加速藥物發現和生命科學的模型開發和部署：來自 EMEA 地區的問答 [S73097a]
Accelerate Model Development and Deployment for Drug Discovery and Life Sciences: Q&A From the EMEA Region [S73097a]
https://register.nvidia.com/flow/nvidia/gtcs25/vap/page/vsessioncatalog/session/1739944897475001i2H5
更新可參考:
https://hackmd.io/rl9VBDohS1amPROCYPYlcA

Anthony Costa,NVIDIA數位生物學總監,Director, Digital Biology, NVIDIA

了解最新發展和下一代 NVIDIA 軟體，以加速藥物發現和生命科學。 BioNeMo 是 NVIDIA 的平台，用於加速最先進的化學、生物和基因組學基礎建模。 NVIDIA Parabricks 是我們最先進的基因組學平台，可加速我們對生物學的理解。結合整個 NVIDIA 全端運算平台，探索生物學和疾病的本質以及開發新藥和療法的技術從未如此快速和便捷。我們將討論兩個平台功能的最新進展、我們未來發展路線圖的各個方面以及利用我們的技術的成功專案的真實演示。
* 了解 NVIDIA BioNeMo 和 Parabricks 的功能
* 了解利用我們的生命科學產品的成功用例
* 深入了解 BioNeMo 和 Parabricks 的未來功能和機會

主題：生成式人工智慧 - 生物學 - 生成式人工智慧
產業領域：醫療保健與生命科學
3 月 21 日，星期五
下午 6:00 - 下午 6:50 中部標準時間

非常感謝大家一大早就來到這裡。我非常感激大家的參與。

Thank you very much for being here bright and early. I really appreciate it.

今天我將為大家帶來一些令人振奮的更新，主要是關於我們在數位生物學（Digital Biology）領域的產品方面所做的事情。我也會談到許多我們的關鍵合作夥伴，他們的支持讓這一切成為可能。所以我想從一張幻燈片開始，這張幻燈片你們可能在週二下午Kimberly Powell——我們的副總裁兼醫療保健部門總經理——的演講中看過。我之所以想展示這張幻燈片，是因為它清楚地呈現了我們在醫療保健和生命科學領域的起點，以及我們如何從明確的應用領域開始，逐步將一個原本屬於水平運算平台（Horizontal Computing Platform）的技術，垂直化（Verticalize）應用到醫療保健領域。

Evidence. And I’ll give you some exciting updates on the things that we’re doing in the digital biology space on the product side. And I’ll also talk a lot about our important partners who make all this possible as well. So I wanted to start with a slide that you may have seen in Kimberly Powell’s—our VP and GM of Healthcare—talk on Tuesday afternoon. The reason I really want to put this up here is that it clearly shows how we started in healthcare and life sciences in areas where there were clear applications first, and verticalize what is otherwise a horizontal computing platform for healthcare.

或許你們會感到驚訝，我們最初專注的領域是影像學（Imaging）和放射學（Radiology）。這些現在已經成為人工智慧（Artificial Intelligence）在醫療保健領域中非常基礎且典型的應用範例。但在過去十年中，我們不斷演進，發現人工智慧技術在各種不同領域中有著驚人的應用。今天，我們將特別聚焦在產品方面的兩個例子。你們會在幻燈片底部看到許多產品名稱，這些是我們生物學團隊需要打造的產品。但我之所以特別想展示這張幻燈片，是因為NVIDIA平台（NVIDIA Platform）和運算技術在生物學中的應用範圍正變得越來越廣泛。

It’s probably a surprise that we started in areas like imaging and radiology. These are now sort of bread-and-butter, canonical types of applications of artificial intelligence in healthcare. But as we have evolved over the last 10 years, we’ve found incredible applications for AI technology in all sorts of different areas. Today, we’re going to talk about two in particular on the product side. You’ll see lots of products listed at the bottom of the slide—these are products that we need to build with our biology team. But the reason I really wanted to put this up here is that the breadth of the NVIDIA platform and computing in general that is relevant in biology is getting broader and broader.

如果你們仔細看這張幻燈片——我可能無法用雷射筆清楚指向那麼遠的地方——但如果你從右邊開始看，你會看到推理加速平台（Inference Acceleration Platform），例如TensorRT。你們今天會看到這樣的例子，我們將其應用於人工智慧模型，特別是在生物學領域，實現了巨大的速度提升。這不僅適用於語言模型（Language Models）——你們可能通常會將TensorRT與這類技術聯繫起來——也適用於結構模型（Structure-Based Models）和藥物設計（Drug Design）。我們在BioNeMo上的許多工作，都是基於過去七、八年間的創新，學習如何擴展基礎模型（Foundation Models）。這建立在我們的Nemo平台之上，這個平台原本是用於大型語言模型（Large Language Models）和影像處理，但我們能夠將這些核心技術應用到生物學領域。

If you actually look at this slide—and I’m not going to be able to point all the way over there with a laser—but if you start from the right-hand side, you’ll see the inference acceleration platform, TensorRT. You’ll see examples of this today, where it was applied to AI models in biology to get huge speed-ups, both for language models—which is where you probably typically think about things like TensorRT—but also in structure-based models and design for drugs. We base a lot of the work that we do in BioNeMo, for example, on the innovations that have come from the last seven or eight years in learning how to scale foundation models. That’s based on our Nemo platform, which is applied to large language models and imaging. But we can inherit a lot of that core technology into what we do in the biology space.

例如像Rapids和GPU-MSA 這樣的技術，它們的應用範圍非常廣泛，而且在理解如何加速對齊（Accelerating Alignment）和加速動態數據處理（Accelerate Data Processing for Dynamics）方面變得越來越重要。這些技術不僅在生物學中有用，也在其他領域展現了它們的價值。


For example, technologies like Rapids and GPU—MSA are examples of technologies that have applications that are very broad, but increasingly important in understanding how to do a lot of accelerating alignment and accelerate data processing for Genomics.


我們將稍微談談這一點。或者根據你的經驗，實際上在使用像AlphaFold這樣的模型時，很多人在加速數據處理流程（Data Processing Pipeline），在你真正進入模型推理（Model Inference）本身之前就已經開始了。所以我不會一一介紹剩下的所有內容，但我認為可以公平地說，這裡展示的每一個技術模塊，我們在數位生物學（Digital Biology）領域中都以某種方式加以利用，這真的非常令人興奮，也很有趣。所以我想展示另一張幻燈片，這張你們可能之前看過。我很喜歡這張幻燈片，因為它告訴我們，我們如何開始生成理解生物學所需的基礎真相資訊（Ground Truth Information），以及我們今天如何大規模利用這些資訊，通過構建不同類型的人工智慧策略（AI Policy）來實現。

We’ll talk a little bit about that. Or in your experience actually using models like AlphaFold, a lot of people were doing that and accelerating the data processing pipeline before you actually get to the model inference itself. So I won’t go through all of the rest of these, but I think it’s fair to say that there isn’t a single box up here that we don’t take advantage of in one way or another in the digital biology space, which is super exciting and probably fun. So I wanna show this—it’s another slide that you might have seen before. And I like this slide a lot because what it teaches us is how we started in generating the core ground truth information that we need to understand biology, as well as how we’re taking advantage of that in a big way today with the types of AI policy we’re building.

