Full transcript of “HUGE breakthrough! This AI discovers unknown molecules”. Click a timestamp to open the video at that moment. Need another video? Use the YouTube transcript generator.
0:00 Life is made of molecules. These are the fundamental building blocks that make up literally everything, including us and other living things and the rest of the world. The thing is, we've only identified less than 10% of all the natural molecules out there. What if there was an AI that can help us find the rest? If we do that, this could lead to huge discoveries in all areas of life, including disease diagnosis, drug discoveries, longevity, new chemicals for batteries and electronics, new materials, and so on. The possibilities are endless. Well, this AI might be the breakthrough that helps us get there. So, this paper is titled self-supervised learning for molecular representations from millions of tandem mass spectra
0:46 using dreams. And this was published very recently. Now, this is a very technical paper, but as always, I'm going to make this easy for everyone to understand. Before we get into that, here's a brief background on what natural molecules even are. So, for the human body and every other living thing out there, it's basically a massive chemical landscape. We can be broken down into different molecules with different functions, and these are crucial for life. For example, there are molecules that are responsible for metabolism, healing, and fighting diseases. The thing is, we've only identified less than 10% of all the natural small molecules out there. There's still a huge unexplored universe of molecules that we don't know about. If we can figure out the rest of this
1:33 90%, it can unlock so many incredible things like new drugs, materials, or chemicals. The problem is historically it's been really hard for us to identify what's out there. The main tool that scientists use to identify these natural molecules is called tandem mass spectrometry coupled with liquid chromatography or LCMS for short. This basically separates a sample's molecules and fragments them to create a spectrum like this. Think of it as like a molecular fingerprint. Now this part is easy. We can take anything in the world and break it down into molecules and map it on a spectrum like this. However, the hard part is interpreting this spectrum or fingerprint. Less than 10% of these
2:18 spectra can actually be identified and matched to known molecular structures. So over 90% is just dark data. We have no idea how to interpret this. So we have a huge amount of information with all the spectra from all the different natural materials that we're able to gather, but we have no idea how to interpret or use this. So in this paper, the researchers used a new AI approach called the dreams neural network to figure this out. The idea is to use all this data that we couldn't read before and train an AI to finally start mapping this hidden chemical universe and hopefully uncover some secrets. Like I said a moment ago, the problem isn't a lack of data. We have a ton of these spectra already. In fact, hundreds of
3:04 millions of these molecular fingerprints from all the natural materials out there. We just don't know how to interpret this or figure out the molecular structure behind it. So here the scientists use a selfsupervised learning method and architecture looks like this. But in the simplest sense, it's basically a neural network that's trained to interpret these spectra. What would be the properties of the molecule behind a spectrum like this? Now, you might be wondering if we have all these spectra, but we don't have the answer. In other words, we don't know the chemical or the structural properties of all the molecules behind these hundreds of millions of spectra. Well, how exactly can we train an AI to learn this? Well, you can think of it as like learning a language just by reading a ton of books in that language. You don't
3:50 necessarily need someone to explicitly tell you that this word means X for every single word. If you spent an eternity just reading tons of books in that language, you'll start to see some words and patterns again and again and eventually you'll start to indirectly figure out the grammar and the meaning of everything. So in a simple sense, that's kind of what this self-supervised learning stage is doing. The AI model is fed a ton of data. In fact, they mined this data set called GNPS and extracted around 201 million of these spectra that are not labeled or identified. And after feeding all of this data to the AI and going through millions of rounds of self-supervised learning, the AI kind of learns the grammar or the language
4:36 behind this. In other words, it learns how these molecules would break apart and then be mapped onto these spectra. And from this, it would also know how to decode the chemical or structural properties of the molecules behind the spectra. Note that this doesn't mean the AI can guess the exact molecular structure just from a spectrum, but it's able to kind of guess its chemical and molecular and structural properties of the material. Now, after the AI has learned to read or decode these spectra, it then goes ahead and decodes these 201 million spectra of natural molecules that we know nothing about. And then it maps each one onto this huge multi-dimensional space called the dreams atlas. So this dreams atlas is
5:21 essentially a map which contains all these 201 million natural molecules as well as the molecules that we already know about. And the position of each dot is based on the similarity of that molecule to others. In other words, molecules with similar properties would be closer together in this map. And if two dots are really far apart, then they would be vastly different in terms of their structures or chemical properties. By the way, this map is kind of like how words of a language are mapped for large language models like GPT and Gemini as well. So for example, let's say we have the word tower. Well, the words gate or building or skyscraper or roof would be roughly in the same space and category as tower. So they would be grouped roughly in the same cluster on this
