Patrick Martin

A roadmap for evaluating moral competence in large language models - Nature This Perspective offers a roadmap for tackling the challenges of the facsimile problem, moral multidimensionality and moral pluralism in large language models.

The distinction between Moral Competence and Moral Performance is key: even if we cannot distinguish them in a model's output, Moral Performance is a function of who own's the LLM.

www.nature.com/articles/s41...

Laser writing in glass for dense, fast and efficient archival data storage - Nature An optical archival storage technology based on femtosecond laser direct writing in glass addresses the practical demands of archival storage.

So... Data Crystals? Sci-fi is just becoming science at this point.

www.nature.com/articles/s41...

Nice! It must have been a tough translation given the amount of word play in that book.

Making large language models reliable data science programming copilots for biomedical research - Nature Biomedical Engineering AI is a powerful tool for facilitating data science programming in biomedicine. This study benchmarks the performance of existing tools and introduces a reliable AI agent copilot.

Making large language models reliable data science programming copilots for biomedical research www.nature.com/articles/s41... #llms #bioinformatics #datascience #coding

PersonaAI developed by @wonkj.bsky.social lab at Cedars-Sinai!

Human in the loop agentic AI system to accelerate hypothesis testing in aging research. It combines a RAG summary of over 600k papers and uses scRNA atlases to convert ideas into verifiable hypotheses!

www.biorxiv.org/content/10.6...

A new sci-fi short story on my Substack: Voices.

What happens when AI tries to find its own voice?

#AI #Scifi #ProjectSuncatcher

pcnmartin.substack.com/p/voices

Another great piece of work by Guangzheng Weng and Kyoung-Jae Won's Lab! Glad to have been part of it.

The model incorporates LLM embeddings and ChIP-seq data in a Transformer-based Graph Neural Network to generate GRN predictions.

advanced.onlinelibrary.wiley.com/doi/10.1002/...

Science in 2050: the future breakthroughs that will shape our world — and beyond Nuclear fusion. People on Mars. Artificial general intelligence. These are just some of the advances that could come by the mid-century mark.

Don't let anyone tell you reading sci-fi is "not work".

www.nature.com/articles/d41...

D&D Coding edition

Wizard = Backend Dev
Artificer = IOS and Android Only
Bard = Front End Dev
Sorcerer = Leet Coder
Warlock = Vibe Coder

Okay buddy.

You could also rake it yourself.

Technological and scientific progress needs to move at the same pace as social progress.

Society is like a big ball of pizza dough. You pull too hard in one direction, you end up with holes in your dough, sauce all over, and a shitty pizza.

Nice and even is how you make the perfect pizza.

I would add: they make your mess someone else's problem.

Along these lines, I believe that, sometimes, it's worth slowing down. It's worth remembering that it is simply not feasible to follow through with all your ideas.

www.oliverburkeman.com/river

I'm just going to say it:

Faster iterations and increased productivity with the use of AI means you should make your "product" of much higher quality. Not producing half baked ideas at a faster rate.

Don't ship more garbage. Ship the same amount but better.

#AI #VibeCoding

I can totally see it as some sort of russian roulette of traps. This was a simple spike coming out of the wall? Now it's snakes. You could use some custom wild magic tables.

I'm adding that to my next dnd game. Shuffle doors. Shuffle cards. Shuffle traps. Endless fun

I have been playing around with the
@chanzuckerberg.bsky.social CellXGene census for mouse data. There are a couple of interesting cell types in there.

One of them is just called "cell". It's a cell and its identity is none of your business.

I have seen a few of these bands live. Welcome to the strange world of metal.

We live in the decade of generative models and biology is no exception.

Genome language models can now generate functional bacteriophages!

Read more in my new sub-stack post where I explore the @arcinstitute.org latest pre-print.

We live in the decade of generative models and biology is no exception.

Genome language models can now generate functional bacteriophages!

Read more in my new sub-stack post where I explore the @arcinstitute.org latest pre-print.

FedscGen: privacy-preserving federated batch effect correction of single-cell RNA sequencing data - Genome Biology Single-cell RNA-seq data from clinical samples often suffer from batch effects, but data sharing is limited due to genomic privacy concerns. We present FedscGen, a privacy-preserving communication-eff...

Federated Integration of single cell data!

Now this is an amazing initiative. No need to share the data.

But, and maybe more importantly, demonstrates that complex tasks such as integration can be done using distributed compute power.

genomebiology.biomedcentral.com/articles/10....

spCorr: flexible and scalable inference of spatially varying correlation in spatial transcriptomics Spatial transcriptomics has transformed our ability to explore gene expression within its tissue context, enabling us to dissect subtle yet biologically significant variations in situ . While numerous...

This is actually a pretty neat idea. Spatially correlated genes.

www.biorxiv.org/content/10.1...

Well. I'm not going to say I am surprised by these results. Come on people. We can do better than this.

www.nature.com/articles/s41...

#AI #Honesty

Often working at the intersection of AI and Biology, I have developed a love-hate relationship with AI.

In this essay, I share some thoughts on the modern AI ecosystem. It’s not all doom and gloom but there is room for improvement.

pcnmartin.substack.com/p/a-love-hat...

Multi-scale and multi-context interpretable mapping of cell states across heterogeneous spatial samples - Nature Communications The alignment of heterogeneous spatial samples has become a growing challenge. Here, authors present a multi-scale, multi-context, and interpretable mapping strategy to map cells across space, time, a...

Vesalius 2.0 is in ‪‪@natcomms.nature.com‬ !

We show how using spatial context improves cell-to-cell mapping across heterogenous spatial samples.

Thanks to the reviewers, editors, and collaborators for making this possible.

GitHub: github.com/WonLab-CS/Ve...
Paper: www.nature.com/articles/s41...

It has been quite fun and interesting trying to run scRNA analysis using only distributed/parallelized packages. Finding approximations to non-parallelizable algorithms is quite the challenge.

It also makes you appreciate the amazing work behind Scanpy from the @scverse.bsky.social and Seurat.

Thank you @qchengx1e3.bsky.social for handling this manuscript! Great to see it out.

Multi-scale and multi-context interpretable mapping of cell states across heterogeneous spatial samples - Nature Communications The alignment of heterogeneous spatial samples has become a growing challenge. Here, authors present a multi-scale, multi-context, and interpretable mapping strategy to map cells across space, time, a...

Vesalius 2.0 is in ‪‪@natcomms.nature.com‬ !

We show how using spatial context improves cell-to-cell mapping across heterogenous spatial samples.

Thanks to the reviewers, editors, and collaborators for making this possible.

GitHub: github.com/WonLab-CS/Ve...
Paper: www.nature.com/articles/s41...

Does this plot look ugly? Yes it does.

Am I happy with because of what it represents? Yes I am.

Simulated Gene Pulse