如果你看幻燈片左下角，在2000年代，你會看到我們早期在基因定序（Sequencing）方面的工作，這是一場革命。這對我們現在正在進行的許多工作以及擴展規模有很大影響，因此工作負載（Workloads）和困難的數據生成（Data Generation）是生成數據的重要應用，例如用於冷凍電顯（Cryo-EM）之類的技術。這些技術正在成為結構研究（Structure Research）的一部分。現在，展望未來的應用，以及所有這些工具與人工智慧的整合，將能夠實現像蛋白質設計實驗室（Protein Labs）這樣的技術。所以我們擁有這個真正的基礎，用於生成和探究從我們實驗室中獲得的數據。但這對我們能夠將所有這些資訊整合起來，並建立有用的生物學表徵（Representations of Biology）至關重要，這些表徵可以幫助人們。

So if you look down on the bottom left, in the 2000s, you see our early work in sequencing—the revolution. That was CISPR for a lot of the work that we’re now doing and scaling up. So workloads and difficult data generation are hugely important applications for generating data, for things like Cryo-EM. And otherwise, it’s becoming structure research. And now, in the future of the application and sort of the integration of all of these tools with AI, it will enable things like protein labs. So we have this real foundation in generating and interrogating the data that comes from our labs. But it’s really fundamental to our ability to then take all of that information and build useful representations of biology that can help people.

因此，AlphaFold是我們第一個這樣的模型，我希望大家能留下來聽後續的專題演講（Talks），因為我們將會有來自一些團隊的精彩分享，他們會在這裡詳細談論過去、現在和未來的相關工作，我很確定這會很精彩。但隨著我們進入新的領域，AlphaFold利用了PDB資料庫（PDB Database），這個資料庫中有大量的結構數據。但現在你會看到許多模型整合了來自幻燈片底部展示的各種不同技術的資訊。所以這真的是個令人興奮的時刻。我會在接下來的半小時左右告訴你們我們如何支持這些工作。

And so AlphaFold was the first of those models. (5:14)I hope everyone sticks around for the upcoming talks because there will be some great presentations by some teams who are going to be here and talk a lot about their past, current, and future work—I’m sure. But then, as we’ve moved into new domains, AlphaFold took advantage of the PDB database that had lots and lots of structures. But now what you’re seeing is lots of models that integrate information that comes from all of the different technologies that you see at the bottom of this panel. So it’s really an exciting time. And I’ll tell you about how we support this over the course of the next half hour or so.

所以我們思考這個問題，我展示這張幻燈片主要是為了給大家一個組織性的原則（Organizing Principle）。在「實驗室循環」（Lab-in-the-Loop）這個主題上有許多變體。但這裡的核心是，我們從生成大量數據開始。你會看到很多人正在生成指數級增長的數據。這並不是因為我們已經做了很長時間就變得不重要，實際上它變得更加重要，無論是在現實世界的環境中，還是在合成數據（Synthetic Data）的環境中。我們需要訓練整合來自不同來源和模態（Modalities）資訊的模型。我們需要開發技術來讓這個過程快速、簡單且高效。然後我們可以使用這些模型來做出有用的預測，無論是作為矽基預測（In-Silico Predictions）的銀標準（Silver Standard），還是甚至在某些情況下生成聯邦加速數據（Federated Accelerator Data）。


(5:48)So we think about this, and I’m putting this slide up mostly as an organizing principle for everyone in their mind. There are lots of variations on this theme of lab-in-the-loop. But the fundamental piece here is that we start with generating a lot of data. You’re going to see a lot of people generating exponentially more data. And this is not getting less important simply because we have been doing this for a long time. It’s actually becoming more important, both in the real-world setting and even in the synthetic data setting. We then train models that integrate information from a variety of different sources and modalities. There’s technology that we need to build there to make that fast, easy, and performant. We then can use those models to make useful predictions, either as a silver standard in in-silico predictions or, in some cases, to even generate federated accelerator data.

我們通常會將我們構建的所有這些模型串聯起來，形成完整的工作流程（Workflows）。對吧？所以我們如何優化一個端到端的流程？這需要讓許多不同的模型能夠互相溝通。然後，一旦你完成了這些，並做出了一些有用的預測，這些預測將驅動你的決策，決定下一個實驗室實驗是什麼，或者你會拿這些資訊去進行一些實驗室實驗，獲取金標準標籤（Gold Standard Labels），然後將其反饋到你的訓練中。

And we usually take all of these types of models that we build and we string them together into entire workflows, right? So how do we optimize something that’s end-to-end? That takes a lot of different models that need to talk to each other in some way. And then, once you’ve done that and you’ve made some useful prediction, that will drive either your decision-making on what the next lab experiment is that you do, or you’ll take that information and get gold standard labels from some lab experiments, measure what you’re working on, and feed that back into your training.

對吧？所以這真的是個舞台設定的介紹，主要是讓你們感受到我們將要談論的幾乎所有內容都符合這個飛輪模式（Flywheel）。我認為可以合理地說，你可以看到這個領域的許多發展，跨越不同應用，特別是在微生物學（Microbiology）中，都可以融入這個模式。

Right? So this is really a stage-setting kind of talk just to give you a sense of how almost everything we’re gonna talk about fits into this flywheel. And I think it’s reasonable to say that you could get a lot of the development that happened in the space across applications, in microbiology, into this paradigm.

首先，我將為你們介紹我們在基因組學平台（Genomics Platform）上的更新。這是一個非常成熟的平台，我們在NVIDIA已經投入並發展了多年。我們非常幸運能夠長期致力於此並提升技術。我們的目標一直是加速高通量分析（High Throughput Analysis），在生命科學（Life Sciences）領域中實現這一點。我們已經通過一項名為Parabricks的技術實現了這一目標。

So first, I’m going to talk to you a little bit about updates that we have in our genomics platform. This is an extremely mature platform at NVIDIA. We’re very lucky to have committed to this and skilled it for a number of years here. And our goal has been to accelerate high throughput analysis in life sciences. And we’ve done that with a technology called Parabricks.

因此，這一直是我們在二級分析（Secondary Analysis）中所做工作的核心，例如加速人們在基因組學生命科學（Genomics Life Sciences）中所做的所有工作。越來越多地，我們現在看到的是，隨著我們生成所有這些數據，並能夠處理這些數據並使其變得有用，數據生成本身與我們附加的人工智慧技術（AI Technologies）之間的聯繫變得越來越普遍。

And so this has been the core of what we’ve done in secondary analysis, for example, to accelerate all the work that people do in genomics life sciences. And increasingly, what we’re seeing now is that as we generate all of this data, as we’re able to process all this data and make it useful, the connection between the generation of the data itself and the AI technologies that we attach to that data is becoming more and more prevalent.