6:08 multi-dimensional chart. Also, the distance and the direction of the differences between different words also have an inherent meaning. So the distance and direction between man and woman would be roughly the same as that of king and queen. So anyways, you can think of this dream atlas as being the same thing, but instead of a graph for all the words in a language, it's a graph of the spectra of all the natural molecules we were able to extract so far. Now everything is quite theoretical. You might be wondering after mapping all of these molecules onto this network graph, what exactly can we do with this? Why is this such a breakthrough? Well, here are some of the key findings. First of all, one really interesting observation is how connected the entire thing is. This suggests that dreams was able to find meaningful
6:55 similarities that link all these molecules together, even for molecules that we've never identified before, which is like over 90% of everything here. In a sense, you can think of this as like a framework or a Wikipedia on any molecule out there. For example, you can take any natural molecule and break it down using this mass spectrometry liquid chromatography technique and then turn it into a spectrum. And then you would plug it through this dreams AI to find its position or embedding in this graph. And from its position, you can see its neighbors, right? The dots that are closest to it. And you can inspect what properties would these neighbors have. And this gives you immediate clues about your unknown molecule. If it's really close to a known molecule, then you would know that it's probably very similar. You can also use this atlas to
7:42 determine how new a molecule is. For example, if it's really far away from anything, then that might mean it's a very different molecule with properties that we've never seen before. Now, remember in this atlas, they also included the 10% of molecules that we've already identified and know about. Well, it turns out that there are a ton of dots that are very far away from any of these known dots. This means that there's still a huge portion of the natural molecules out there that are truly distinct from what we know. It's an indicator of how much chemical novelty is still out there just waiting to be discovered. That means that there could be drugs that are way more effective or way better chemicals or materials for batteries and electronics. Just to give you some examples, we haven't even scratched the surface of
8:28 what's possible. So, this atlas isn't just remapping the known world, but it's actually showing us the shape and scale of what has not been discovered yet. Now, again, everything I've said so far is quite theoretical. So, let's actually go over some real examples of what Atlas was used for. So, here the researchers took a ton of different and diverse food items like oranges, grapes, tomatoes, avocados, ground coffee, water, milk, potato chips, Oreo cookies, chicken, turkey, ground beef, etc. And they broke all these down into molecules and then used that liquid chromatography technique to fragment the molecules into spectra. The spectra were then plugged through this dreams AI and plotted on this network graph. And what they found was that all the food items clustered perfectly according to basic food
9:15 taxonomy. All the plant-based foods were clustered together. The animal-based foods were also near each other and then beverages also formed another distinct cluster. So this proves that the AI can correctly classify molecules based on their properties. The AI would see that for example ground beef and turkey would be more related to each other than for example coffee which is all the way over here. Note that this AI knows nothing about the food sources where the molecules came from. It was only given the spectra of all these molecules but yet it was able to perfectly capture the biological and chemical relationships of all these food items. So this is an incredibly cool and powerful validation. Here are some other key findings. Apparently they found a really close linkage between psoriasis which is a
10:02 skin disorder that causes red scaly patches and the fungicide esox strobin. So again here is the entire atlas. But if you zoom in on this point you'll see this. You'll see a lot of different unrelated molecules like citrus tree and then hand over here face arm pesticide mix fruits and vegetables. But then we have arm psoriasis over here and it's quite closely linked to exoxy strobin which is a common chemical used in agriculture. So this might suggest that exposure to this chemical could cause this skin condition because they're so closely related in this atlas. Now note that this atlas just maps correlations. So you can't really conclude that there's any causal link here. But the relationship of these two things might be worth investigating further. So in a
10:49 sense, Atlas kind of acts like a discovery engine. It finds surprising molecular similarities between two very different things. And so we can then investigate further how these two things might be related. Here are some other interesting findings from this atlas map. So they found that this plant metabolite, which is basically a chemical that's created by the plant to help it grow and survive, it kind of appears among all these different types and species of plants that are seemingly very unrelated to each other. but they all share this same plant metabolite. So, is there something going on there? It also identified a whole family of lipids or fats that's not only really close to type 2 diabetes, but also for brain cancer, lung cancer, and renal cancer. Again, we can't say anything
11:35 conclusive here, but this atlas is like a hypothesis engine. We can then look at this family of lipids further and see how exactly it's related to all these different cancers. And from that, maybe we could come up with new cures to fight these types of cancers. Now, just like regular AI models, dreams can also be fine-tuned further to predict specific properties based on the spectrum of a molecule. So, the researchers fine-tuned dreams to predict how relevant a molecule is to the Leinsky's rule of five. This is basically a set of rules that determines whether a chemical or molecule is a suitable drug and can be absorbed well in the human body. This might not seem like much, but imagine taking a ton of natural molecules, which
