Scholarlink

Context <p>Insects navigate the world through odourant receptor (OR)–mediated chemo-sensation. While of primary importance, insect chemo-sensation is poorly understood due to the lack of experimental structures surveying the binding between ORs and their ligands. Computational approaches empowered by deep learning can, in principle, provide an accurate prediction of OR structures and an understanding of the key determinants shaping OR–ligand interaction. However, a comprehensive survey of these models’ capabilities is missing.</p> Methods <p>Here, we evaluate the ability of different models to predict OR structures and their binding. We found that while AlphaFold2 and Boltz2 can accurately predict the structure and pose of insect ORs, existing deep learning models struggle to accurately predict ligand binding affinity and differentiate between binders and nonbinders: with high prediction variability even between very similar targets. To increase accuracy and ameliorate the identified limitations, we tested a weighted combination of models, demonstrating a meta-model with improved accuracy. We show that, when applied to the prediction of naphthalene binding, this approach can efficiently predict binding affinities to a series of highly responsive ORs, with empirical validation of the computational predictions. These results not only demonstrate the opportunity for meta-models to increase accuracy of insect OR prediction but provide insights on the generalised prediction of membrane protein complexes and ligands.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From data to scent: validating an ensemble AI model that predicts novel insect odourant interactions

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Leroy Bird,
  • Brady Owen,
  • Alaigne Mare,
  • WenJuan Huo,
  • Sarah Wilson,
  • Andrew Kralicek,
  • Jonathan Good,
  • Colm Carraher

摘要

Context

Insects navigate the world through odourant receptor (OR)–mediated chemo-sensation. While of primary importance, insect chemo-sensation is poorly understood due to the lack of experimental structures surveying the binding between ORs and their ligands. Computational approaches empowered by deep learning can, in principle, provide an accurate prediction of OR structures and an understanding of the key determinants shaping OR–ligand interaction. However, a comprehensive survey of these models’ capabilities is missing.

Methods

Here, we evaluate the ability of different models to predict OR structures and their binding. We found that while AlphaFold2 and Boltz2 can accurately predict the structure and pose of insect ORs, existing deep learning models struggle to accurately predict ligand binding affinity and differentiate between binders and nonbinders: with high prediction variability even between very similar targets. To increase accuracy and ameliorate the identified limitations, we tested a weighted combination of models, demonstrating a meta-model with improved accuracy. We show that, when applied to the prediction of naphthalene binding, this approach can efficiently predict binding affinities to a series of highly responsive ORs, with empirical validation of the computational predictions. These results not only demonstrate the opportunity for meta-models to increase accuracy of insect OR prediction but provide insights on the generalised prediction of membrane protein complexes and ligands.