所以，如果我們看看我們的基因組學生態系統（Genomics Ecosystem），我們在整個生態系統中跨越所有這些目標進行合作。我們參與了每一個基因組學儀器平台（Genetic Instrument Platform），並與這些合作夥伴密切合作。我們也在所有高內涵篩選（HCS, High-Content Screening）和雲端平台（Cloud Platforms）中，這裡列出的當然不是全部。然後，我們與產業和研究組織密切合作，這些組織是我們與合作夥伴共同部署技術的消費者。

So if we look at our genomics ecosystem, we work across all of these goals across our entire ecosystem. So we are in every single genetic instrument platform and work very closely with those partners. We are in all of the HCS and cloud platforms that you see here, and this is certainly not an exhaustive list. And then we partner very closely with industry and with research organizations that are consumers of the technologies that we work with our partners to deploy.

一個非常棒的例子是，我們與合作夥伴如何合作，例如我們的儀器合作。這是與Roche在SPX平台（SPX Platform）上宣布的工作，Kimberley Powell也有參與。如果有機會，我無法在這裡放一段影片，但如果你去YouTube搜尋並觀看這些技術運作的影片，第一次看到時真的很震撼。這真的很驚人，我們將基因定序（Gene Sequencing）中原本非常昂貴的活動，簡化成可以在一小時內完成多個基因組（Genomes）的定序，並保持合理的覆蓋率（Coverage）。

So one really incredible example of this, and an example of how we work with, for example, our instrument partners, is work that was announced on the SPX platform from Roche, and this was also mentioned by Kimberley Powell. I couldn’t put a video on this, but if you go to YouTube and just type this in and actually look at the videos of how this technology works, the first time I saw it, it’s really mind-blowing. It’s really amazing—we’ve taken what was an incredibly expensive activity in gene sequencing and reduced that into something that can be done with many genomes in an hour at reasonable coverage.

所以這是下一個前沿，在我們可以生成的基因定序（Sequencing）數據類型上。我們與Roche的合作整合了我們在Parabricks上的工作，以及其他加速技術，例如TensorRT。這是實現如此快速運作的平台。

And so this is the next frontier in the types of data that we can generate in sequencing. And our partnership with Roche is integrated work that we’re doing in Parabricks as well as other acceleration like TensorRT. It’s the platform to enable this to work as fast as it does.

Parabricks平台（Parabricks Platform）或基因組學平台（Genomics Platform）是用於基因組學（Genomics）的。我不會將我們限制在Parabricks上，這一點很快就會變得明顯。我認為，Parabricks是一個GPU加速的軟體套件（GPU-Accelerated Software Suite），用於基因組學（Genomics）。總體來說，如果你很熟悉這個平台，你會看到我們在二級分析（Secondary Analysis）中所做的許多工作。這是我們的核心技術。我們有非常快速的對齊（Alignment）。你可以在幻燈片左邊看到，我再次說明，我沒有雷射筆，但我指的是左邊，我們投入了大量資源，這是最新的進展，讓平台的演算法（Algorithm）速度快得多。

So the Parabricks platform or the genomics platform for genomics. And I’m not going to restrict ourselves to Parabricks. And that will be obvious, I think, in just a moment. Parabricks is a GPU-accelerated software suite for genomics. Overall, if you know the platform well, you will see lots of work that we’ve done in secondary analysis. This is our bread and butter here. We have incredibly fast alignment. You see on the left-hand side of the slide—and again, I’m indicating I don’t have a pointer when I don’t—but on the left-hand side, we’ve committed a lot of resources, and this is the most recent effort to make the platform’s algorithm much, much faster.

我們也有很棒的變異識別（Variant Calling）技術。你會看到這兩者如何互動，我們將它們整合成藍圖（Blueprints）。我們稍後會談論這一點，重要的是，正如我之前提到的，我們正在與NVIDIA更廣泛的技術集合進行互動，以實現人工智慧與動態（AI plus Dynamics）的革命，這是我們看到的結果。產業的數據處理（Data Processing）能力隨著Parabricks 4.5的成熟而提升。

We have incredible variant calling technologies in there as well. And you’ll see how the interaction of those two, we put together into blueprints. We’ll talk about it a little bit, and then importantly, as I mentioned before, we’re engaging with a much broader set of technologies in NVIDIA to enable the AI plus dynamics revolution that we’re seeing happen as a consequence. The data processing capabilities of the industry, as it matures, are enhanced in Parabricks 4.5.

就像我說的，我們專注於三個主要領域。在幻燈片左邊這裡，我列出了一些項目，我們會詳細介紹這些，但重點是核心性能改進，針對我們已有的技術，並在這裡中間部分將這些技術擴展到我們的Blackwell平台（Blackwell Platform）。我會向你展示一些關於Blackwell平台上更快、更便宜的Parabricks工作流程（Workflows）的內容，然後我會在這裡給你一個預覽，關於我們在涉及變異識別和深度學習（Variant Calling and Deep Learning）的工作流程中獲得的額外加速。

Like I said, there are three main areas that we focused on. On the left-hand side here, I have a number of items, and we’ll go through all of these in detail, but focusing on core performance improvements for the technologies that we already have, and extending those in the middle here to our Blackwell platform. So I’ll show you a little bit about that—both faster and cheaper on Blackwell for Parabricks for many workflows—and then I’ll give you a little preview here of additional acceleration that we get for workflows involving variant calling and deep learning.

這是我們與下一代Blackwell GPUs合作的例子，也就是我們所謂的RTX 6000系列。我在這裡展示了兩種不同的加速場景（Vignettes）。在中間，你可以看到。

So this is an example of work that we’re doing with our next generation of Blackwell GPUs, our RTX 6000 series. And so I show here two different vignettes of the acceleration that we’re getting. So in the middle, you see.

相較於L40S，我們上一代的GPU，在Smith-Waterman對齊（Smith-Waterman Alignment）上的表現，然後在右邊，你可以看到一個完整的遺傳基因組工作流程（Full Germline Workflow），在我們的案例中，這比以往快了135倍。現在在一個4x RTX 6000系統（4x RTX 6000 System）上，相較於一個96核心CPU實例（96-way CPU Instance），我們也獲得了相較於上一代平台的逐步改進。

Get versus L40S, our last generation of these GPUs for a Smith-Waterman alignment. And then on the right-hand side, you see a full germline workflow, which, in our case, is 135 times faster. Now on a 4x RTX 6000 system compared to a 96-way CPU instance, and then we get incremental improvements over the previous generation of the platform as well.

如果我們也看看我們在遺傳基因組（Germline）和變異識別（Variant Calling）方面的工作，相較於基準線（Baseline），相較於L4和L40S，我們在這個平台上的運行時間快了6倍，用於變異識別（Variant Calling），同時也有很大的改進，這些是最近發布的Parabricks核心功能的重大改進，並持續為投資者和加速提供支持。

If we look also at the work that we’re doing in germline and variant calling, we see compared to baseline on our L4 and L40S a 6 times faster run time on this platform for variant calling along with great improvements to core functionality in Parabricks that were released recently and continue for investors and accelerate.

