related to machine learning and bioinformatics, especially for graph representation learning,generative models and computational protein design.

Summer Intern

Designed the PCBs and adjust the PID parameters of the controllers of the products (several terrariums). Tested and assembled the battery test instruments.

Projects

Fold-based De Novo Protein Design

This motivation of this project is to design the protein sequences given the target protein folds (topology of the structures). By learning the relationship between the protein sequences and folding structures we would like expand the sequence space for known folds or even discover novel folds which is called de novo design. We applied kernel-PCA method to learn the protein represeantations is a low dimensional space and trained generative models to learn the conditional sequence distribution. We developed a conditional Wasserstein generative adversarial network (cWGAN) and improved it to the guided cWGAN (gcWGAN) with a pretrained oracle for the feedback. On test set covering over 100 test fold. Our methods successfully design the proteins on 3.5 times more targets than the SOTA methods. The paper was accepted to JCIM as a cover paper and available here, and the project is available here.

CASP14-CAPRI.

I attended the CASP14-CAPRI competition which is to predict the protein-protein or protein-DNA docking structures (prediction tasks) and rank the predicted models (scoring tasks). I took charge of the scoring tasks and assited on the prediction tasks. I scored the predicted assmblies with EGCN which is a GNN-based model developed by our group. I transformed the molecules into graph formats and retrained the models for varies versions for different purpose. The models will provide scores for ranking. Our team got the 7th for the scoring task and the 15th (only 1-point shy of the top 10) for the prediction tasks in Round 50, and the published paper is available here. Currently I am still attending the challenges of the following rounds./p>

Intellectual Disability (ID) Panel Challenge of CAGI6

I attended the CAGI6 competition with my group and took charge of the ID panel challenges. I was asked to make ID predictions given multiple gene mutations for each sample (patient). For a patient there can be over 4 hundred mutations (while many of them are unique) covering 74 genes and it is a few-shot learning process with 150 training smaples for 400 test samples. I transformed the sample representation problem as sequence embedding problem, and applied pretrained model like DNABert for the representations. And I applied pre-learned gene-gene interaction (GGI) networks as the startpoint, and then finetuned severl GNN models on the training samples. The outputs will be sent into an MLP for the final predictions. The challenge is under review and the description is available here.

Contact

• Twitter
• Facebook
• Instagram
• GitHub

Elements

i = 0;

while (!deck.isInOrder()) {
    print 'Iteration ' + i;
    deck.shuffle();
    i++;
}

print 'It took ' + i + ' iterations to sort the deck.';