Notes on artificial intelligence, foundation models, computational biology and the future of scientific discovery. Updated when I have something worth saying, not on a schedule.
Series: AI for Biology (since Oct 2024)
A retrospective on a multi-agent drug-discovery system: what worked, what didn't, why the orchestrator is most of the engineering work, and why autonomous-agent-scientist is a worse idea than it sounds.
A research note on the protocol of a paper sent to Briefings in Bioinformatics: five chemical foundation models, five downstream tasks, the same head — and what surprised us when we ran the comparison.
A working researcher's retrospective on ESM-2, ESMFold, AlphaFold 3, ESM-3 and Boltz: which ones shipped into our pipeline, which didn't, and the calibration problem nobody talks about.
A write-up of the IWANN 2025 paper. Why pretraining on unlabeled SMILES buys you cross-dataset CCS generalization — and why it doesn't, by itself, fix the cross-instrument problem.
A technical deep-dive into deep learning for CCS prediction on METLIN-CCS. Architectures, baselines, evaluation traps, and what actually generalizes across instruments.
The kickoff post for this series. Why a PhD candidate in AI who mostly works on molecular property prediction decided to spend some of his evenings writing things down for public reading.
Most of these posts are side effects of my main work — see the publications page for the peer-reviewed version, or the projects page for engineering artifacts. The writing is where the trade-offs, dead-ends and design reasoning live, before (or without ever) becoming a paper.
If you'd like to be notified when a new post goes up, the most reliable way is to drop me a message via the contact page; I don't run a separate newsletter infrastructure.