Three papers from the last fortnight, in Nature, Cell and Genome Biology. Bioinformatics gets a lot of tool announcements and very few of them change anyone's day. These three do.
One turns the world's single-cell datasets into something you can actually search. One is a model that guesses how a cell will react to being poked, including in cell types it never saw poked. One is the unglamorous kind of paper that quietly tells you the tool you have been using for years is the wrong one for your data.
1. The single-cell atlases are finally searchable
Here is a problem that sounds like it should already be solved. You have a sequence in hand: a mutation, an odd splice junction, a stretch of a virus. You want to know whether it shows up anywhere in the enormous pile of single-cell data the field has generated. Which cells, in which tissue, in which disease.
You cannot ask that question. Standard pipelines line reads up against a reference genome, count how many landed on each gene, and throw the raw sequence away. What gets deposited and shared is the count table. If you want the sequences back you are downloading and reprocessing millions of very large files, and the pile grows by petabytes a year.
Malva indexes the raw sequence space instead of the counts. It is reference-free, meaning it does not need a genome to compare against, and species-agnostic, meaning it does not care whose genome it would have been. The index currently covers around 74 million cells from thousands of experiments across health and disease, and you query it directly for a sequence, a mutation, a splice junction, a pathogen, or where a given transcript sat in space.
The consequence the authors lean on is that identity falls out of sequence composition alone. You can pick out cell types, and work out how similar two cells are, without ever mapping anything to a reference. They also show it wired up to neural networks for larger automated searches.
What to watch: a search engine only knows what is in its index. Seventy-four million cells is a lot and it is not everything, so a sequence that returns nothing means it is absent from Malva, not absent from biology. And this is discovery, not quantification. It tells you where to look; you still go and do the careful measurement.
2. Predicting how a cell reacts, in cell types you never tested
A perturbation experiment means doing something to a cell and reading what changed: knock out a gene, add a drug, hit a signalling pathway, then sequence. The dream is to stop running most of them. If a model can tell you which of ten thousand possible perturbations are worth the bench time, the screen gets small enough to actually do.
Models have been trying this for a while and keep failing in the same way. They learn a cell type, do well on it, and fall apart on the next one. Since the whole point is to predict what you have not tested, that failure is the failure.
State is built around that problem. Rather than predicting what happens to a single average cell, it predicts across sets of cells, so the variation inside a population and between experiments is part of what it models rather than noise it has to survive. Its underlying cell embeddings were trained on observational data from 167 million cells, which is data about cells sitting there being themselves, no perturbation involved.
The numbers they report:
- More than 30% better at telling perturbation effects apart on large datasets, compared with the baselines.
- Better accuracy at naming which genes actually changed, across genetic, signalling and chemical perturbations.
- It can flag strong perturbations in cellular contexts where it saw no perturbations at all during training. That is the generalisation everything else has been missing.
They also released Cell-Eval, a framework for scoring models like this, which matters more than it sounds. A field full of models and no agreed test is a field where every paper wins.
What to watch: 30% better than a baseline is not the same as right. Flagging that a perturbation will be strong in an untested cell type is a narrower claim than predicting the response in detail, and the authors frame the work as setting up the scaling problem rather than solving it.
3. Your genome annotation tool is probably the wrong one
Annotation is the step where a bacterial genome stops being a string of letters and starts being a list of genes with names and probable jobs. Everything downstream inherits it. Get it wrong and the error propagates silently through every comparison you make afterwards.
Four tools do most of this work: Prokka, Bakta, EggNOG-mapper and PGAP. People pick between them by habit, by what their old lab used, or by what the last paper they read used. Until now there was no systematic comparison to pick differently.
This group ran all four across 156,033 genomes. Not just clean ones: archaea, thousands of bacterial species, deliberately frameshifted genomes, and metagenome-assembled genomes, which are the fragmented, sometimes contaminated assemblies you get from sequencing a whole community rather than a cultured isolate.
The answer is not one tool. It splits:
- Bakta for high-quality bacterial genomes.
- PGAP for archaea, and for the awkward cases: metagenome-assembled, fragmented or contaminated.
- For Gene Ontology terms, the functional labels, PGAP covers a broader range of terms while EggNOG-mapper assigns more terms per feature.
So the right question is not which tool is best, it is which tool suits the state your genomes are in. If you are working with MAGs and reaching for the tool you learned on clean isolates, this paper is telling you something.
What to watch:for most of these genomes there is no gold-standard annotation to check against, so "excels" here is comparative. It tells you which tool to prefer over the others, not how close any of them are to the truth.
The thread
Two of these make data usable that was already sitting there. Malva does not generate a single new read; it makes the reads the field already deposited answerable. The annotation benchmark does not build a tool at all, it just runs the existing four at a scale nobody had bothered with. A surprising fraction of progress in bioinformatics looks like this.
The third points somewhere less settled. Worth reading alongside State is a comment piece that ran in Nature Methods the same week, asking how you would even test a biomedical foundation model: whether these things can be refuted, or verified, or only judged on whether they turn out to be useful. It is not a research paper and it reports no result. It is the question the whole category has to answer, and it is being asked by people who build benchmarks for a living.