(184k) Gaussian Process Regression Modeling to Predict ?/?-Peptide Selectivity for Antifungal Drug Development

Since the introduction of insulin to treat type 1 diabetes 100 years ago, large strides have been made in the field of peptide therapeutics to treat a range of other diseases such as cancer, multiple sclerosis, and HIV.¹ A pressing issue in medicinal research today is leveraging the knowledge of current peptide synthesis techniques to design antifungal drugs. Given the recent development of drug-resistant strains to current treatments on the market, there is large interest in developing antimicrobial peptides (AMPs) that work via membrane disruption and are therefore less likely to lead to resistance.² However, naturally sourced α-helical AMPs have low stability in vivo due to their susceptibility to proteolytic cleavage, and they exhibit low selectivity for fungal versus human cells.³ Consequently, new approaches to design synthetic peptides are needed to tackle these challenges.

One promising class of synthetic peptides is α/β-peptides, where α denotes traditional amino acids while β denotes modified amino acids with an additional backbone carbon. An advantage of these materials is the large array of α/β combinations that can be tested, each with differing stability profiles, distributions of side chains, etc. The goal in designing α/β-peptides for antifungal applications is to strike a balance in maximizing the activity against fungal pathogens – quantified as decreasing the minimum inhibitory concentration (MIC) of peptide required to prevent pathogenic cell growth – while minimizing hemolysis of human red blood cells – quantified as increasing the peptide concentration required to lyse 10% of red blood cells (HC₁₀).⁴ Maximizing the ratio of HC₁₀ to MIC through a metric known as the Selectivity Index (SI) is desired during the rational development of new antifungal peptides.^3,4 Recent experimental studies have shown that only a handful of measured quantities such as helicity and liquid chromatography retention time can be utilized to accurately predict fungal activity and hemolysis metrics.^3,5 However, these experiments require α/β-peptide synthesis and consequently have a large time and financial cost, limiting their utility in large-scale model development.

To address this challenge, we hypothesized that computationally derived descriptors could be utilized to more quickly and accurately predict HC₁₀ and MIC measurements prior to α/β-peptide synthesis. Using a dataset of 150 published α/β-peptide sequences as a starting point, we built a Gaussian Process Regression (GPR) workflow to predict HC₁₀ and MIC values by calculating 2D molecular descriptors with the RDKit cheminformatics toolkit and simplified SMILES string representations of α/β-peptide sequence. We developed and implemented an iterative model training approach to predict new peptide experimental labels using a design space motivated by the most selective (highest SI) sequences in the initial training set. Peptides were selected for synthesis based upon GPR predictions, with measured HC₁₀ and MIC values then used to update GPR model parameters to guide the next round of predictions. The advantage of this workflow is that GPR provides estimations of uncertainty for prospective sequences, allowing us to probe both promising new sequences with low uncertainty and high predicted SI while also testing sequences with high uncertainty to expand the design space (e.g., new amino acids) for future rounds. We show that this approach identifies new selective α/β-peptides and sequence features important to HC₁₀ and MIC predictions. These results demonstrate the potential of this approach to screen for new highly selective sequences for peptide drug development.

(1) Muttenthaler, M.; King, G. F.; Adams, D. J.; Alewood, P. F. Trends in peptide drug discovery. Nature Reviews Drug Discovery 2021, 20 (4), 309-325. DOI: 10.1038/s41573-020-00135-8.

(2) Benfield, A. H.; Henriques, S. T. Mode-of-Action of Antimicrobial Peptides: Membrane Disruption vs. Intracellular Mechanisms. Frontiers in Medical Technology 2020, 2, Mini Review.

(3) Lee, M.-R.; Raman, N.; Gellman, S. H.; Lynn, D. M.; Palecek, S. P. Incorporation of β-Amino Acids Enhances the Antifungal Activity and Selectivity of the Helical Antimicrobial Peptide Aurein 1.2. ACS Chemical Biology 2017, 12 (12), 2975-2980. DOI: 10.1021/acschembio.7b00843.

(4) Murugan, R. N.; Jacob, B.; Ahn, M.; Hwang, E.; Sohn, H.; Park, H.-N.; Lee, E.; Seo, J.-H.; Cheong, C.; Nam, K.-Y.; et al. De Novo Design and Synthesis of Ultra-Short Peptidomimetic Antibiotics Having Dual Antimicrobial and Anti-Inflammatory Activities. PLOS ONE 2013, 8 (11), e80025. DOI: 10.1371/journal.pone.0080025.

(5) Chang, D. H.; Lee, M.-R.; Wang, N.; Lynn, D. M.; Palecek, S. P. Establishing Quantifiable Guidelines for Antimicrobial α/β-Peptide Design: A Partial Least-Squares Approach to Improve Antimicrobial Activity and Reduce Mammalian Cell Toxicity. ACS Infectious Diseases 2023, 9 (12), 2632-2651. DOI: 10.1021/acsinfecdis.3c00468.