我認為這真的很酷，這將引導我們進入一個關於我們在單細胞動態（Single Cell Dynamics）方面工作的討論。但我們有一個庫，實際上是與RAPIDS Single Cell合作開發的。這是一個很棒的庫，提供了巨大的速度提升。如果你熟悉Rapids，這是我們長期以來用於加速高數據生態系統（High Data Ecosystem）中數據科學（Data Science）的平台。你可能熟悉像cuML和cuDF這樣的庫，它們是Pandas和scikit-learn的直接替代品。

I think this is super cool, and this will lead us into a conversation around the work that we’re doing in single cell dynamics. But there’s a library that we have that is in collaboration with RAPIDS Single Cell. And it’s an incredible library, huge speed-ups. If you’re familiar with Rapids, which has been our platform for accelerating data science in the high data ecosystem for quite a long time. So you may be familiar with libraries like cuML and cuDF, which are drop-in replacements for things like Pandas and scikit-learn.

所以RAPIDS Single Cell利用了我們在這方面學到的很多經驗，來擴展和優化這些技術，用於單細胞生物學（Single Cell Biology）中常見的工作流程（Workflows）。這些是針對Scanpy中可用功能的GPU優化版本（GPU-Optimized Versions）。左邊的圖表展示了驚人的速度提升數字，例如UMAP的速度幾乎快了700倍，PCA的速度提升了近兩個數量級。這裡實際上是一個1100萬單細胞數據集（11 Million Single Cell Dataset），而不是100萬，這些數字真的很驚人。所以我們看到端到端的性能，運行時間表現接近200倍的提升，如右邊所示，還有我們長期以來通過與Rapids合作開發的更快的個別組件加速（Component Acceleration）。

So RAPIDS Single Cell takes advantage of a lot of what we’ve learned in doing that to extend and optimize those for common types of workflows in single cell biology. These are GPU-optimized versions of what’s available in Scanpy. And so the figure on the left just gives you incredible speedup numbers like UMAP of almost 700 times faster, nearly two orders of magnitude for PCA. Here is actually an 11 million single cell dataset, not a 1 million, but it’s incredible figures to look at. So we see end-to-end performance, run-time performance of approaching 200x, as you see on the right-hand side, with much faster individual component acceleration that we’ve been working on for some time through the work that we do with RAPIDS.

好的。所以在我們擴展Parabricks平台（Parabricks Platform）方面有幾件事情，以及一些非常重要的、甚至是革命性的性能改進，針對我們現有的工具，並持續投資於新工具。

Okay. So a few things there on what we’re doing in extending the Parabricks platform and some really important, in some cases, revolutionary performance improvements for the tools that we have and continuing to invest in new tools.

我想在這場演講中強調的一個主題是，我們在基因組學（Genomics）方面的工作如何與公司在其他領域的工作相關聯，特別是當我們將傳統上應用於其他領域的人工智慧技術（AI Technology），也許是生物學（Biology），應用到這個領域時。

One of the themes that I’d like to lean into for this talk is how the work that we’re doing in genomics relates to the work that we’re doing in all sorts of other areas in the company, especially as we’re applying AI technology classically applied to other areas, perhaps biology, into the space.

我們有很多關於這方面的公告，我在這場演講中無法全部涵蓋。但多模態（Multi-Modality）工作的交集，例如我們可能在MONAI中為病理學（Pathology）和高內涵影像（High-Content Imaging）進行的工作，以及我們在BioNeMo中將要做的工作。現在我將在幻燈片的下一部分深入探討，使用人工智慧模型（AI Models），我們正在利用加速這些Transformer模型（Transformer Models）的基礎知識，並將其擴展到我們需要處理和理解生物數據（Biological Data）的獨特架構中。

And so we have a lot of announcements around this, and I won’t be able to get to all of them in this talk. But the intersection, the multi-modality intersection of work that, for example, we might do in MONAI for imaging, for pathology, for high-content imaging, and the work that we’re going to do in BioNeMo. And now I’m going to lean into quite a lot in the next section of this slide with AI models where we’re taking the fundamentals of how we accelerate these transformer models and extending those into the uniqueness of the architectures that we need to process and understand biological data.

然後我們通過嵌入（Embedding）等方式整合這些經驗教訓，以實現對生物學（Biology）更全面和完整的理解。我之前提到過我們在RAPIDS Single Cell方面的工作，這對於處理所需的數據非常重要，以便將這一切整合起來。

And then we’re integrating those learnings through embedding as an example into a larger and more complete understanding of biology. And I mentioned before the work that we’re doing in RAPIDS Single Cell, which is really fundamental to processing the data that is required for putting this all together.

好的，接下來從我們的基因組學平台（Platform for Genomics）轉移到另一個主題，我將談論我們正在進行的許多工作。這是我們的生物學平台（Platform for Biology）。我們將首先從BioNeMo開始談起。

Okay, so moving on from our platform for genomics, I’m going to talk about a lot of work that we’re doing. And this platform for biology is in our platform for biology. So we’re going to talk about BioNeMo first and foremost to start with here.

這個圖表旨在讓你了解我們在人工智慧（AI）和生物學（Biology）方面的整個平台，我們稱之為BioNeMo。BioNeMo已經推出大約兩年了。我們已經用它完成了一些非常了不起的事情。就像這樣，也許我會再次播放這個動畫給大家看，這樣你們就可以看到這個很酷的動畫。

So this figure is intended to give you a sense of the entire platform that we have for AI and biology, and we call that BioNeMo. BioNeMo has been around for just about 2 years. And we’ve done some really remarkable things with this. And it’s like, maybe I’ll play this again for everybody. So you can just see the cool animation.

但有三個主要組成部分對大家來說很重要，需要先理解。首先，我們有庫（Libraries）和一個訓練框架（Training Framework）。這些是針對生物數據（Biological Data）獨特挑戰的加速庫（Accelerated Libraries）。你會在演講中看到幾個例子，但這並不是全部。這些庫被用於一個訓練框架中，這是我們對如何利用NVIDIA加速（NVIDIA Acceleration）的看法。我們在過去七八年學到了很多經驗教訓，正如我之前提到的，如何利用這些Transformer模型（Transformer Models），以及越來越新穎和更具異國情調的架構，這些架構對生物學問題特別有用，並在其中訓練人工智慧模型（AI Models）。

But there are three major components of this that are important for everyone to just get their head around. The first is that we have libraries and a training framework, right? And so these are accelerated libraries for the types of unique challenges that we have in biological data. So you’ll see a couple of examples of that, though not exhaustive, as I move through the talk. Those are used within a training framework, which is our perspective on how you take advantage of NVIDIA acceleration. A lot of work that we’ve learned over the last 7 or 8 years, as I mentioned earlier, and how you take these types of transformer models and increasingly novel and more exotic architectures that are particularly useful for problems in biology and train AI models there.

