Sagace
Models for living systems
It names the ability to read a situation correctly and choose the right action inside it. Intelligence is becoming abundant, but discernment can become scarce. We want to bring machine learning into scientific organizations that make consequential decisions currently without its lights.
Worked on experiment design for protein engineering and lead optimisation. Worked on hepatotoxicity and single-cell biology. Fascinated by emergence, in living systems and artificial ones, and the ways each explains the other.
Works on deep-learning interpretability and on the scalability and efficiency of deep learning models. Has a strong background in secure and efficient infrastructure deployment.
We have run projects outside of school for as long as we have been in it. We are two today, and we only plan to work with people who want to learn the field of AI for Science from inside real problems. We are curious minds, passionate about science and solving problems.
We want to be dynamic: we work for six weeks on a given project and organize ourselves a hackathon in the middle. We value regular checkpoints and autonomy between them.
A confidential deliverable, an open-source contribution wherever the work would allow one, a case study, and a proposal for what to do next.
We price so that a small team can afford us. A clearly scoped deliverable, defined by your needs, with you covering our computational and infrastructure costs.
We would never accept work we already know inconclusive, even if profitable, or engagements we cannot walk away from. When a client is wrong, we say so: their budget has a better use.
These are three systems we have already built and that can run in your browser. We have tried to package our previous work on what we were curious about. They are not yet production-grade, but they cannot wait to be.
Per-pathway predictors, one for CYP inhibition and metabolism, transporters and nuclear receptors, composed into a liver-injury call you can point a mechanism at, with a live screen against 193 marketed drugs.
Predicted HepG2 single-cell response read as pathway-level stress, combined with exposure to give a triage decision, plus an applicability domain that says when to abstain.
A walkthrough of one of our most recent pieces of work. We detail the datasets, methods and results behind the active-learning demo. A preprint on the same work is in preparation.
AI should be cooperative
Systems that keep a human inside the loop are harder to design than systems that remove them, yet they are achievable. Capability has outrun application: the distance between the frontier labs and the people who could use these tools is the most wasteful gap in the field, and it is where we want to work.
We are taking on a small number of focused six-week projects. Bring us the problem that sits just below the line of your roadmap.