About this role
Summary: Lead the development of foundation models for molecular structure prediction using mass spectrometry data. Set modeling direction, design architecture experiments, and collaborate with computational chemists to build evaluation pipelines.
Overview
Novogaia is an applied AI drug discovery company. We build machine learning systems that decode the chemistry of natural organisms, starting with fungi, to find the next generation of medicines.
We are a small team of AI engineers, computational biologists and chemists building foundation models for molecular structure prediction from mass spectrometry data. We are seeking a machine learning research engineer to lead the development of these models: setting modeling direction, designing and running architecture experiments, and owning and implementing innovative architectural improvements. You will also work with our computational chemists to design our internal benchmark and evaluation pipelines.
The Role
In your first year, you'll take the lead on Gaia-02 model development, building on the work done to train [Gaia-01](https://www.novogaia.bio/gaia-01)while moving us towards an innovative, best-in-class architecture.
- Lead modeling decisions for Novogaia's core tasks: de novo molecular structure generation, molecular and spectral retrieval, mass spectrum simulation, molecular formula prediction, and compound property/class prediction
- Serve as the technical authority for Novogaia’s machine learning program: set modeling and experimentation standards through your own best practices, and lead architecture reviews and experiment design across the team, including:
- Making and experimenting with architecture decisions for our spectrum encoders and structure decoders, justifying each choice with experimental evidence
- Analyzing model behavior and failure cases, and explaining findings to researchers and non-technical stakeholders alike
- Working with engineers to turn training and evaluation code into reproducible, scalable pipelines
- Develop new ways to represent mass spectra and molecular structures, such as how peaks, relationships between peaks, and molecular graphs are encoded as model inputs, and design our internal pipelines to benchmark model performance
- Collaborate with our computational chemists to build a trustworthy benchmark for structure prediction models, including hard negatives and leakage-safe splits
- Follow current research in machine learning and computational mass spectrometry, and propose and prototype relevant approaches from other fields
- Translate lessons from experimentation into research priorities, data requirements, and R&D direction
What We Require
- Research or applied experience in machine learning, with direct experience owning foundation model development and generating new modeling ideas
- Depth in model architecture: you have designed or modified architectures such as transformers or graph neural networks, and have improved model performance through changes to the architecture, the training objective, or the data representation rather than hyperparameter tuning alone
- Exceptional technical communication skills, including the ability to explain evaluation findings clearly to both researchers and non-technical stakeholders
- Ability to analyze model behavior and interpret computational results critically
- Strong proficiency in Python and experience with reproducible, containerized pipelines
- Familiarity with cheminformatics representations (SMILES, InChIKey, molecular fingerprints), or willingness to pick these up quickly
- Familiarity with computational mass spectrometry or eagerness to learn
- Ability to work in-person in London
What We Value
- Strong scientific judgment and ability to navigate and operate novel research directions
- Motivation to build infrastructure other people can confidently rely and build upon
- Comfortable being the person who tells the team a result doesn't hold up
- Curiosity, low ego, and a willingness to learn the chemistry side quickly