Abstract: FR-PO0773
Machine Learning-Guided Refinement of Antibodies Targeting Podocyte Integrin α3 for Development as Novel Therapeutics for FSGS
Session Information
- Glomerular Diseases: Cell Homeostasis and Novel Injury Mechanisms
November 07, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Glomerular Diseases
- 1401 Glomerular Diseases: Mechanisms, including Podocyte Biology
Authors
- Balza Pineda, Santiago, Allosite Therapeutics, Miami, Florida, United States
- Lopez-Rodriguez, Darlah M., Allosite Therapeutics, Miami, Florida, United States
- Gonzalez-Lerma, Paola, Allosite Therapeutics, Miami, Florida, United States
- Youssef, Mohamed A., The University of Texas Medical Branch at Galveston, Galveston, Texas, United States
- Gupta, Vineet, The University of Texas Medical Branch at Galveston, Galveston, Texas, United States
Background
Integrin α3β1 is essential for podocyte adhesion to the glomerular basement membrane. Loss of α3β1 function leads to podocyte detachment and proteinuria, contributing to focal segmental glomerulosclerosis (FSGS). We recently identified novel, conformation-specific antibodies targeting integrin α3β1 that increase podocyte adhesion, reduce cell loss, and provide protection from kidney injury in FSGS model systems. Here, we utilized state-of-the-art machine-learning (ML) tools to further their affinity, cross-reactivity, and developability for therapeutic development
Methods
Antibody CDRs of the identified α3β1-activating antibodies were grafted onto the trastuzumab scaffold. Framework positions were optimized using an ML model trained on Ab sequences and HCDR3 was targeted for mutagenesis using an NNK-based combinatorial library. Four rounds of phage display selection against recombinant human α3β1 were followed by next-generation sequencing (NGS) of enriched eluted pools. A multi-step filtering strategy considering sequence liabilities, humanness, and ML model likelihoods was applied to score 11 candidates. These were tested via ELISA, flow cytometry, and kinetic BLI assays. Functional assays included scratch wound healing, ligand binding potentiation, and kidney tissue staining. A composite score incorporating KD, species cross-reactivity, and cell binding was used to train an ML model, then evaluated by leave-one-out cross-validation (LOOCV)
Results
Zero-shot grafted antibodies retained α3 binding and exhibited improved expression yields. Of 11 variants, 9 selectively bound human α3 over other integrins; four with improved human α3 affinity, and one which preserved high-affinity to mouse. Functional assays confirmed rescue of podocyte adhesion and potentiation of ligand binding. ML training captured sequence-function relationships with ~25% LOOCV accuracy and was used to generate new prospective variants with further improved characteristics. Six of these are being prioritized for further validation in vivo
Conclusion
This ML-guided pipeline enabled rapid, multi-parameter optimization of α3 antibodies beyond classical screening. The approach accelerates therapeutic maturation and may be extended to integrins implicated in other kidney diseases, such as Alport syndrome