Almost all organisms, from single-cell bacteria to complex species, can detect chemicals through the pairing of chemical ligands and their receptors. Humans and other vertebrates mostly detect chemicals through our senses of smell and taste. Chemicals that we can sense by smell, called odorants, are ligands that bind to odorant receptors in the body, primarily in the nose.
几乎所有的生物，从单细胞细菌到复杂物种，都可以通过化学配体及其受体的配对来检测化学物质。人类和其他脊椎动物大多通过我们的嗅觉和味觉来检测化学物质。我们可以通过嗅觉感知的化学物质，称为气味剂，是与体内气味受体结合的配体，主要是在鼻子里。
Gaurav Ahuja’s lab at the Indraprastha Institute of Information Technology, Delhi, studies this chemodetection — how it influences complex behavioral outputs and the genetics behind these processes. The lab recently released an artificial intelligence–driven prediction tool for olfactory decoding and authored a paper in the Journal of Biological Chemistry detailing the construction of the tool and data behind it.
印度德里Indraprastha信息技术研究所的Gaurav Ahuja实验室研究了这种化学检测——它如何影响复杂的行为输出以及这些过程背后的遗传学。该实验室最近发布了一个用于嗅觉解码的人工智能驱动的预测工具，并在《生物化学杂志》上发表了一篇论文，详细介绍了该工具的构建及其背后的数据。
The researchers named their tool OdoriFy; it is open-source, accessible to other researchers and highly interpretable. An interdisciplinary team of authors spanning computational biology and computer science and several institutes helped on the project. Co-first authors Ria Gupta, a fourth-year undergraduate student who worked on the deep learning behind the model and interpretability, and Aayushi Mittal, a second-year doctoral student who spearheaded data collection and design, share enthusiasm for the tool’s potential uses.
研究人员将他们的工具命名为OdoriFy;它是开源的，可供其他研究人员使用，并且具有高度的可解释性。一个由计算生物学和计算机科学领域的作者组成的跨学科团队以及几个研究所参与了这个项目。共同第一作者里亚·古普塔(Ria Gupta)是一名四年级本科生，她致力于模型和可解释性背后的深度学习，而Aayushi Mittal是一名二年级博士生，负责数据收集和设计，他们对该工具的潜在用途充满热情。OdoriFy’s four modules or prediction engines — Odorant Predictor, Odor Finder, Odorant Receptor Finder, and Odorant–Odorant Receptor Pair Analysis — are available through a user-friendly website. Ahuja and team believe the use of cutting-edge neural network architecture, a series of algorithms that make up the artificial intelligence approach, helps distinguish their tool.
OdoriFy的四个模块或预测引擎-气味预测器，气味查找器，气味受体查找器和气味-气味受体对分析-可通过用户友好的网站获得。Ahuja和他的团队认为，使用尖端的神经网络架构，以及构成人工智能方法的一系列算法，有助于区分他们的工具。The data set behind OdoriFy is one of the largest curated data resources to date. The team manually checked and cross-checked olfactory information of more than 5,000 odorants, 800 nonodorants and 6,000 interaction pairs (between odorant and receptor) — a massive effort to read the scientific literature and document their findings. Mittal said the team “had so many sleepless nights, holding meetings asking ourselves, how can we approach this problem? How can we solve this?”
OdoriFy背后的数据集是迄今为止最大的精选数据资源之一。该团队手动检查和交叉检查了5000多种气味剂、800种无气味剂和6000对相互作用对(气味剂和受体之间)的嗅觉信息——这是一项巨大的努力，以阅读科学文献并记录他们的发现。米塔尔说，团队“度过了许多不眠之夜，召开会议，问自己，我们如何解决这个问题?我们如何解决这个问题?”“There’s a concept in machine learning called garbage in–garbage out — good data in, good data out,” Ahuja said. Without highly accurate input data, their precisely designed computer model wouldn’t be as strong as a predictive tool. As a result, OdoriFy consistently outperforms other models in the olfaction field and scores high across a number of validated metrics for measuring accuracy in prediction.
Ahuja说:“机器学习中有一个概念叫做垃圾输入，垃圾输出——好的数据输入，好的数据输出。”如果没有高度精确的输入数据，他们精确设计的计算机模型就不会像预测工具那样强大。因此，OdoriFy始终优于嗅觉领域的其他模型，并在许多经过验证的预测准确性度量指标中得分很高。Scientists understand that humans’ ability to distinguish odors is combinatorial. “Nature has developed ways to deal with the fact that we’re exposed to billions of chemicals, but we have only a limited genome and therefore a limited number of odorant receptors,” Ahuja said. “One receptor can recognize more than one odorant, and one odorant can be recognized by more than one receptor.” So, while humans have only about 400 functional genes for odorant receptors, the combinatorial effect gives us the ability to detect many more than 400 odorants.
科学家们明白，人类区分气味的能力是组合的。阿胡贾说:“大自然已经开发出方法来处理我们接触数十亿种化学物质的事实，但我们只有有限的基因组，因此气味受体的数量有限。”“一个受体可以识别不止一种气味，一种气味也可以被不止一个受体识别。”因此，虽然人类只有大约400种气味受体的功能基因，但组合效应使我们能够探测到400多种气味。A tool such as OdoriFy that can predict both odorants and odorant-receptor pairing can help open doors for researchers working across this field of chemodetection and novel applications. Ahuja and team already have been contacted by companies and researchers who have used the tool and are interested in further collaboration. One of the most interesting avenues of pursuit is the application to cancer, as human tumor cells are known to express certain odorant receptors.
像OdoriFy这样可以预测气味剂和气味-受体配对的工具可以帮助研究人员在化学检测和新应用领域打开大门。已经有公司和研究人员联系了Ahuja和他的团队，这些公司和研究人员使用了这个工具，并对进一步的合作感兴趣。最有趣的研究方向之一是将其应用于癌症，因为已知人类肿瘤细胞表达某些气味受体。“Working on this project made us all realize how important olfaction is and how important our tool is for the public,” Gupta said.
古普塔说:“这个项目让我们都意识到嗅觉的重要性，以及我们的工具对公众的重要性。”
Leia Dwyer is a Boston-area biotech and pharmaceutical industry professional.
莱娅·德怀尔是波士顿地区生物技术和制药行业的专业人士。
原文链接：Researchers make sense of scents (asbmb.org)