然後我們有一個有效的方式來部署這些模型。我們是NVIDIA微服務生態系統（NVIDIA’s Microservice Ecosystem）的重要成員。我們已經推出了許多模型（Models），涵蓋了各種不同的使用案例。其中一些是你可以在框架中訓練的模型類型，目前主要是基於序列的模型（Sequence-Based Models）。但我們也有針對蛋白質結構（Protein Structure）、預測（Prediction）、蛋白質生成（Protein Generation）和設計工作流程（Design Workflows）的模型，多組學複雜工作流程（Multi-Omics Complex Workflows）、姿態預測（Pose Prediction），例如分子對接（Docking）之類的技術。

Then we have an efficient way to deploy those models. And so we are very much a strong member of NVIDIA’s microservice ecosystem. We have launched many, many models across a variety of different types of use cases. Some of those are the types of models that you can train in the framework, which today are primarily sequence-based models. But we also have models across protein structure, prediction, protein generation, and design workflows, multi-omics complex workflows, pose prediction, things like docking.

我會告訴你我們的庫和框架如何教我們如何讓整個生態系統更快、更好、更高效。然後當我們轉到右邊的藍圖（Blueprints）時，單個模型（Models）可能很有用。它們在大規模應用時通常非常有用，特別是當你從這些遊戲中生成大量數據時。但現實世界的工作流程往往需要整合許多不同的模型和技術。

And I’ll tell you a bit about how our libraries and our framework teach us how we can make that entire ecosystem much better, faster, and more performant. And then as we move over to blueprints on the right-hand side, the individual models can be useful. They can be useful at scale. They’re often very useful when you’re generating lots of data from these games. But real-world workflows tend to integrate lots of different models and different technologies together.

所以我們將所有這些工作流程整合成藍圖（Blueprints），我們會提供這些藍圖，並以多種不同方式提供。我稍後會告訴你這一點。我們正在跨越從基因組學（Genomics）到化學（Chemistry）再到生物學（Biology）的所有領域進行這項工作。並且你會看到一些與合作夥伴的工作，這些工作真正將這一切擴展到推理（Reasoning）和語言模型（Language Model）類型的操作，我稍後會提到。

And so we integrate all of these workflows into blueprints, which we make available, and we make those available in lots of different ways. I’ll tell you about that in a little bit. And we’re doing this across everything from genomics to chemistry to biology. And you get some work with some partners that really extend this into reasoning and language model type operations, which I’ll mention just about.

那麼BioNeMo框架（BioNeMo Framework）是什麼？BioNeMo框架擴展並改進了訓練超大規模模型（Large-Scale Models）的傳統。所以你如何處理一個人工智慧模型（AI Model），如何將其分解到一個大型集群（Large Cluster）上？

So what is the BioNeMo framework? The BioNeMo framework extends and improves on a legacy of training very large-scale models. So how do you take an AI model, how do you break that up across a large cluster?

你如何訓練這個模型？我們需要為生物、化學和基因組數據（Biological, Chemical, and Genomic Data）做一些特別的事情，這與我們在大型語言模型（Large Language Models）或影像空間（Imaging Space）中的做法不同。我們可以訓練高達數十億參數規模的模型。我們最近與Arc Institute合作發布了Evo-2（Arc Institute Evo-2），這是一個新的架構，擴展到400億參數的人工智慧模型（AI Model），這在生物學中是一個非常大的模型。

How do you then ask, how do you train that model? And one of the specific things that we need to do for biological, chemical, and genomic data that are different from how we do things in, for example, the large language or imaging space. And so we can train models up to many billion parameters scale. We recently released work with the Arc Institute Evo-2, which is a novel architecture that extends up to 40 billion parameters for an AI model, which is a very large model in biology.

我們為生物學應用（Biological Applications）和所需的要求（Types of Requirements Needed）構建了定制架構（Custom Architectures）。這包括支持長序列（Long Sequences）和對非常長序列的推理（Reasoning）。因此，非常大的輸入上下文（Input Context）是其中一個重要且獨特的方面，當然不是唯一的一點，這與對生物數據（Biological Data）的推理有關。

We build custom architectures bespoke for biological applications and the types of requirements needed. These include support for long sequences and reasoning along very, very long sequences. So very large input context is one of the things, certainly not the only thing, that is important and unique about the reasoning over biological data.

我們以模組化的方式（Modular Fashion）來處理這一點，這意味著我們認識到這是一個快速發展的領域。我們知道我們需要找到通用的方法（Generalize），這樣我們就可以加速所有正在構建這些模型的人，而不僅僅是專注於特定的模型架構（Model Architecture）。所以每次我們發現如何讓某個部分更快時，我們都會試著找出如何在適當的層次上公開這些改進，讓每個構建模型的人都能利用這些優勢。我會給你展示幾個例子，結合我們如何部署這些工具，我們有許多方式可以做到這一點，包括開源（Open Source）可用、PyPI可用（PyPI Available）、作為容器（Containers）可用，以及作為企業版本（Enterprise Versions）提供。

We approach this in a modular fashion, which means that we recognize that this is a very fast-moving field. We know that we need to find things that generalize, so that we can accelerate everyone who is building these types of models and not just commit to a specific model architecture. And so we try to find, every time we figure out how to make something faster, we try to figure out how to expose that at the right level so that everyone building models can take advantage of that. And I’ll show you a couple of examples of that in concert with how we deploy these tools, which we have many ways of doing—both open source available, PyPI available, as containers, and available as enterprise versions of this as well.

我們也非常注重社群驅動（Community-Driven）。這是一個研發（R&D）領域，我們與在場的許多人合作，我認識並與這裡的許多人一起工作，他們今天來聽這場演講。我們從你們所有人那裡獲得反饋，這些反饋告訴我們未來需要解決的最重要的問題是什麼。部分做法是我們在公開環境中開發（Developing in the Open），這樣你們可以貢獻意見，告訴我們下一步需要做什麼，我們可以以合作的方式做到這一點。我們正在逐步建立工作小組（Working Groups），圍繞我們在這個框架中進行的一些工作，確保我們能正式滿足這些需求。

We’re also quite community-driven. This is an R&D field, and we work with many people in the room, many people I know and work with here that are here for the talk. And we get feedback from all of you that teaches us what direction we need to go and what’s going to be the most important problem that we can solve for you in the future. And part of that means that we’re developing this in the open so that you could contribute, you could tell us exactly what we need to do next, and we could do that in a collaborative way. And we’re gradually building working groups around some of the work that we’re doing in this framework to make sure that we’re meeting those needs formally in the future.

好的，那這實際上意味著什麼？讓我們舉一個非常簡單的模型例子。

Okay, so what does this actually mean? So let’s take just a very straightforward example of a model.

我相信大家都知道ESM-2（ESM-2）。在左邊，你有6.5億參數的變體（650,000,000 Parameter Variant），在右邊，你有30億參數的變體（3,000,000,000 Parameter Variant）。你可以看到非常不錯的結果。在不同模型大小範圍內，ESM-2 在我們的A100和H100系統（A100 and H100 Systems）上獲得了相當的加速，大約是1.7倍到1.8倍的加速（1.7x to 1.8x Speedup）在A100上，相較於標準實現（Standard Implementation），在H100上則超過2倍的加速，這是你可能從基準計算（Computation of the Baseline）中獲得的原始表現。

I’m sure everyone knows ESM-2. On the left-hand side, you have the 650,000,000 parameter variant. On the right-hand side, you have the 3,000,000,000 parameter variant. And what you see is quite nice. An equivalent speedup across a range of model sizes for ESM-2 on our A100 and H100 systems, so roughly a 1.7x to 1.8x speedup on A100, and over 2x speedup on H100 compared to the standard implementation that you might get from the original computation of the baseline that we can interpret here.

