High-fitness AAV Mutant Proposal

A skill performs automated multi-round optimization of a 28-amino acid segment of the VP1 capsid protein of Adeno-Associated Virus (AAV) to discover mutants with improved DNA packaging fitness and high sequence diversity.

When to Use This Skill

Design novel AAV mutants with improved DNA packaging fitness.
Run computational iterative directed evolution.
Perform fast mutation search guided by an oracle model.

Example prompts:

“Design AAV mutants with higher DNA packaging fitness.”
“Run multi-round mutation optimization for AAV.”
“Generate 96 AAV variants with improved fitness.”

Prerequisites

Python 3.9+
PyTorch
NumPy / Pandas
Protein sequence analysis tools
Protein language model tools (ESM2)

Core Capabilities

This skill can:

Download initial AAV sequences if they were not provided by users.
Download and execute an in-silico oracle AAV prediction model.
Generate controllable mutants within 4 point mutations for each round.
Use ESM2 embeddings to represent protein sequences.
Optimize mutation proposals based on oracle feedback.
Maintain population diversity using average pairwise Hamming distance.
Perform multi-round optimization and return the best mutants.

Workflow

Download initial AAV sequences from https://cloud.tsinghua.edu.cn/f/992109032d8049689a6d/?dl=1 and use them as the starting pool.
Download the oracle AAV prediction model from https://cloud.tsinghua.edu.cn/f/80bbc575ec3f4e63a0af/?dl=1, and the configuration file from https://cloud.tsinghua.edu.cn/f/09ea0869b74b4d2ca53e/?dl=1.
Execute code for oracle loading and scoring:

import torch
from omegaconf import OmegaConf

# ===== ORACLE MODEL LOADING =====
def load_oracle_model(ckpt_path, cfg_path):
    with open(cfg_path, 'r') as fp:
        cfg = OmegaConf.load(fp.name)
    oracle = BaseCNN(**cfg.model.predictor)
    state_dict = torch.load(ckpt_path)
    oracle.load_state_dict(torch.load(ckpt_path))
    oracle.eval()

# ===== ORACLE SCORING FUNCTION =====
def score_sequence(oracle, sequence: str) -> float:
    results = oracle(sequence).detach()
    return results.cpu().numpy()

Compute ESM2 embeddings for all sequences to represent sequence features.
Proposal: for each round, propose 96 × 4 candidate mutants from the current population using only point mutations with ≤4 mutations per sequence.
Evaluation: evaluate all candidate sequences using the oracle scoring function. Use oracle feedback from previous rounds to bias mutation proposals toward directions that increase predicted fitness (fitness gradient exploitation).
Selection: rank sequences by predicted fitness and select the top 96 mutants, while maintaining diversity measured by average pairwise Hamming distance.
Repeat proposal, evaluation, and selection until 10 rounds are completed, or best fitness does not improve for 3 consecutive rounds.
Collect the best 96 mutants discovered across all rounds and sort them by predicted DNA packaging fitness, and export the results as a CSV file following the specified output format.

Output Format

The final result must be a CSV file with two columns:

sequence	fitness
AAV_mutant_sequence	predicted_fitness

Requirements:

Exactly 96 sequences
Sorted by fitness in descending order
Sequences must be valid AAV mutants

Example:

sequence,fitness
ADMEIIQVNPYSSEQYGDVATPLYHGTG,0.96
ADMEIRQVNPYSSEQYGDVATPLQHGTG,0.93
ADSELASTNPVSTELYGIVATNLMAQAS,0.92
...