scRNAseq cell type annotation is notoriously messy. Despite so many algorithms, most researchers still rely on manual annotations using marker genes. In a new preprint accepted at ICML GenAI Bio Workshop, we ask if reasoning LLMs (DeepSeek-R1) can help with cell type annotation🧵

just realized that most deep domain experts are very similar to reasoning LLMs: very knowledgable, most of the times right in their judgement, sometimes wrong, but unaware/unwilling to admit they are wrong 🤯.

Join millions who have switched to Grok.

and they don’t like eachother.

first mover advantage 101

what's the best AI agent for science right now?.

RL scaling is meant to hit a wall in real-world tasks because such tasks are designed to be done by many people together & include lots of redundancies & inefficiencies. anybody who worked among many people knows that it’s seldom the “best” (by a clear metric) solution that wins.

extremely excited to share my lab's research of AI for cancer genomics at the neurIPS 2025 workshop on multi-modal foundation models in san diego.

This R packages can transform any plot using color pallets from painting at the Museum of Modern Art in New York.

I disagree with lots of what François (used to) say, but I think he's super right here.

computational biology/bio ML has changed more in the past 2 years than in the last 10. “i have experience with CNNs for images, can I get an AI research job” doesn’t work anymore. if you’re trying to solve a pure prediction problem, you need to constantly know SOTA AI methodology.

I might get this wrong, but why on Earth do people use chatGPT to generate “in-depth” blog posts on a technical topic? Doesn’t this literally beat the purpose? Like what is the point, what value does this bring to this world, what can one learn/gain from this? I don’t get this.

7. as more & more scRNAseq data will be generated, we believe that general-purpose reasoning LLMs s.a DeepSeek-R1 can be reliably employed for automating cell type annotation & also multiple other scRNAseq analysis tasks, streamlining these analyses across the scRNAseq community.

5. reasoning LLMs are better annotators than non-reasoning LLMs (DeepSeek-V3, GPT4o). 6. overall, the best performing models on single-tissue real-world situations were the DeepSeek-R1 classifiers, especially when prompted with scGPT's large list of labels to choose from.

4. reasoning LLMs (DeepSeek-R1-0528) are competitive & interpretable cell type annotators. Their performance is sometimes higher & sometimes lower than that of alternative tools. But, on most real-life scenarios & especially on less standard cells, they are as good or even better.

Putting everything together, we found that:.1. benchmarking cell type annotation tools is inherently difficult. 2. cell typing should be both accurate, but also prone to discovery. 3. regardless of the approach - some cells are really easy to annotate, while others are difficult.

We tested 5 more tissues:.- lung: highest accuracy for scTab & Macro-F1 for regular DeepSeek-R1.- cerebellum: highest accuracy & Macro-F1 for DeepSeek-R1.- trachea & kidney: highest accuracy & Macro-F1 for scTab.- peripheral retina: highest accuracy for R1 with scGPT classifier

On breast cells, DeepSeek-R1 in scGPT classifier mode was most accurate, while scTab had highest Macro-F1. Regular DeepSeek-R1 was particularly poor in accuracy (2.9%), partially due to granularity issues or mistaking the basal & luminal lineages in cell typing.

For pancreatic cells, DeepSeek-R1 classifier with access to scGPT labels had 80.9% accuracy (50.7% for regular DeepSeek-R1), while scGPT's accuracy alone was only 3.4%. This further suggests that it is hard for expert models to generalize, while LLMs are highly adaptable.

In contrast, when given specific indications to choose from either scGPT’s or scTab’s labels, DeepSeek-R1 in classifier mode correctly labeled a much larger fraction of blood cells, with even greater granularity than the ground truth data, providing additional biological context.

In blood, all models except DeepSeek-R1 & its classifiers struggled with hitting the granularity level of the ground-truth data (e.g predicted “leukocyte” for the ground-truth “naïve B cell”), suggesting generalization challenges on novel datasets, despite lots of training data.

Diving into tissue-specific metrics:. For blood cells, the two DeepSeek-R1 classifiers had highest accuracy & Macro-F1, while scTab & scGPT performed poorly, despite almost complete cell type label overlap (99.92%) with scTab & blood the most frequent tissue in scTab’s training.