12:20 we know nothing about, and then we can just plug it through this AI to determine which ones can potentially be used for drug candidates. This is a super powerful tool. Thanks to Monica for sponsoring this video. Monica is an AI assistant that provides access to the best AI tools all in one place. This includes the top AI models like GPT, Deepseek, and Gemini, plus the top image generators like Flux and Stable Diffusion and the top video generators like Cling and High Law. This is a killer deal because you can use all of these in one platform for way cheaper than if you used all of these separately. You can use this as a browser extension on your desktop or on your mobile device. And their browser
13:06 extension is contextaware. So you can interact directly with any web page you're on. You can easily get Monica to reply to emails or summarize an article. So for example, let's say I'm on this page. I can get Monica to summarize this really technical article in just one click. And here's something even cooler. You can even generate a mind map from this content which can help you understand ideas more visually. Note that this is way more convenient than just copying everything and going to chatpt to ask it to summarize the article. You can also use Monica to summarize YouTube videos. So here's an example. Let's say I'm watching this YouTube video. I can get it to generate a summary or even generate a podcast in just one click. Let's try this one. It even gives you a list of highlights with
13:52 the correct timestamps. Plus, if you click on this again, it's generating a mind map for you so you can understand everything more visually. With Monica, everything is just a few clicks away. But don't take my word for it. They have over 10 million users worldwide with a 4.9 star rating on the Chrome store. Try it out for free via the link in the description below. And when you're ready to upgrade, use my code AIR10 for a nice discount. Full details in the description below. Here's another example. They also fine-tuned dreams to predict the presence of florine in the molecule. Now, why is this particularly important? First of all, florine is used in many industries and everyday life. If a molecule contains florine, it often means the compound is very stable and
14:39 resistant to heat and chemicals. Such compounds are widely used in pharmaceuticals for more effective drugs, but also non-stick coatings like Teflon and also refrigerants, high performance plastics, and electronics manufacturing like semiconductors, just to name a few examples. But here's the problem. It's really hard to determine whether a molecule has florine. Previous methods weren't really able to predict this well, but Dreams was able to handle this very well. So they fine-tuned the model to predict florine presence from these molecular spectra and the results were quite impressive. When dreams flags that a molecule contains florine, it's correct 91% of the time. Whereas older methods like Sirius only gets around 51% precision. So this is a huge
15:25 improvement. From dreams, we can plug over like 200 million different natural molecules into this and filter out the ones that contain florine with pretty high certainty. And these would be super stable molecules that can be used in a ton of different applications. And these are just some examples that they mentioned in the paper. You can potentially fine-tune the model further to look for molecules of certain properties. For example, ones that could fight cancer or ones that could be anti-aging or molecules that could fight pollution or digest plastic. Remember, we only know less than 10% of what's out there right now. There's still a huge chemical universe that we can tap into. Scientists can use this atlas and find a ton of new drug candidates by looking at
16:10 regions or clusters that are near drug candidates that we already know about. Or they can tap into areas which are really far away from anything we know about. And this can uncover molecules with completely new chemical properties that are unlike anything we've seen before. This is only the tip of the iceberg. Plus, right now, this AI can only take a spectrum and kind of guesstimate the chemical properties and composition of that molecule. What's even better is if we can eventually train an AI that can take this spectrum and predict the full structure of the molecule. That would be the ultimate goal. We're not there yet, but Dreams is definitely a step in that direction. Now, here's the awesome thing. They've actually released the code to this. Everything is on GitHub and HuggingFace. This is under the MIT license, and it
16:56 gives you all the instructions on how to use Dreams to map a molecular spectrum onto a network graph. It also gives you instructions on how to fine-tune Dreams to look for specific properties. All the instructions are over here for you to get started. So, that sums up my review of this paper that was released recently. It's quite a technical paper, but hopefully this video made it easy for you to understand. Let me know in the comments what you think of this. And if you're interested in this kind of stuff, I've actually made a few other videos that are quite similar. So, check out this video about Google's Alpha Evolve, which has already discovered a ton of breakthroughs in math and physics, or this video, which goes over another AI that can accelerate discoveries in material science. As always, I will be on the lookout for the
17:43 top AI news and tools to share with you. So, if you enjoyed this video, remember to like, share, subscribe, and stay tuned for more content. Also, there's just so much happening in the world of AI every week, I can't possibly cover everything on my YouTube channel. So, to really stay up to date with all that's going on in AI, be sure to subscribe to my free weekly newsletter. The link to that will be in the description below. Thanks for watching and I'll see you in the next one.
