WEBVTT 00:00.000 --> 00:07.440 So, hey, and I guess you get to introduce yourself. 00:07.440 --> 00:08.440 Absolutely, hey, everybody. 00:08.440 --> 00:09.440 My name is Ben Busby. 00:09.440 --> 00:13.120 I'm going to try to cram three lightning talks into a talk. 00:13.120 --> 00:16.920 First, I'm going to tell you about some cool stuff in videos building the bioinformatics 00:16.920 --> 00:17.920 space. 00:17.920 --> 00:22.560 Second, I'm going to talk about what would happen if you wanted to sequence and distribute 00:22.560 --> 00:24.440 millions of genomes. 00:24.440 --> 00:28.960 And then third, if any of you are interested in knowledge graphs or rag, I'm going to talk 00:28.960 --> 00:31.160 about cramming knowledge graphs and rag to get it. 00:31.160 --> 00:32.160 Everybody good with that? 00:32.160 --> 00:33.640 All right, let's roll. 00:33.640 --> 00:38.760 So yeah, I'm Ben Busby here, some disclosures, but I will say that a lot of us, with a bunch 00:38.760 --> 00:42.960 of academic institutions and stuff, but a lot of us, and in video are really from the open 00:42.960 --> 00:43.960 source community. 00:43.960 --> 00:49.640 So, part of it, I remembered my galaxy t-shirt, but I did forget my NFC or socks. 00:49.640 --> 00:55.540 Anyway, so yeah, we work at an Nvidia, we work with pretty much everyone in the sort of 00:55.540 --> 01:01.060 industrial bioinformatics ecosystem, and also hundreds and hundreds of academic institutions, 01:01.060 --> 01:06.100 really trying to help speed up algorithms and bioinformatics. 01:06.100 --> 01:11.380 It's one of the main things we do, and really across the entire sort of biological science 01:11.380 --> 01:15.140 as ecosystem, it's not shown here, but including agriculture. 01:15.140 --> 01:19.020 I'll talk about a few of these things, but today I'm not going to talk about robots, if 01:19.020 --> 01:22.100 you want to talk about robots after we can do that. 01:22.100 --> 01:28.980 Anyway, so yeah, Nvidia builds all kinds of stuff and also accelerates all kinds of open 01:28.980 --> 01:33.500 source algorithms so you can use them faster and faster and then cramps them together 01:33.500 --> 01:35.100 in convenient ways. 01:35.100 --> 01:39.700 So that's something to think about, but for those of you who are bioinformaticians, 01:39.700 --> 01:46.540 there's about 20 massively sped up regular bioinformatics algorithms. 01:46.540 --> 01:51.340 Here are about 10 of them, and if you're interested, check out Parabrics. 01:51.340 --> 01:57.060 By the way, some of them are in NF Coral, ready, and others are in galaxy. 01:57.060 --> 02:01.340 So that's a really nice thing that I'll talk a little bit about some things we're doing 02:01.340 --> 02:03.540 in single cell, as well as models in a minute. 02:03.540 --> 02:08.780 But if you're a bioinformatician, you want to go faster, Google for Parabrics. 02:08.780 --> 02:10.860 What do I mean by massively faster? 02:10.860 --> 02:19.580 So on an RTX Pro 6,000, BWA is about 135 times faster than it is on a relatively comparable 02:19.580 --> 02:25.020 CPU, so there's a lot of speed up which actually turns into being cost savings if you're 02:25.020 --> 02:29.260 sequencing or mapping a lot of genomes. 02:29.260 --> 02:34.280 So for single cell, things are even faster, so the average single cell workflows about 02:34.280 --> 02:41.640 200 times faster on GPU, you can check it out, and so that's to me that's really, really 02:41.640 --> 02:42.640 cool. 02:42.640 --> 02:45.760 So if you're interested in millions of cells or particularly hundreds of millions of 02:45.800 --> 02:49.680 cells, GPU really just makes sense. 02:49.680 --> 02:52.840 One of the really cool things about this is this all works in Jupiter notebooks too, 02:52.840 --> 02:56.240 so you don't have to know Kudai anything like that, you can just run it. 02:56.240 --> 03:00.080 And then also a lot of people don't know that there's accelerated data science. 