Google Uses Deep Learning to Predict the Olfactory Properties of Molecules

2019-10-25

Smell is a common sense shared by animals and plays a crucial role in how organisms respond to the external world. For humans, our sense of smell is closely related to our ability to enjoy food. Smell allows us to sense a variety of odors in our daily lives, and although it is important, smell has not received as much attention from machine learning researchers as vision and hearing.


Human olfaction is achieved by activating 400 different types of olfactory receptors (ORs), olfactory sensory neurons (OSNs), located in a small piece of tissue called the olfactory epithelium. These OSNs send signals to the olfactory bulb and then to other tissues in the brain. Based on precedents from visual and auditory deep learning, it should be possible to directly predict the final sensory outcome of input molecules even without knowing all the complex details involved in the system. Solving the odor prediction problem could help discover new synthetic odorants, reducing the ecological impact of using natural plants. Examining the generated olfactory models may even lead to new discoveries in olfactory biology.Small odor molecules are the basic components of fragrances, making them the simplest form of odor prediction problem. However, each molecule can have multiple odor labels. For example, vanillin has labels such as sweet, vanilla, creamy, and chocolate, with some components being more prominent than others. Therefore, odor prediction is also a multi-label classification problem.


Google researchers have utilized graph neural networks (GNNs), which are deep neural networks designed to operate on graphs. As input, they directly predict odor labels for individual molecules without specifying any rules. They have demonstrated that this method significantly improves odor prediction performance compared to the state-of-the-art techniques and is a promising direction for future research.



Graph Neural Networks for Odor Prediction



Each node is represented as a vector, with each entry in the vector initially encoding some atomic-level information.



For each node, we look at neighboring nodes and collect their information, which is then transformed into new information for the central node using a neural network. This process is performed iteratively. Other variants of GNN utilize edge and graph-level information.



The odor prediction graph of GNN. We transform the structure of a molecule into a graph and input it into GNN layers to learn better representations of nodes. These nodes are simplified into single vectors and passed to a neural network for predicting multiple odor descriptors.


Since molecules resemble graphs, with atoms forming vertices and bonds forming edges, GNNs are a natural choice for understanding them. But how do we convert the structure of a molecule into a graph representation? Initially, each node in the graph is represented as a vector using any preferred characterization (atomic identity, atomic charge, etc.). Then, in a series of message-passing steps, each node propagates its current vector value to each neighboring unit. The update function then collects the set of vectors sent to it and generates an updated vector value. This process can be repeated multiple times until all nodes in the graph are summarized into a single vector through summation or averaging. This single vector representing the entire molecule can then be passed to a fully connected network to learn molecular features.This representation is unaware of the spatial positions of atoms, so it cannot distinguish between stereoisomers, which are molecules composed of the same atoms but with slightly different molecular structures, such as (R)- and (S)-carvone. However, we found that even without distinguishing stereoisomers, odor can still be predicted effectively. In odor prediction, GNN consistently outperforms previous state-of-the-art techniques.


Performance example of GNN on odor descriptors, measured by AUROC score. Close to 1.0 indicates better performance. In most cases, GNN's performance significantly surpasses the domain-standard baselines and is comparable to other metrics such as AUPRC, recall, and precision.



Learning from the Model and Extending to Other Tasks


In addition to predicting odor labels, GNN can also be applied to other olfactory tasks. For example, classifying odor labels with limited data. For each molecule, a learned representation can be extracted from the intermediate layers of the model, which are optimized for odor labels, called "odor embeddings." These can be seen as olfactory versions of color spaces, such as RGB or CMYK. To understand if these odor embeddings are useful for predicting related but different tasks, researchers designed experiments to test the embeddings on related tasks. The performance of odor embeddings was then compared to common chemoinformatics representations, which encode structural information of molecules but are odor-agnostic, and it was found that odor embeddings are generally applicable to some challenging new tasks and even meet the requirements of state-of-the-art techniques in certain aspects.


2D snapshot of the embedding space, highlighting some example odors. Left: Each odor clusters in its own space. Right: Hierarchical nature of odor descriptors. Shadow and contour regions are computed using kernel density estimation with embeddings.



Future Work


In the field of machine learning, smell remains one of the most elusive senses. The possibilities for future research are diverse, ranging from designing new odor molecules that are cheaper and more sustainable to digitizing smells, and even enabling people without a sense of smell to experience odors. We hope to raise more awareness of this problem in the machine learning community by ultimately creating and sharing high-quality open datasets.

RECENT AI NEWS

RECENT AI TOOLS