我們的擴展性也非常好。左邊的圖表是標準化的（Normalized），針對單個用戶節點（User Nodes）。我們獲得了非常高的性能擴展（Performance Scaling），基本上是完美的擴展（Perfect Scaling）。BioNeMo相較於你可能從標準實現中獲得的表現有顯著提升。

We also scale incredibly well. So the figure on the left is normalized to individual user nodes. And we get incredibly high performance scaling, basically perfect scaling. And BioNeMo compared to what you might get from the standard implementation that you can get up.

我們也實現了各種預訓練優化驅動（Pre-Training Optimization Drivers）。這些是更大的改進（Larger Improvements）。這些是功能上的改進（Functional Improvements），無論是在性能（Performance）還是能力（Capabilities）方面，當我們在一個真正穩健且可重複的平台上進行工程設計時，例如BioNeMo框架（BioNeMo Framework），這是基於多年投資和理解如何擴展這些模型的傳統。

And we also implement all sorts of pre-training optimization drivers. So these are larger improvements. These are really functional improvements both in terms of performance and in terms of capabilities that you get when we engineer something on a really robust and repeatable platform, like the BioNeMo framework, which is based on a legacy of many, many years of investment and understanding how to scale these models.

我們也擴展了我們在Transformer空間（Transformer Space）或人工智慧模型空間（AI Model Space）中針對序列（Sequences）所能做的事情。這是一個架構的例子，最初可能是從Evo-1開始使用的，這是Arc Institute去年發布的模型，去年是的，然後是我們今年與Arc Institute合作發布的Evo-2，這是非常有趣的合作，讓我們學到了很多關於如何在BioNeMo框架（BioNeMo Framework）中構建新架構的知識。這真的很重要，因為這些架構對於我們在生物學（Biology）中需要做的事情特別有用，在這種情況下，就是對非常長的序列（Very Long Sequences）進行推理（Reasoning）。

We also then extend what we can do in the transformer space or in the AI model space for sequences. And so this is an example of an architecture straight, you know, starting. It was perhaps first used in Evo-1, a model that was released by the Arc Institute last year, yeah last year, and Evo-2, which we released together this year as a collaboration between us, which is extremely fun and taught us a lot about how we build new architectures into the BioNeMo framework. And this is really important because these architectures are particularly useful for what we need to do in biology, which is reasoning, in this case, across very, very long sequences.

你可以在這裡看到一些性能基準（Performance Benchmarks），在A100、H100和B200（A100, H100, and B200）上，我們在BioNeMo框架（BioNeMo Framework）中獲得了革命性的性能改進（Evolutionary Performance Improvements），再次展現了極好的擴展性（Scaling）和非常好的性能，特別是當我們處理非常長的上下文（Very Long Context）時。這是我們與在場的幾位作者一起發布的，今年早些時候。我們與Arc Institute以及許多學術和產業合作者一起，歷史上也參與了我們發布的論文。

And you see here a number of performance benchmarks on A100, H100, and B200, where we’re getting evolutionary performance improvements in the BioNeMo framework again, with extremely good scaling and really good performance as we move out to a very, very long context. Like so, this was published together with a few authors in the room here today, earlier this year. We have historically worked between the Arc Institute and ourselves, in addition to a number of academic and industry collaborators that were also on the paper that we published.

它完全是開源的（Open Source）。對吧？所以現在社群的結果是，你在BioNeMo中有一個真正可重複、穩健的平台。你可以去獲取並訓練7億和40億參數的檢查點（Train Checkpoints）。你有一個平台來訓練間隙（Train the Gaps）、微調（Fine-Tune），然後最終部署這些模型。我們也給你一個例子，在一個可部署的NIM（Deployable NIM）中，它也使用了Evo模型，並帶有我們與Arc Institute設計的API。

And it’s available completely open source, right? And so the consequence for the community is now you have a really repeatable, robust platform in BioNeMo. You can go and get to train checkpoints at 7 and 40 billion parameters. And you have a platform to train the gaps, fine-tune, and then eventually deploy these models. And we give you an example of that in a deployable NIM that takes the Evo model as well with an API that we designed with the Arc Institute.

很棒的新架構在BioNeMo之上，這讓我們能夠做一些以前無法做到的事情。簡單介紹一下我們如何看待BioNeMo中的模組（Modules）。這是兩個例子，但並非全部。

So great new architectures on top of BioNeMo that enable us to do things that we otherwise couldn’t do before. Briefly getting into the way that we think about the modules in BioNeMo, these are two examples and they are not exhaustive.

但我們想要做的是，就像我之前說的，我們想要找到常見的加速和痛點（Common Acceleration and Pain Points），我們可以為社群貢獻這些，並讓這些對開發模型的人來說是可用的（Accessible）。所以其中一個是我們最近納入BioNeMo框架的套件（Package），但也可以獨立安裝。你可以直接用pip安裝（Pip Install That）。這就是為什麼我在右上角放了那個小東西。

But what we want to do is, like I said before, we want to find the common acceleration and pain points that we can contribute to for the community and make those accessible for people developing models on top of. And so one of those is a package that we recently included in the BioNeMo framework, but was also independently installable. You could just pip install that. And that’s why I had that little thing on the top right.

我從PyPI網站（PyPI Website）中提取了一個套件，它允許你使用Interpolate庫（Interpolate Library），這對於構建和探索生物學（Biology）中的架構擴散和流匹配模型（Architectural Diffusion and Flow-Matching Models）特別有用。這些模型在文本到圖像生成空間（Text-to-Image Generative Space）中很有名，但它們在序列和結構設計（Sequence and Structure Design）中也非常重要。你會在演講後面看到很多模型。到我結束這裡時，它們正在利用潛在擴散（Latent Diffusion）和擴散（Diffusion）類型的方法來進行分子設計（Molecular Design）。我們在這裡支持連續和離散模型（Continuous and Discrete Models）。

I pulled from the PyPI website, which allows you with an Interpolate library, which is particularly useful for building and exploring architectural diffusion and flow-matching models in biology. So these are famous in the text-to-image generative space, but they’re also really important in sequence and structure design. You’ll see a lot of models later in the talk. By the time I finish here, they are taking advantage of both latent diffusion and diffusion types of approaches for molecular design. So we support continuous and discrete models here.

所以你可以在框架中使用這個，你也可以使用這個獨立模組（Independent Module）來設計這些模型。我們也構建了很棒的數據加載器（Data Loaders）。這是一個發布的庫，我們已經在這個領域投資了一段時間，類似於你在RAPIDS Single Cell中看到的東西。所以這個名為BioNeMo SCDL的庫，特別適用於加載單細胞數據（Single Cell Data），以便在BioNeMo中訓練人工智慧模型（AI Models）。

And so you can use this in the framework. You can also use this independent module for designing these. We also build incredible data loaders. And so this is a library released. It’s actually where we’ve been investing in a space for some time, similar to what you’ve seen in things like RAPIDS Single Cell. And so the library called BioNeMo SCDL, which is particularly useful for loading single cell data for the purpose of training AI models in BioNeMo.

