AI Lab Research, AIDD (AI Drug Design)

ProtesinSora: Protein Moldecular Dynamics Trajectory Generation Based on Diffusion

(Supervisor: Quanquan Gu)

• Proposed a framework for protein molecular dynamics trajectory generation based on Diffusion.
• Designed three novel metrics for protein molecular dynamics trajectory evaluation.
• Improved the accuracy and realism of protein molecular dynamics trajectory through physical prior knowledge and particular loss functions.
• Researched methods for long molecular dynamics protein trajectory generation.

MSM State Transitions	MD Trajectory Generation

Department of Automation & School of Integrated Circuits, Tsinghua University

Tracking and classification of microglia under wide field microscope

(Supervisor: Qionghai Dai)

• Trained U-Net for tracking and segmentation for video of microglia in motion recorded by wide field microscope and contrastive learning network (SimSiam) for classification of cell subtypes.
• Researched new algorithms of data enhancement for positive samples needed in Simsiam based on GAN.
• Designed transfer learning method between two neural networks.

Modeling DNA sequences with natural language models

(Supervisor: Xiaowo Wang)

• Explored efficient methods for DNA sequence tokenization with reference to the BPE algorithm.
• Built Neural Networks with multi-layer CNN, LSTM and CRF to perform various downstream tasks related to DNA sequences.
• Fine-tuned the DNABERT family of models and found motifs in DNA sequences as the attention scores recorded.

Computing-in-memory chip with memristor

(Supervisor: Bin Gao)

• Applied convolutional neural network on chip based on memristor implementation.
• Tested the memristor chip and counted the number of bad tracks using the given SDK.