Free papers detail | 11th Asia Pacific Tele Ophthalmology Society Symposium

Title
A Generative Framework for De Novo Discovery and Potency Optimization of Antimicrobial Peptides For Ocular Infection

Authors
Yue Wu, Aowen Wang, Duanhua Cao, Mingguang He

Presenting
Yue Wu

PURPOSE:
The crisis of antimicrobial resistance (AMR) is a global health issue. Antimicrobial peptides (AMPs) offer a promising alternative; however, existing computational methods often lack adaptability or struggle to navigate the vast chemical space effectively from limited experimental datasets.

METHODS:
We present a novel generative framework designed to rapidly design and optimize potent AMPs through directed latent-space navigation. To address the scarcity of labeled data, we implemented a two-stage training strategy: initial pre-training on massive peptide sequences followed by domain-specific fine-tuning on validated AMP datasets. To achieve sequence evolution, we formulate the optimization task as a diffusion-based bridge problem. It allows the model to learn deterministic trajectories that steer low-activity sequences toward high-potency regions while preserving structural integrity and essential physicochemical properties.

METHODS:

We present a novel generative framework designed to rapidly design and optimize potent AMPs through directed latent-space navigation. To address the scarcity of labeled data, we implemented a two-stage training strategy: initial pre-training on massive peptide sequences followed by domain-specific fine-tuning on validated AMP datasets. To achieve sequence evolution, we formulate the optimization task as a diffusion-based bridge problem. It allows the model to learn deterministic trajectories that steer low-activity sequences toward high-potency regions while preserving structural integrity and essential physicochemical properties.

RESULTS:
Our framework demonstrated optimal predictive and optimization performance. In large-scale benchmarks, it achieved a superior median activity improvement, significantly outperforming existing methods. Experimental validation confirmed the practical efficacy of the framework, with synthesized candidates exhibiting potent activity against Gram-positive and pathogens, including Staphylococcus aureus, with MIC values as low as 4 ug/mL. Notably, in a case study involving de-extinct peptide templates, the model achieved high success rate in potency enhancement. At therapeutic concentrations, candidates exhibited high biosafety with minimal hemolysis and high viability in human corneal epithelial cells (HCECs) and human keratocytes (HKs).

RESULTS:

Our framework demonstrated optimal predictive and optimization performance. In large-scale benchmarks, it achieved a superior median activity improvement, significantly outperforming existing methods. Experimental validation confirmed the practical efficacy of the framework, with synthesized candidates exhibiting potent activity against Gram-positive and pathogens, including Staphylococcus aureus, with MIC values as low as 4 ug/mL. Notably, in a case study involving de-extinct peptide templates, the model achieved high success rate in potency enhancement. At therapeutic concentrations, candidates exhibited high biosafety with minimal hemolysis and high viability in human corneal epithelial cells (HCECs) and human keratocytes (HKs).

CONCLUSIONS:
This framework effectively steers peptide sequence evolution toward high-potency manifolds by combining large-scale generative pre-training with directed latent-flow optimization, providing a computational solution to accelerate the design of next-generation AMPs.