Poster Presentation The 16th Australian Peptide Conference 2025

Developing a machine learning model to predict peptide-ligand binding affinities to improve plasma protein purification (#226)

Elizabeth Phua 1 2 , Andrea Robinson 1 2 , Katya Pas 1 2 , Zhenghui Xu 3
  1. Chemistry, Monash University, Clayton, VIC, Australia
  2. ARC Training Centre for Green Chemistry in Manufacturing, Monash University, Clayton, VIC, Australia
  3. CSL Innovation Pty Ltd, Melbourne, VIC, Australia

Blood plasma proteins maintain physiological and chemical homeostasis in the human body. Isolated from plasma donations, plasma proteins are necessary to treat conditions from immunodeficiencies to rare diseases.1,2 However, on scale, these proteins are isolated from laborious multi-step cold ethanol fractionation processes, resulting in low product recoveries and large volumes of waste.2

Alternative purification methods like affinity chromatography (AC), are being explored to improve purification efficiency, to meet the rising demand for plasma protein therapies.4 However, challenges in obtaining peptide-affinity ligands with high specificity and selectivity towards targeted plasma proteins must be addressed to make purification processes more cost-effective and streamlined.

The current strategy for compound discovery relies on resource-intensive ‘trial-and-error’ methods requiring extensive synthetic endeavour to identify small numbers of ‘hits’ within large arrays of low-quality and low-quantity ligands.6 Alternatively, machine learning (ML) techniques can be used to accurately predict the properties of ligands, expedite selection of ligands with requisite functionality, and design bespoke ligands.7,8

A ML model is being developed to predict binding affinities of peptide-affinity ligand candidates for their use in AC towards the isolation of targeted plasma proteins. A dataset of varying ligands and their binding affinities has been collated from literature, open-sourced data bases, and results from novel experimental data. This dataset has been used to train, test, and validate several ML models based on widely used fingerprint descriptors that represent the chemical structure and information on bond types and atom connectivity.

This presentation will highlight the key challenges faced in the development of a robust ML model, ensuring comparability across reported and currently collected datasets, and overcoming difficulties relating to the synthesis, purification, and analysis of peptide-ligand candidates. This industry-aligned project will facilitate maximal therapeutic yields from donations and increase treatment availability for patients who depend on them for survival.

  1. A. Farrugia, Transfus Med Rev; 2010, 24(1), 53-63.
  2. J. Curling, N. Goss & J. Bertolini, Production of Plasma Proteins for Therapeutic Use; 2013, 1-28.
  3. P. Akbarzadehlaleh, A. Kamani & K. Shamsasenjan, Sep Sci Technol; 2022. 57(15), 2491–2499.
  4. T. Burnoff & M. Radosevich, J Biochem Biophys Methods; 2001, 49(1-3), 575-586.
  5. D.G. Brown & J. Boström, J Med Chem; 2016, 59(10), 4443-4458.
  6. K. Low, R. Kobayashi & E. I. Izgorodina. J Chem Phys; 2020, 153(10), 104101.
  7. J. Shim, Z.-Y. Hong, I. Sohn & C. Hwang, Sci Reports; 2021, 11, 4416.
  8. S. Nazari and A. Abdelrasoul, Scientific Reports, 2025, 15, 3474.