Why I Started Writing About AI and Biology
A short note on why I'm starting to publish technical writing about foundation models, self-supervised learning and computational biology — after several years of heads-down research.
In short
- I've spent the last few years building ML systems for biology without writing anything public about the process.
- Starting this series to share the parts of the work that don't fit in a paper: failed experiments, design choices, framing.
- Posts will be AI-for-biology only, written for working researchers and engineers — not science journalism.
The usual reason, said honestly
The honest reason is that I've been accumulating notes. After a PhD in physics, a master's in synthetic biology at the CSIC, a few years teaching CS and now a PhD in AI, I have a folder of observations about how machine learning actually behaves when you hand it molecular data instead of ImageNet — and most of those observations never make it into a paper.
Some of them are embarrassingly small. Things like: don't trust the validation MSE on a token-level chemistry model until you've checked what it predicts for octanol. Some of them are less obvious: that the gap between self-supervised pretraining on SMILES and a useful molecular property predictor is much wider than the leaderboards suggest, because the leaderboards reward settings that don't transfer cleanly to a wet-lab pipeline.
None of those by themselves is a publication. Together, they're the part of the job I actually want to read about other people's work, and rarely do.
What this series is, and what it isn't
This series is for working researchers and ML engineers who have to actually deploy models, not just benchmark them. The default length will be short — between five and fifteen minutes of reading — and the default tone will be direct. I'll try to make the posts standalone enough that you can read them in any order, with a short list of references at the bottom of each one.
It is not a survey of the field. I've read Cheng et al.'s review of molecular representation learning probably a dozen times and I still can't quote it from memory. The literature is deep and well-reviewed elsewhere; the gap I want to fill is the pragmatic one.
Fair warning. My first language is Spanish, and grammar is not my strongest skill in English either. If you spot something sloppy in one of these posts, I'd rather you tell me than let it linger. The posts will get better over the year.
Where to start
If you're new here, the most useful entry points are probably:
- Predicting Collision Cross-Sections: What I Learned From 61,863 Molecules — the closest thing to a proper technical writeup of work I co-authored; goes deep on data, baselines and the gotchas of CCS prediction.
- Self-Supervised Transfer Learning for Collision Cross-Section Prediction — same domain, but showing why pretraining on unlabeled molecules buys you cross-dataset generalization.
- Evaluating Chemical Foundation Models for Metabolite Property Prediction — a research note on the trap that any future foundation-model-for-chemistry paper needs to address.
A note on cadence
I'll post about once every two or three months, not weekly. Each post will be something I'd genuinely want to read, written mostly on weekends and evenings. If I post more often, something has gone wrong and I should be in the lab instead.
If any of this is useful to your own work, leave a comment on the relevant paper, or drop me a line through the contact page. I'd love to hear what you're building.