[CS] Songhao Jiang Candidacy Exam/May 17, 2024
meganwoodward at uchicago.edu
meganwoodward at uchicago.edu
Fri May 3 09:27:54 CDT 2024
This is an announcement of Songhao Jiang's Candidacy Exam.
===============================================
Candidate: Songhao Jiang
Date: Friday, May 17, 2024
Time: 10 am CT
Remote Location: https://uchicago.zoom.us/j/97665214384?pwd=UG81Z2ljUEVkMmpkVVMzSFR2QUFYdz09 Meeting ID: 976 6521 4384 Passcode: 043310
Location: JCL 298
Title: Leveraging Transformer Models for Accelerated Drug Discovery
Abstract: In the realm of AI-accelerated drug discovery, particularly in de novo drug design, significant challenges include unpredictable drug responses in clinical trials, biases in predictive models, and the opaque nature of AI methodologies that complicate the understanding of a drug's mechanism of action. These issues have limited the progression of AI-discovered drugs into clinical trials and regulatory approval. Concurrently, the development of me-too drugs, which involve modifications of existing drugs within the same therapeutic class, presents a less risky and potentially more effective avenue. However, the potential of AI to enhance their development remains largely underexplored.
This dissertation aims to transform the development of me-too drugs through the application of AI, with a focus on transformer and large language models (LLMs). It introduces innovative frameworks that utilize the representation learning and generative capabilities of transformer models to refine and expedite the me-too drug development process. These methodologies, referred to as "drug optimization", seek to further accelerate the production of effective me-too drugs.
This work makes four significant contributions to the field: (1) It proposes two fusion methods that integrate transformer models with graph neural networks, enhancing the precision of binding affinity predictions. (2) It assembles a comprehensive dataset of 10 million binding affinity values across a diverse array of proteins and drugs, providing an invaluable resource for model training and validation. (3) It proposes two generative models for drug optimization, fine-tuned through reinforcement learning, with the goal of automating and expediting the creation of effective me-too drugs. (4) It introduces an innovative bidirectional GPT model for molecular textual sequences (SMILES), enabling precise generative mask infilling for targeted drug optimization. And by conducting comprehensive evaluations on real world viral and cancer target proteins, we demonstrate that the proposed drug optimization frameworks can consistently enhance existing molecules/drugs.
Advisors: Rick Stevens
Committee Members: Rick Stevens, Ian Foster, and Fangfang Xia
More information about the cs
mailing list