同樣，它是可以獨立安裝的（Independently Installable），作為一個獨立的、非常高效的記憶體庫（Memory-Efficient Library），基本上與AnnData API有一對一的映射（1:1 Map），並且速度提升了大約5倍（Five X Speedup）。所以你可以看到如何做到這一點的例子，對於那些熟悉這一點的人來說，這看起來會非常類似於你已經做過的事情，但我們已經在BioNeMo中實現了這一點，類似於我們已經做過的。

And so again, it’s independently installable, but as a standalone, a very nice memory-efficient library for doing this with basically a 1:1 map to the AnnData API and a speedup of about five X. So you can see an example of how to do this. For those of you that are familiar with this, this will look very, very similar to what you’ve already done, but we’ve implemented this in BioNeMo sort of similar to what we’ve already done both at this point.

因此，這大幅加快了進度。當你在思考如何為這些模型提供資料時，這一點非常重要，不論是在訓練階段，還是在大規模應用時。我們提供了許多入門教學，我們線上有大量的文件資料。有很多方法可以獲取這些資源，我會在下一頁投影片中向你展示。但請務必上網查看我們的說明文件，看看教學內容，去了解我們正在做的事情。你可以即時看到我們工程團隊的工作進展。這是公開且透明的。我們致力於打造這個框架(framework)，也請你們來和我們交流，提供意見反饋，告訴我們需要做什麼，才能讓這個框架變得更好。

So it speeds it up considerably. What's important as you're trying to figure out how to feed these models, uh, as you're training, but then also as you're operating at a large scale. So lots of tutorials to get started. We have a ton of documentation online. Um, there are many ways to get access to this; I will show you in the next slide. But please go ahead and take a look online in our documentation. Look at tutorials, go have a look at what we're doing. You can see what our engineering team is doing in real time. Uh, that's open and transparent. We are about the framework that we're building, um, and please come talk to us and give us feedback and tell us what we need to do, um, to make this framework even better.

所以，你可以用各種不同的方式來使用這些資源。我提到過開源(open source)可以幫得上忙，這是大家所期望的。我們以非常寬鬆的許可證，例如Apache許可證(Apache license)，來開發這個開源項目。我們還有一個打包在NGC上的開源版本。這非常方便，因為我們為你處理了所有的依賴堆疊(dependency stack)，你只需要下載容器(container)，一切就能運作。非常推薦這樣做。這不會帶來額外的費用，你可以直接下載並使用。如果你在使用這個框架時需要企業級支援，我們也提供這種服務，你可以聯繫我們的銷售團隊，他們會與我們合作，為你提供與這個框架及容器相關的完整商業支援。

Um, so you can consume things in a variety of different ways. I mentioned the open source can help out, um, which is as everyone expected to be, um, developed with a very permissive license, an Apache license. We also have an open source version that is packaged up on NGC. Um, this is very convenient because we take care of the entire dependency stack for you, and you just run the container and everything works. Um, so highly recommended. There's, uh, no additional costs associated with this. You can go download it and use it. And then if you do need enterprise support for the framework, we do provide that, and you can get in touch with our sales team, who will come through us, and we can give you, um, we can give you full support in business for the framework and the container that we built here.

好，接下來是倒數第二部分。現在我們為你提供了一個框架(framework)，讓你可以打造這類型的模型。那麼，我們如何實際調整這些模型呢？我們又該如何在大規模應用中充分利用這些模型呢？

Okay, um, so the second last section here is: okay, so now we have a framework for you to build these types of models. What do we actually do to adapt models? And how do we take advantage of these models at large scale?

我們參與了影像推論(video inference)和微服務(microservice)領域的工作，例如在感測器啟動(sensors inception)方面，我們實際上是這個領域中最活躍的發布者之一。因此，這些影像相關的服務真的是全棧解決方案(full stack solution)。你會得到一個優化過的模型，包含所有依賴項，全部打包在一個你可以下載並部署的單一容器(container)中。這對那個模型來說是合理的。我們花了很多心思來優化這些模型，例如，像02或04 SM這樣的模型提供了多種選擇。你稍後會看到我給出的具體性能數據。我們確保模型能夠運作，並且我們定義了API，讓它易於大規模部署並整合到你的環境中。我們也有相關範例，看起來就像是最佳實踐(the Prince)。

So we’ve been a part of what we’re doing in the video inference like microservice space, sensors inception, or one of the most prolific publishers in this space actually, um and um and so the image of respect and services really are the full stack solution so that you have an optimized model with all of those dependencies, all as a single container that you can download and deploy. And that makes sense for that model. Right, so we take lots of pains to optimize the models, so for example, of 02 or 4 SM offer a number of miles. You’ll see later where I have very specific performance figures. We ensure that the model works. The model is, um, and that we define API and make it easy to deploy at scale and then integrate into your world. And we have examples of those. Of course, look like it. Those are called the Prince.

好，現在快速回到動態領域(dynamic space)。這是基因學(genetics)開發分支(limbs)和藍圖(blueprints)的新動向。我們這週發布了首批兩個NIMs，針對動態領域。第一個是對齊(alignment)微服務(microservice)，它易於部署，用於標準對齊。我們還發布了變體(variant)命名的版本，用於變體識別(variant calling)。這些是可以下載的NIMs，你可以在我們的網站上找到。所有你熟悉並喜愛的推論微服務(inference microservice)生態系統的特點，這裡都適用。我們還將這些整合到我們的首個基因分析(genetics analysis)和單細胞分析(single cell analysis)參考中。左邊你可以看到一個範例，它整合了今天已經談到的許多技術，針對單細胞，包括快速單細胞分析(rapid single cell)，成為完整的影像藍圖(blueprints)。右邊也有類似的故事，針對變體的段落，用於端到端分析(end-to-end analysis)。當你思考這些藍圖時，關鍵要記住它們對我們來說是參考。它們教你如何將這些加速的微服務串聯起來，建構出有用的工作流程(workflow)。但它們也很有彈性，你可以加入其他容器或其他技術，插入這些工作流程，打造出最適合你的藍圖。

Okay, so um, just moving very quickly back into the dynamic space. Um, this is a new motion for genetics um in developing limbs and blueprints. Uh, we released our first two NIMs uh for dynamics uh this week. Um, the first is an alignment name uh microservice that is easy to deploy for standard alignment. Um, and we’ve released a variant name uh for variant calling. Um, so these are uh downloadable NIMs. You can get them on our website all. All the same uh types of things that you come to know and love about the inference microservice uh ecosystem applies here. And then we’ve also integrated these into our first reference for uh genetics analysis and for single cell analysis. Uh, so the left-hand side, you see an example that integrates a lot of technologies that we’ve already talked about here today uh for a single cell that includes rapid single cell um uh into an entire video uh blueprints. Um, and we have a similar uh story on the right um with the paragraphs of the variant uh for end-to-end analysis. And the critical thing to keep in mind when you’re thinking about these blueprints is different for us. Are references. What they do is they teach you how to take all of these accelerated microservices and build them together into a useful workflow. But they’re also flexible uh in the sense that you can take other containers, other technologies that plug them into these workflows so that you can string together exactly the blueprint that makes sense for you.