03:00.080 --> 03:06.880 If you Google Rapids.ai, there's a pandas, there's a numpy, and I think most importantly, 03:06.880 --> 03:10.720 there's a psychic learn version called QML. 03:10.760 --> 03:14.720 In addition, in addition, in addition, in addition, in addition, in addition, in addition, 03:14.720 --> 03:18.720 in addition, in addition, in addition, in addition, in addition, in addition, in addition, 03:18.720 --> 03:22.720 in addition, in addition, in addition, in addition, in addition, in addition, in addition, 03:22.720 --> 03:26.720 in addition, in addition, in addition, in addition, in addition, in addition, in addition, 03:26.720 --> 03:30.720 in addition, in addition, in addition, in addition, in addition, in addition, in addition, 03:30.720 --> 03:34.720 in addition, in addition, in addition, in addition, in addition, in addition, in addition, 03:34.720 --> 03:38.720 in addition, in addition, in addition, in addition, in addition, in addition, in addition, in addition, 03:39.200 --> 03:43.600 I think I'm more or less on time, but the vision is that, you know, in five years for people 03:43.600 --> 03:49.760 with very common ideologies of disease, we can really move towards hitting them pharmacology 03:49.760 --> 03:55.120 more or less straight off the sequencer. But I think something we ignore all the time, we think 03:55.120 --> 03:59.120 about variants of this variants that variant disease blah, blah, blah, but, you know, I mean, 03:59.120 --> 04:05.520 about 8% of variants have single disease, single variant associations, according to NHGRI, 04:05.520 --> 04:11.280 and other places, the vast majority of diseases, multi-factorial. 04:11.280 --> 04:16.080 Variant annotation is very common, there's 300 variant annotators, in this tool called open 04:16.080 --> 04:20.880 provide, but still we're missing a lot, we have a lot of variants of unknown significance, why, 04:20.880 --> 04:25.680 probably because of lack of contextualization, because there's actually polygenic. 04:25.680 --> 04:32.000 So here's a bunch of reasons that diseases are multi-factorial, but why don't we think about this 04:32.000 --> 04:36.720 very well? Well, it turns out humans don't like to think about multi-factorial causation. 04:36.720 --> 04:42.400 Deep learning, fortunately, does not have this problem. So, but anyway, so we decided to try to 04:42.400 --> 04:47.920 attack this problem head on in a very, very simplistic way. Some of these things, thank you. 04:47.920 --> 04:52.160 Some of these things are seem quite complicated, but we thought we could do it simply, 04:52.160 --> 04:56.480 and by me, I mean, me and a bunch of really clever graduate students at CMU, 04:57.440 --> 05:02.720 we downloaded a thousand genomes split them by habitat blocks, or sorry, 05:02.720 --> 05:08.720 recombination sites into habitat blocks, and clustered them. What we saw is in Puerto 05:08.720 --> 05:14.640 Reakins, Great Britain's and Humb Chinese, around TNF, you get 9, 8, and 3 clusters of 05:14.640 --> 05:20.800 habitat blocks. Looking at basic ancestry, HLA, you get 17, 15, 13. What does this mean? 05:20.800 --> 05:26.240 This means that genome is discretizable. So basically what we can do is we can go from a linear 05:26.240 --> 05:32.320 genome to a pan genome, but we can make this pan genome without making millions and millions 05:32.320 --> 05:37.360 of alignments, which I think is really important, because that is actually computationally heavy, 05:37.360 --> 05:42.720 even on a GPU. So basically what we're doing is approximate hashing coding for very large genomics 05:42.720 --> 05:48.960 by using these haploblocks. So we go basically from genome graphs to approximate hashing coding, 05:48.960 --> 05:54.560 then what we can do is label these haploblock clusters with snips, and then use them 05:55.200 --> 06:00.320 use convert them to binary strings, and then use these to come up with feature weights for different 06:00.320 --> 06:05.440 parts of the genome in a particular disease of phenotype. So that's something we're particularly 06:05.440 --> 06:11.120 excited about at this particular point. And I would like to note that a bunch of this was actually 06:11.120 --> 06:19.200 built in the Elixir biohagathon in Berlin, in November. So these are a bunch of things that 06:19.200 --> 06:23.680 Nvidia does, but I want to talk about actually knowledge graphs, and that's going to be sort of 06:23.840 --> 06:30.400 the third thing we talk about. So really thinking about the phenotype side of clustering of disease 06:30.400 --> 06:34.800 subtypes, it would be really nice if we could do disease clustering. Why? Because we've known 06:34.800 --> 06:39.520 all the ideologies of breast cancer for 30 years, or the four major ideologies I should say. 06:40.560 --> 06:45.040 But this is what a snowmed graph looks like right now. This is what a graph could look like. 06:45.040 --> 06:50.800 So there is a future there. And I think something I call pig rag could be part of the future 06:51.440 --> 06:56.400 in full disclosure. I didn't do any of the