# Deep learning to design immune-evading viral vectors for gene therapy

> **NIH NIH F32** · HARVARD MEDICAL SCHOOL · 2021 · $66,390

## Abstract

Project Summary
The delivery of gene therapies (DNA or RNA) using viral vectors is a promising therapeutic avenue
for several indications, including genetic conditions, cancer, and neurodegenerative disease.
Adeno-associated viral (AAV) vectors in particular have demonstrated efficacy and safety in
several clinical trials, and have been approved for the treatment of two monogenic diseases by
the FDA. However, a major bottleneck in expanding the reach of AAV-based therapies is
optimizing vectors to avoid immune detection while preserving their function and tropism. AAV
vector applications are significantly limited by immune targeting of viral capsids within a patient's
body, resulting in accelerated clearance, loss of efficacy and dangerous hypersensitivity
responses. As AAV is widely prevalent in the human population, these effects are often the result
of immunological memory and thus more severe than a naïve response. Adverse effects are
difficult to predict through pre-clinical experiments, increasing the risk posed to clinical trial
patients. While immune recognition epitopes can sometimes be mapped to specific viral capsid
residues, even high-throughput library-based strategies to ablate these epitopes are plagued by
the creation of non-functional variants. To address this challenge, we propose to use a generative
statistical model trained on natural sequence variation to design `smart' immune-evading AAV
capsid libraries enriched in functional variants. This approach will integrate statistical models of
the functional constraints on protein sequence with state-of-the-art predictors of T-cell and
humoral immunogenicity to produce diverse libraries of therapeutically useful vector sequences.
Data gathered from initial designed libraries will inform model optimization to iterate towards
further improved viral vectors. Deimmunized, diverse vector libraries will address immune evasion
at the initial stages of gene therapy development, accelerating progress towards safe and
effective therapies. An algorithm that builds on observed sequences to generate diversity subject
to immunogenicity constraints has broad implications for the development of all protein-based
biotherapeutics, such as antibodies. This approach combines the advantages of protein therapies
optimized through natural selection with precise control over desirable characteristics for human
health applications.

## Key facts

- **NIH application ID:** 10313084
- **Project number:** 1F32GM141007-01A1
- **Recipient organization:** HARVARD MEDICAL SCHOOL
- **Principal Investigator:** Nicole Thadani
- **Activity code:** F32 (R01, R21, SBIR, etc.)
- **Funding institute:** NIH
- **Fiscal year:** 2021
- **Award amount:** $66,390
- **Award type:** 1
- **Project period:** 2022-03-01 → 2023-05-24

## Primary source

NIH RePORTER: https://reporter.nih.gov/project-details/10313084

## Citation

> US National Institutes of Health, RePORTER application 10313084, Deep learning to design immune-evading viral vectors for gene therapy (1F32GM141007-01A1). Retrieved via AI Analytics 2026-05-25 from https://api.ai-analytics.org/grant/nih/10313084. Licensed CC0.

---

*[NIH grants dataset](/datasets/nih-grants) · CC0 1.0*