Molecular foundation models
General-purpose representations of molecules and their measurements that transfer across chemistries and downstream tasks.
Molecular machine learning
Tomo Oga
I build machine learning models for molecules: foundation models that learn chemical structure, generative models that propose it de novo, and evaluations that check whether either generalizes. I'm currently a researcher at Matterworks.
01
Research
I'm a third-year Data Science and Mathematics student at Northeastern University and a machine learning researcher at Matterworks in Somerville, MA, where I work on foundation models for molecules.
Most of chemical space is unexplored, and most of the molecules that matter in biology have never been characterized. I'm interested in models that learn chemistry well enough to reason about molecules outside their training data, and in evaluation that tells us honestly when they can.
Structure from data
Proposing molecular structures de novo from experimental observations, beyond what reference libraries already contain.
Benchmarks that hold up
Structure-disjoint splits that still measure out-of-distribution generalization as molecular datasets grow by an order of magnitude.
02
Publications
NeurIPS 2026 · Workshop on AI Data Readiness for Scientific Discovery
Critical Mass: Structure-Disjoint Splits Drift With the Libraries They Are Built On
Molecular ML measures out-of-distribution performance with structure-disjoint splits: threshold a pairwise distance, then keep connected components together. These rules are fixed at the scale of the library they were designed on. Taking spectrum-to-structure prediction as a case study, we make the exact MCES-10 split computable on 227,690 molecules, nearly 8× the library it was set on. At this scale the similarity graph percolates: the holdout nearly vanishes, and what remains is systematically heavier. Naive large-library splits therefore conflate structural novelty with extrapolation to larger chemistry.
2,259×
fewer integer programs than all-pairs MCES, with a provably identical partition
95.5%
of a 227,690-molecule library falls in a single connected component
0.881
probability that a held-out compound outweighs a training compound
max achievable holdout under MCES-10 → 4.5%
03
Experience
-
Machine Learning Researcher · Matterworks
Somerville, MA
Building foundation models for molecules, trained on mass spectrometry data. My current work is on de novo structure elucidation and on how molecular models should be evaluated at scale.
-
Machine Learning Researcher · Gyori Lab for Computational Biomedicine
Northeastern University, Boston, MA
- Integrated clinical trial data from multiple registries into a computable knowledge graph for TrialSynth, grounding biomedical entities to standard identifiers.
- Proposed a MeSH prefix-tree restriction for named entity recognition, raising entity precision by about 17%.
- Explored LLM-based extraction of genomic variants from trial eligibility criteria for precision oncology.
-
B.S. Data Science and Mathematics · Northeastern University
Khoury College of Computer Sciences
Earlier projects
- ConvLSTM for MLX Convolutional LSTM (Shi et al.) implemented for Apple's MLX array framework. 2024
- Video frame interpolation Reproduction of residual ConvLSTM frame interpolation (Suzuki et al.) in PyTorch. 2023
04
Contact
Interested in molecular ML, generative chemistry, or evaluation? I'm always glad to talk.
oga.t@northeastern.edu
github.com/tomo-oga
linkedin.com/in/tomo-oga