Machine Learning Powered Platform for Discovery of Peptide Ligands for Peptide-Drug Conjugates

Robust Diagnostics Llc, Phoenix AZ

Investigators

Abstract

Project Abstract Peptide drug conjugates (PDCs) are an emerging approach for targeted cytotoxic delivery, offering higher tumor penetration and lower manufacturing costs compared to antibody-drug conjugates (ADCs). However, the discovery and optimization of PDCs remains inefficient due to the linear nature of peptide hit-to-lead workflows. Existing peptide discovery methods, such as phage or mRNA display not only require several enrichment cycles but yield hits that lack selectivity and stability resulting in additional time for downstream optimization. Moreover, data from display techniques suffer from high false positive rates, lack of negative controls, and amplification biases, limiting their utility for machine learning (ML)-driven improvements. To address these challenges, we propose to join the complementary strengths of Robust Dx’s peptide microarray expertise and Koliber’s ML optimization capabilities to create a next-generation peptide discovery system capable of identifying PDC leads within three months. The proposed discovery platform utilizes a peptide array designed for enhanced protease stability through the incorporation of mixed L- and D-amino acids, with sequences selected using unsupervised ML approaches to achieve high diversity and uniform coverage of the design space. The platform utilizes advanced assays developed by RobustDx to accurately measure target and off-target binding while also estimating kinetic parameters. Preliminary modeling by Koliber using random, L only array datasets shows feasibility of ML model training and ability to discover distinct binding patterns enabling rational selection of diverse hits. In this Phase I proposal, we aim to demonstrate that the designed peptide arrays can effectively identify stable and specific binders to a promising PDC target, Nectin-4. Furthermore, we will show that the discovery data can be utilized to train ML models to further optimize potency and specificity. In Aim 1 we will produce a 2,000-peptide microarray of mixed L- and D-amino acids and discover hit peptides for Nectin-4. In Aim 2 we will validate the hits via SPR for potency, selectivity and evaluate serum stability. And in Aim 3 we will demonstrate that we can use ML models trained on the discovery data to improve the hits to achieve sub µM KDs and over 5-fold specificity. Successful completion of Phase I will establish the feasibility of a system that can rapidly and reliably result in protease stabilized, highly selective and potent PDC leads. In Phase II we will further improve the library designs, refine the ML models to enhance prediction accuracy and begin to commercialize the platform through partnering with biopharma companies to address a broader range of targets and advance novel PDC candidates to the clinic.