01⌁
Computational olfaction
I train models on human odor ratings and test them on molecules the models have not seen before.
- Principal odor maps
- Mixture perception
- Odor intensity
Neuroscientist who writes a lot of code
I study
smell.
I run experiments and build models to study why things smell the way they do. Most of the code I write is for checking one against the other.
Molecule
Model
Percept
01 Molecule → model → odor ratings
Currently
Head of NeuroscienceOsmo
Research ProfessorArizona State University
Based inPhoenix, Arizona
01 / What I do
Predicting smell is harder than predicting color.
Wavelength tells you a lot about color. There is no equivalent measurement for smell. A molecule can smell different at another concentration, and two people may describe it differently. I collect those judgments and train models on them.
02◇
Models that can be tested
I write tests that run a model on new data and score the result.
- Model evaluation
- Human-in-the-loop workflows
- Research infrastructure
03∿
Scientific software
I maintain the datasets and testing software used in several of these projects.
- Pyrfume
- SciUnit & NeuronUnit
- Neuroinformatics
02 / The path here
How I got into
olfaction.
My PhD work was experimental. During my postdoc I started writing more modeling and data-analysis code. At ASU, human olfaction became the main subject.
2002—08
University of Pittsburgh
PhD in Neurobiology. I studied synaptic plasticity and activity in small networks of hippocampal neurons.
2008—13
Carnegie Mellon
As a postdoc, I recorded from the olfactory bulb in vivo and helped develop NeuroElectro.
2013—
Arizona State University
I built tools for testing neuroscience models and worked on how humans and rodents perceive odors.
2022—
Google Research → Osmo
I joined the Principal Odor Map project at Google Research and now lead neuroscience at Osmo.
03 / Selected research
What happens when
odors are mixed?
Human-levelprospective odor description
The Principal Odor Map
We asked a trained panel to describe 400 new molecules, then gave the same job to the model. It performed about as well as the median panelist. I focused on the experimental design and on making sure the predictions and measurements could be compared fairly.
2026Odors smell like their componentsMost mixtures in this study were well described by averaging the smells of their components.Mixtures
2025A quantitative framework for odor intensityA model of how strong single molecules and mixtures smell.Intensity
2023A principal odor map unifies diverse tasksA graph neural network predicts odor from molecular structure and works on molecules it has never seen.Odor map
2017Predicting human olfactory perceptionWe gave teams the same odor data and compared how well their models predicted held-out human ratings.DREAM
View more publications
04 / Software
Software I maintain.
Much of this software exists because I needed to know where a number came from and whether it had changed.
01PyrfumeData commonsA collection of published odor datasets, plus Python code for loading them.↗
02SciUnitFrameworkA framework that runs a model prediction and scores it against an experimental result.↗
03NeuronUnitNeuroinformaticsThe SciUnit idea applied to neurons, ion channels, and electrophysiology measurements.↗
04DREAM OlfactionCommunity challengeCode and data from the 2015 odor prediction challenge.↗
05 / Collaborations
I also work in
large groups.
The DREAM challenge gave the same prediction problem to teams around the world. In 2020, the Global Consortium for Chemosensory Research ran a survey in 32 languages and found that recent smell loss was a strong predictor of COVID-19.
Get in touch
If your project involves smell, there’s a good chance I’ll want to hear about it.