04/02/2021
By Aravindan Balaji

The Kennedy College of Sciences, Department of Computer Science, invites you to attend a Master's thesis defense by Aravindan Balaji on "PepGAN: A Pathway to High Quality Peptide Drugs."

Candidate Name: Aravindan Balaji
Degree: Master's
Defense Date: Thursday, April, 22, 2021
Time: 10 a.m. to noon EST
Location: Contact Aravindan_Balaji@student.uml.edu at least 24 hours before the defense to request access to the zoom meeting.
Thesis Title: PepGAN: A Pathway to High Quality Peptide Drugs

Committee:

  • Wenjin Zhou (Advisor), Department of Computer Science, University of Massachusetts Lowell
  • Benyuan Liu, Department of Computer Science, University of Massachusetts Lowell
  • Mohammad Arif Ul Alam, Department of Computer Science, University of Massachusetts Lowell

Brief Abstract:

Computational Drug Design is gaining traction in the pharmaceutical industry since it provides a good starting point for new drugs and speeds up the time to market of a drug. Peptides are short chains of between two and fifty amino acids linked by peptide bonds. They are used in immunotherapy to treat multiple types of cancer. Therapeutic peptides react with targeted receptors without any side effects making them a better alternative to small molecule drugs. The ability to generate better peptide drugs with short time to market will be a big boost to cancer treatment.

At present, Generative Adversarial Networks is used with reinforcement learning techniques to generate drug sequences but the binding affinity and active atom distance is not satisfactory since the generator does not have enough guidance from discriminator on how to improve its generation quality. Similarity of CTLA4 drugs generated to FDA approved drugs is also below the threshold range (50% - 90%) which needs to be improved. To combat these issues, we propose a GAN framework called PepGAN in which discriminator reveals the features of the generated sequences to generator so that it gets a better idea of improvement needed rather than just sending a scalar reward back to generator. Moreover, we include a Judge module which selects the best candidate amino acid to be generated next at each generation step using cosine distance between real data and 20 natural amino acids eliminating the randomness in sampling. Peptide generation experiments for PD1, PDL1 and CTLA4 show a great improvement in binding affinity and active atom distance compared to its predecessor GANDALF and similarity of CTLA4 drugs now fall within the desired range.