engineering on this project. I just made this graphic, 06:56.400 --> 07:04.960 and I did that with Google Gemini. So not all that much. But anyway, so yeah, I mean, most of you know 07:04.960 --> 07:10.720 what graph rag is, anybody not know what graph rag is? It's written up there. So anyway, 07:12.000 --> 07:18.640 great. So basically what we've done here to do Retrieval Augmented Generation for an LLM. 07:19.520 --> 07:25.120 A lot of people do that now. But basically what we've done is put a pie towards geometric 07:25.120 --> 07:31.280 gen in in front of the LLM. And you can retrieve a subgraph. You can see who's made what contribution, 07:31.280 --> 07:37.600 et cetera, et cetera. And then you can concatenate the embeddings from a gen and from an LLM. 07:37.600 --> 07:43.600 Why is that important? Because then you can also do document retrieval. So you can go ahead and 07:43.600 --> 07:48.400 merge unstructured data with structured data. But the really cool thing here is you can put 07:48.400 --> 07:53.120 out another knowledge graph so you can actually do validation as well as asking natural language 07:53.120 --> 07:58.720 queries. We're doing a lot of experiments where we look at you know one hop to five hop prompting 07:58.720 --> 08:04.480 and bin it and so on and so forth. Anyway, and it's you do need to do some fun tuning of the 08:04.480 --> 08:11.920 decoder. But we tested this out with Neo4j. You can check out this blog. And this was in the 08:12.000 --> 08:19.520 start prime data set. And you can actually you actually double the accuracy of retrieval 08:19.520 --> 08:27.200 augmented generation by putting a gen in front of the LLM. So that is all I wanted to tell you today. 08:27.200 --> 08:35.360 I think I managed to get all of this in nine minutes. They told me I had 10. So we should have five 08:35.360 --> 08:41.520 minutes for questions. Hopefully you guys have a bunch of questions. But so this is kind of my vision 08:41.520 --> 08:46.000 of the future that we can discretize the human and actually we're trying this with peanut 08:46.000 --> 08:53.120 and sorghum as well. Human and agricultural genomes, we can use those to develop models for 08:53.120 --> 09:01.600 phenotypes that we can actually represent in knowledge graphs to treat people. So treat people faster. 09:02.320 --> 09:04.640 So that's it. And happy to take any questions. 09:11.520 --> 09:25.120 All right. So somebody asked me when will Kuda be released as open source. So there's a full open 09:25.120 --> 09:32.320 source stack and I will say I have no idea. I work in a completely, I work in a very 09:33.280 --> 09:42.720 distant corner of in video where we really focus a lot on genomes. I certainly could ask 09:43.760 --> 09:48.800 whether they're going to tell me I don't know. But one thing I will say is that I work with a lot 09:48.800 --> 09:53.360 with the open source community and there's so many tools developed around Kuda that I think what we 09:53.360 --> 09:58.240 see is open source developers will develop first around Kuda just because it's so much easier 09:58.240 --> 10:04.240 and then moved to other sort of GPU support systems. Yeah. 10:28.640 --> 10:37.920 So I don't know how to answer a question based around low-level stuff, but we can discuss around 10:37.920 --> 10:44.720 beers, I think. But that said, so all of the paratrox modules, we spend a lot of time on 10:45.520 --> 10:51.680 reproducibility and matching to the original algorithms. And so people can check all of that out. 10:51.680 --> 10:57.200 And I mean, for example, we work a lot with the original development team. So for example, 10:57.200 --> 11:03.600 with the variant deep somatic, those sorts of things, to get matching with the GPU processes and 11:03.600 --> 11:08.400 CPU, we also do a lot of work with algorithm developers to try to get similar matching. 11:09.280 --> 11:11.760 And actually optimization as well. Yes. 11:27.200 --> 11:41.600 Yeah. So that's a great question. So somebody is saying, is there 11:41.600 --> 11:46.800 out of stations, server? One thing I would say is, I mean, in genomics, oh yeah, yeah. 11:46.800 --> 11:56.560 I'm stepping back and repeating the question. So if I am remembering the question correctly, 11:56.560 --> 12:04.720 it's, so are we allowing people to go ahead and verify that 12:26.640 --> 12:36.000 special. 12:36.000 --> 12:40.800 Are there signed to add a station artifacts? And I do not know the answer to that question. 12:40.800 --> 12:44.800 So, yep. Anybody else?