I find this not only refreshing, but at some level astounding.
Nature.
Welcome back to the Nature Podcast.
This week, how mysterious skin structures could help regeneration.
And the AI that helps with literature reviews.
I'm Benjamin Thompson.
And I'm Nick Perchichow.
First up on the show, new research has been uncovering more about mysterious structures in the skin known as reet ridges, which could help unlock ways to help human skin regenerate, combat ageing, and may even give us clues as to why humans aren't furry.
Reet ridges are microscopic peaks and valleys found just below the surface of our skin.
They're thought to harbour the stem cells that can help our skin heal and seem to play a structural role in holding the different layers of skin together.
But despite this, not much has been known about them.
So one of the questions has been, what is it?
And is it a vestigial structure?
Do we even need it?
This is Ryan Driscoll, a skin researcher at Washington State University.
Part of the reason that we didn't know a lot about them is because there's so few models that are easily accessible and ways to actually study it that most of the understanding about how it works inside of us as humans is just unknown.
But that may be starting to change, as Ryan and his colleagues are publishing a new study in Nature that details what may be the best animal to model reed ridges.
And if any of you are skin researchers listening to this, you may be saying well, there's one animal that's an obvious candidate.
And you'd be thinking along the same lines as Ryan.
So I don't think it's a secret, at least amongst skin biologists, that pig skin is more similar to human skin.
The biggest problem is accessibility to the pigs and affordability to work with the pigs and then kind of also doing a little bit of the smelly work as well.
I didn't really want to work with pigs either, even though I knew, as well as the rest of the field knows, that pigs are the better model, because it's really tough to work with them.
So I searched for everything else other than pig.
While it's been thought that pigs are the ideal model for human skin, Ryan and the team wanted to be sure.
After all, pigs and humans are not exactly the closest of relatives.
There are animals such as monkeys that are a lot closer evolutionarily to humans.
So, with a blank slate, Ryan and the team saw skin from dolphins monkeys, naked mole rats, mice and even grizzly bears.
They examined the thickness of the skin, hair density and, of course, whether or not there were reet ridges, as they're not present in every animal.
And with all this examination, they discovered that the animal that had the closest match to human skin was Pigs.
Which is not a surprise.
Not that there weren't some surprises.
It turns out that the skin of grizzly bears was particularly similar to humans.
Grizzly bears have reed ridges.
I could not believe that at all.
Grizzly bears are not exactly close relatives of humans and looking at them they don't exactly seem like they have similar skin.
But Ryan and the team found that their skin was a better match for humans than that of rhesus, macaques or common marmosets.
By looking at all these animals, the team also found some clues about hairlessness.
Generally speaking, it seems that the presence of reet ridges was associated with an animal having thicker skin and less hair.
Animals like dolphins have more and larger reet ridges compared to that of hairier animals, although there were exceptions to this.
Bears are quite hairy and did have reed ridges, and the naked mole rat lacked reed ridges and has relatively thin skin, despite famously being pretty hairless.
The reet ridges then could be to do with making the skin tougher.
When animals had less fur, it's possible that they didn't have the same protection from that hairy layer, and so maybe reet ridges filled that need for the less hairy animals like humans.
But that remains to be seen.
What was clear, though, was that pigs were the best model for human skin.
And with that, the team were able to try and tease out some of the key questions about these reed ridges, such as when do they form?
This is a really important thing to know because it's like laying the blueprint, right?
If you don't have the blueprint at a molecular and cellular level.
You know it's like chance, trying to get something to regenerate.
Knowing how reet ridges form at this level of detail could help skin scientists to understand how to get them to regenerate if they're damaged.
Past a certain age, they don't grow back, and Ryan thinks that may be why skin regenerates less well when we get older.
And that's why there was another surprise.
It had been thought that reet ridges formed during embryogenesis, a very early stage of development, as this is when other parts of the skin, like hair follicles and sweat glands form.
But Ryan and the team showed that, in pigs at least, reet ridges form after birth.
And they showed exactly how this happened.
We found a different molecular mechanism that's distinct from the normal epidermal appendages.
Unlike sweat glands or hair follicles which form similarly, the reet ridges form via their own mechanism, developing where sweat glands or hair follicles don't.
And Ryan and the team found a group of molecules that were involved in reet ridge development bone morphogenic proteins, or BMP.
Reet ridges are thought to be the places where the stem cells needed to regenerate skin gather.
And so these structures may play an important role in the process.
Now Ryan has some molecular clues and a good model animal.
He hopes they can uncover how to restore reet ridges and skin.
You know, one of the biggest unanswered questions is if you want to get human skin to regenerate, how do you get reet ridges to regenerate?
And so now we have that blueprint.
Ryan's next step is to use some of the molecular evidence that he's found to develop a model of reed ridge formation in mice to figure out if they can use this knowledge to help coke skin to regenerate.
They've really opened the door to essentially determine the importance of these organs and then use that to improve the regeneration of skin.
This is Yana Kamberov, a skin scientist who wasn't associated with this new study.
