EvoDiff Unleashed: Microsoft's Breakthrough in Open Source Protein Generation

0:00The landscape of protein design, which is a cornerstone in understanding and treating diseases, I think is on the brink of a transformation. So proteins, for those that don't know, essentially serve as the natural molecules executing critical cellular functions. But creating them in a lab has often been a really, like, it's really costly and it's very complex. However, this is all potentially about to change. Microsoft claims to have simplified this really intricate process with its newly introduced framework, which is called EvoDiff. So traditionally, protein design has kind of a bunch of extensive computational and human resources that are needed in order to do it.

0:40And scientists have to conceptualize a protein structure capable of performing specific bodily tasks. And then once they do that, they have to determine the sequence of amino acids that would likely fold into that structure. Okay, so kind of complex scientific stuff. But protein folding is essentially as proteins have to adopt. They have to essentially adopt these precise three dimensional shapes to function as they're intended. So that's what protein folding is. So EvoDiff offers a radical kind of departure from what has been the norm for so long. So according to Microsoft, the general purpose framework can churn out, you know, high fidelity diverse proteins based solely on a protein sequence.

1:23So this completely sidesteps the often very cumbersome task of requiring structural information about the target protein. And this open source framework holds promise in essentially applications ranging from generating enzymes for novel therapeutics to also kind of like facilitating new industrial chemical reactions. So Kevin Yang, who is a senior researcher at Microsoft, he envisions EvoDiff as a groundbreaking tool in protein engineering. That's what he said. He said, quote, We envision that EvoDiff will expand capabilities in protein engineering beyond the structure function paradigm towards programmable sequence first design.

2:01So he also said with EvoDiff, we're demonstrating that we may not actually need structure, but rather than protein sequence, rather that or rather that proteins protein sequencing is all that you need to controllably design new proteins. Really, really interesting stuff. So I think at the core of EvoDiff is around a 640 million parameter model trained on a comprehensive data set spanning various species and functional classes of proteins. So for the, you know, uninitiated, parameters in an AI model are essentially learned from training data and essentially dictate the model's competence. So data sources for the model include the OpenFold data set for sequence alignment and Uniref50, which is essentially a subset of data from the renowned Uniprot consortium database.

2:52So drawing parallels to cutting edge image generating models like stable diffusion and Dolly2, EvoDiff operates as a diffusion model. This is really interesting because, you know, these diffusion models, for those that don't know, essentially what that really means is it's like if so we'll go to like images, but it's really interesting. They've moved this over to protein and science and other areas. So this is kind of like, in my opinion, this is another testament to like how cool and impactful a lot of these a lot of these advances in AI are because they're not just like for, you know, generating images on mid journey or generating text on chat GBT.

3:27like the way that they're built and the architecture of these tools is now being used in so many other things and actually i i'll also just as a side note say i think it is really incredible people people just hear the word ai and in their brain they're like yeah ai is just like a computer doing stuff um and i think they don't really realize the fundamentally like how incredible it is that we have image ai and also text generating generating ai coming up at the same time because these are actually fundamentally different in how these things function like of course you need data it in for both of them but for diffusion models which is what this new um evo diff is using but how it works for images essentially diffusion model means that like when it when an image is rendering it's kind of like chat gbt where it's predicting the next token but instead of predicting the next token it's predicting the next pixel in an image and so you can imagine it where it's like a square and you've probably seen this on if you've used something like mid-journey where you like look at the image and it's like blurry and it slowly kind of like comes into focus right so that's what a diffusion model is doing it's essentially rendering it where it like renders like a really fuzzy bunch of pixels in a square and then it's like if we were trying to do x y and z like how would the pixels change what's the prediction and it slowly like almost like comes into focus it's diffusion and it's uh coming into focus of what it's actually supposed to do predicting all the pixels on the placements super super interesting stuff um so it's really cool because that same technology is now getting moved into other areas like evo diff so essentially how evo diff works is that it refines a protein made mostly of you know quote unquote noise and then gradually filtering out the distractions to arrive at an accurate protein sequence such as you know the same thing that diffusion models are kind of how they're not like confined to obviously proteins or applications really stretch across a bunch of different domains including music and speech synthesis synthesis so really that diffusion model i think is very very like influential in the sense that like of course we discovered it to create images but now we're literally using it for image generation we're using it for protein generation we're using it for like music creation speech creation all sorts of really interesting things and it's kind of like this diffusion model that's that's doing all that and really really cool stuff in any case evo diff cannot only create new proteins but it can fill in the gaps in existing protein design so that's according to Ava Amini, which is another researcher, a senior researcher at Microsoft.

5:52So the framework's versatility allows it to essentially generate protein amino acid sequences meeting specific functional criteria and even to synthesize quote unquote disordered proteins. So those don't typically fold into a final structure, but they still play vital roles in biology and diseases and stuff. So very, very interesting. And while EvoDiff appears promising, I think it's fairly essential to note that the research has yet to kind of undergo peer review. So this is coming out of Microsoft. And then Sarah Alamadari, who is a data scientist at Microsoft, cautioned that there's still, quote, a lot more scaling work needed before commercial application and also said, quote, this is just a 640 million parameter model and we may see improved generation quality if we scale up to billions of parameters.

6:40So I think looking ahead, the EvoDiff team plans to kind of validate the generated proteins in the lab. If successful, this is going to pave the way for the framework's next iteration, which is opening new vistas in protein engineering and healthcare innovation, all sorts of really exciting things. So definitely a story we're going to continue following, and we're really excited to see how this continues to advance and play out.