She was impressed by the scope of the work Ryan and the team did, especially their efforts to track down skin samples from so many different animals.
I think what's really kind of unique or really powerful about the study is just the breadth of species that were examined and using that comparative approach, and how powerful it is to actually give us biological information about the emergence of a trait.
So that for me was something that differentiates this paper quite a bit from others in the literature.
Typically, skin researchers have tended to think about all the different parts of the skin being related to each other, like hair follicles and sweat glands having similar development mechanisms, for example.
So Jana was intrigued by the result that reet ridges seem to have a distinct way to form.
What was really fascinating for me to see was that the reet ridges are not abiding really by the rule.
They're their own entity.
They're their own class, essentially, of appendages.
The one thing that Jana thought the paper didn't explain is how humans lost their fur.
Ryan and the team made this quite clear too.
While there is an association between having reet ridges and not being furry, we can't quite say yet whether they explain humans' hairlessness.
It's possible that reed ridges allowed humans to develop tougher skin when they lost their fur, but it's hard to know whether humans lost their fur and then developed reed ridges or the other way around.
Maybe one day we'll be able to figure it out.
I would say it's too soon to tell because we can't go.
It'd be wonderful to take a time machine and be able to get a biological sample from a long time ago to see what that was actually saying.
I think it remains unanswered and hopefully a really interesting question that other scientists can then utilize this understanding to actually go back and really think about domestication, evolution and the acquisition of unique appendages.
That was Ryan Driscoll from Washington State University in the US.
You also heard from Yana Kamberov from the University of Pennsylvania, also in the US.
For more on that story, check out the show notes for some links.
Coming up, the AI system that aims to take the hallucinations out of literature reviews.
Right now though, it's time for the research highlights with Dan Fox.
A pungent weed has been genetically edited to become a potentially profitable winter crop.
Field pennycress is a plant that grows quickly and can survive harsh cold temperatures.
But it isn't particularly useful.
Its distinctive smell led to it gaining the nickname stinkweed, and two molecules found in its seeds can be toxic to humans and animals.
Researchers experimented with editing pennycress using CRISPR for more than a decade, eventually finding a combination of gene variants that made a crop suitable for animal feed, with seeds that produce an oil similar to rapeseed or canola.
Besides increasing farm profitability, growing crops throughout the year can help to improve soil health, sequester more carbon and reduce atmospheric greenhouse gas emissions.
Sniff out that research in Nature Plants.
Astronomers have released the most ambitious cosmic map to date and confirmed that matter in the universe is less clumpy than theory would predict.
From 2013 to 2019, the Dark Energy Survey team repeatedly imaged a large section of Earth's southern sky.
This allowed the team to collect the positions, colours and shapes of around 150 million galaxies and detect more than 1500 supernova explosions.
They looked at four aspects of the data.
The brightness and other characteristics of the supernovas.
The clustering of galaxies across space and time, the evolving size of remnants of pressure waves generated in the universe's infancy and the distortion of images of background galaxies by large concentrations of intervening invisible dark matter.
The results refined previous measurements to confirm that gravity has not clumped galaxies together as much as observations of the early universe would lead us to anticipate, were the standard theory of the universe's evolution correct.
You can find that pre-print on Archive.
You might hear this described as a literature review, or perhaps evidence synthesis.
But either way, it's about combing through the existing research before you set off on your own work.
Of course, this can be a time-consuming endeavor, and more and more frequently, researchers are turning to LLM-backed AI systems to do the legwork.
Ask them a question, and they get searching and respond with a written answer.
But these systems do have drawbacks.
They frequently hallucinate, inventing references or research that doesn't exist.
One of the reasons for this is that these systems rely on statistical patterns and they can put together plausible-sounding answers based on the dataset they're trained on.
This training set could be from a broad range of sources, not just scientific papers, and it may only cover a certain time period.
In our next story, we hear about a team who are trying to do something about this.
They have a paper in this week's Nature describing OpenScholar, their open-source system that consists of a small LLM that's augmented to specialize at this task and which they say can outperform existing commercial systems.
One of the members of the team is Hannah Haji-Sherzi from the University of Washington in the US.
I called her up to hear about the system and she laid out how OpenScholar has been designed.
So here is a very simple idea.
It's called retrieval augmented generation.
It became very popular over the past year.
The idea is this.
Let's add a really high quality data store, a collection of scientific documents.
So then, in order to answer my question, make sure you use this gigantic knowledge base, not just hallucinate from your own knowledge.
So Open Scholar consists of a language model and this kind of information repository.
So a user asks a question.
The retrieval system first tries to figure out which scientific articles are most relevant, and then they start to rank them.
Which one is more relevant?
Now, language model comes into play.
Given this collection of a few articles that are relevant, let's generate a response.
So what you're saying, then is it kind of hones down on what it thinks is the useful information and then uses that to generate a response?
100%.
Think of it as a draft.
Now we start refining it.
We use the LLM to read it.
Is it good enough?
Do we think we cover some important information?
Or that document or scientific article that we retrieved, is it relevant?
Overall, asks a lot of questions and then refines it.
So we do multiple iterations of this and then use the final kind of generated response as okay.
Now, this is the synthesized knowledge.
This is the final piece of answer to you.
And to test how good this final answers to queries was.
I think, as part of your work, you also describe a benchmarking tool as well Scholar QA Bench, which you designed to measure the efficacy of different systems that look to synthesize scientific data in a way that you describe.
How does this system work?
How does it benchmark things?
So in general, when we build these language models, they generate long responses.
Even generic domains, not just scientific domains.
It's very hard to evaluate if their responses are good or not, do they hallucinate or not, and so on.
As part of Open Scholar, we actually built this benchmark to understand the limitations and then start improving it.
So, basically what we did, we worked with scientists, with experts in physics, computer science, biomedicine and neuroscience, to collect expert written questions.
And then we asked these scientists to write responses.
And then we asked them, what is important to measure here?
And it tells us metrics that we need to really care about when we evaluate these systems.
And then finally, when we build a final system, we use again those experts to really go deeply, compare generated responses by two systems side by side to see which one is better.
So you put those questions into the computer and see what it chucks out, and then compare those to a scientist doing the same job.
Is that fair?
Absolutely.
What did your results show then?
So our new system decreases hallucinations significantly.
Its citation accuracy is as good as human experts in generating citations and being accurate on those.
Our final open scholar system outperforms proprietary models like GPT-40 more than 5 across different metrics on our benchmarks.
Right, so it outperforms other systems in your measures of correctness.
But you also show that humans prefer the computer outputs in around 50 to 70 of cases, depending on how OpenScholar is used.
What is it about the system that allows it to provide the numbers that you show in your paper?
So that's the role of that data store that I described like that information repository, and making sure that language model dips into it.
I mean, of course, there are folk who would say that you've designed your system.
You've also designed a benchmark to test how good it is.
And it said it is the best one.
What would you say to them?
That's the best question.
I agree.
So we made a demo for OpenScholar and decided, let's put it in the wild.
Let's see if scientists use it.
To our surprise, in the first few days more than 30000 queries came in and we saw a lot of evaluations from the community that way.
Internally, we started evaluating the responses.
Do we see any issues or not?
And so on.
So this is kind of how we got the positive feedback from the community.
But obviously, we are still improving it.
And it has to be said that this system isn't perfect.
It is not perfect.
Yeah, for example, you say in your paper you know it doesn't consistently retrieve the most representative or relevant papers.
So this isn't necessarily the complete article.
Absolutely.
This is just a start.
And you do note some other limitations to OpenScholar as you describe it in this paper as well.
I mean, I think you mentioned some of the fields it looked at computer science physics, neuroscience and biomedicine.
Those fields are relatively narrow.
You only use open access papers as well, which is obviously a huge amount of research papers.
But there are many, many more that aren't necessarily accessible to this system.
So we've mostly focused on these four domains for evaluation.
This is something that we can say we are good at.
We have only collected open access papers and our repository is still static, meaning that there are more papers that could be added, but we are working with this static collection of documents at this point.
And of course, you've talked there about GPT-4.0.
We're now a little bit further down the road.
And of course, there are new iterations of GPT and of various other systems as well.
Have you looked forwards at how your system compares to more modern LLMs?
If you compare with GPT-5 just out of the box.
So our model is still better because GPT-5 is not designed again to deal with citations.
However, these current systems have something called deep research, which are connecting to the knowledge repositories and try to ground their responses based on those knowledge repositories.
This is great.
It really improves the citation accuracies and so on in these most recent systems.
They are pretty much using the same process that I just described to improve question answering.
And that makes OpenScholar one among many, really.
There's lots of these things that are in existence from various companies.
What do you think makes OpenScholar stand out?
Obviously, your paper's out.
What are the key things for you?
So still, a lot of this work from these companies are designed to answer any type of question.
And as a result, still hallucinations exist.
So OpenScholar still stands in this bigger collection of AI systems because it is a specialized for science.
It is really designed to kind of deal with scientific papers as this knowledge repository and so on and so forth.
And you say you've used it yourself.
Do you think it's been useful to you as a researcher?
Absolutely.
Yeah, it was actually very useful.
And this is how we kind of decided let's release the demo, because we found very interesting responses.
One thing for us was we needed to kind of instruct scientists how to collaborate, query and how to ask question.
Because for example, we came with some queries like, hi, how are you?
This is not the type of queries that it is designed for.
It's more like we are looking for a scientific question and then we want to kind of use the output.
And we got actually very interesting responses and kind of testimonies of, oh, this was very useful.
We love to continue using it.
Are you maintaining it?
Like these type of questions was very exciting to see.
Hannah Hachicherzy from the University of Washington there.
To read her paper, look out for a link in the show notes.
That's all for this week's show.
Don't forget to look out for the briefing show on Friday.
That'll be on the same feed as this one, so it should just appear in your podcast player of choice.
In the meantime, you can reach out to us on social media.
We're at Nature Podcast.
I'm Benjamin Thompson.
And I'm Nick Petrichow.
Thanks for listening.