在告訴你我們與一些合作夥伴合作的範例之前，我先簡單提一下。你還有幾天時間可以分享這些QR碼，從我們的合作夥伴那裡獲得一些免費點數(credits)。你可以用這些點數部署單細胞分析藍圖(single cell analysis blueprint)和動態分析藍圖(dynamics analysis blueprint)，並獲得計算資源，讓你體驗我們提供的技術。這項活動到3月22日結束，所以請把握機會。

Very briefly before I tell you about some of those uh examples that we work on with a few partners, I just wanna plug this. You have a couple of days left um uh to share these QR codes and get some free credits uh from our partner so um where you can deploy the single cell analysis blueprint and the dynamics analysis blueprint um and get access to the computer so that you can play around with the technology that we could. Um, so uh I think that this ends, I guess, by March 22nd, um and so please do it.

在NIM生態系統中，我們還在結構生物學(structural biology)和小分子設計(small molecule design)方面發布了很多內容。接下來我會講得更技術性一點，告訴你我們如何讓這些模型更高效，甚至在某些情況下性能更好。我們研究了像AlphaFold這樣非常知名的模型，還有其他可能正逐漸出名的模型。這些模型中，有些人可能聽過像DiffDock這樣的，用於完美結構生成(protein structure generation)。還有新的模型，我稍後會介紹。

So in the NIM ecosystem, we’ve also published a lot of things in structural biology, in uh in small molecule design. And so here’s where I’m going to get a little bit more technical and just tell you about some of the uh some of the incredible things that we can do uh to make these models much more efficient, in some cases, much more performing as well. Uh, so we worked on models like AlphaFold, obviously very famous at this point, uh other models that are probably, you know, getting to that level of fame, things like that. And all is that uh some people might know here like DiffDock uh for protein structure generation and in general, what’s the new model which I will tell you about in a second?

重要的是，這些都以NIMs的形式提供，作為影像輸入(video inputs)和微服務(microservices)。但在每個案例中，我們所做的是拿到了模型架構(model architecture)，並學會如何讓它變得更高效。對比一對一，你會得到完全相同的結果。例如，對於AlphaFold，我們現在有一個NIM，在單一GPU推論(single GPU inference)上速度提升了五倍，比你可能公開取得的標準AlphaFold實作還要快。我們在DiffDock上也做了驚人的工作。我們甚至將這些成果整合成一個名為Covariance的函式庫(library)，用來加速各種類型的等變模型架構(equivariant model architectures)。我們在DiffDock上達到了七倍的速度提升。我們還創造了新的資料集(data sets)，訓練下一代模型(next generation models)，同時提升這些模型的準確度和性能。我們的擴散模型(diffusion model)速度大約快了兩倍。Grok是下一代模型，雖然在性能比較上不太一樣。但相關的完整論文將在今年晚些時候於ICLR上展示，達到了基於片段(fragment-based)的最新準確度(state-of-the-art accuracy)。

The important thing is these are all available as NIMs, as video inputs, microservices. Um, but in every single case, what we’ve done is we’ve taken that model architecture and we’ve learned how to make it way more efficient. Right? Uh, 1 to 1, you get exactly the same results. Um, but for example, an AlphaFold, uh, we have now a NIM that is five times faster, um, uh, for a single GPU inference, um, uh, than the standard implementation of AlphaFold that you might get publicly and DiffDock. We’ve done incredible work here. And we’ve actually calculated this into a library called Covariance to accelerate all types of equivariant model architectures. And we’re, you know, pushing 7 times faster in the DiffDock. And there we’ve actually invented new datasets, trained next-generation models, and improved the accuracy and performance of these models as well. Our diffusion were about 2 times faster in that model. Grok uh is the next-generation model. Um, and so comparing quite the same way in terms of its uh in terms of performance. Uh, but this entire paper out this is going to be featured. ICLR later this year, um, and um, and is achieving state-of-the-art accuracy for a fragment-based approach in a second.

好，我想談談三個我們非常驕傲的合作整合案例，我們與這些夥伴密切合作。首先，我們與BioVia緊密合作，部署了一個Moment 16的NIM，作為左上角的展示。這裡的重點在於，BioVia實作的是一種「自選預測(choose your own oracle)」的方法。如果你從事分子設計(molecular design)，我們有一個Moment和優化器(optimizer)，讓你可以引入外部預測(external oracle)，也就是你自己做的預測，來引導模型的演進，以及生成新的候選分子(candidate molecules)，然後將結果回饋給Moment。你可以在BioVia平台上靈活地做到這一點。我們很快將Moment NIM整合成一個端到端藍圖(end-to-end blueprint)，這不是我們公開發布的那個，而是利用相同技術客製化的藍圖，讓他們的客戶能快速應用，並在他們的平台上實現生成化學設計(generative chemistry design)。我們還與科學城(City of Sciences)合作。

Okay, um, so I wanna just talk about three partner integrations uh that are really that we’re extremely proud of and that we’re working very closely. Um, so we work very closely with BioVia um to uh deploy a um uh a Moment 16 as a NIM here on the top left. And the important thing here is that what BioVia implements um is a choose-your-own-oracle kind of approach. Um, if you’re doing molecular design, uh, so we have a Moment. We have the optimizer uh that allows you to sort of take an external oracle, so some prediction that you make that guides the evolution of the model and how it generates new candidate molecules, um, and feed that back into Moment. Um, but you can do that in a flexible way on the BioVia platform. And so they integrated very quickly, uh, our Moment NIM into an end-to-end blueprint, a customized blueprint, not the one that we published, the one that uh takes advantage of the same technologies in order to quickly stand this out for their customers and enable this generative chemistry design on their platform. Um, we’ve also worked with um uh with the City of Sciences.

重要的是，這些都以NIMs的形式提供，作為影像輸入(video inputs)和微服務(microservices)。但在每個案例中，我們所做的是拿到了模型架構(model architecture)，並學會如何讓它變得更高效。對比一對一，你會得到完全相同的結果。例如，對於AlphaFold，我們現在有一個NIM，在單一GPU推論(single GPU inference)上速度提升了五倍，比你可能公開取得的標準AlphaFold實作還要快。我們在DiffDock上也做了驚人的工作。我們甚至將這些成果整合成一個名為Covariance的函式庫(library)，用來加速各種類型的等變模型架構(equivariant model architectures)。我們在DiffDock上達到了七倍的速度提升。我們還創造了新的資料集(data sets)，訓練下一代模型(next generation models)，同時提升這些模型的準確度和性能。我們的擴散模型(diffusion model)速度大約快了兩倍。Grok是下一代模型，雖然在性能比較上不太一樣。但相關的完整論文將在今年晚些時候於ICLR上展示，達到了基於片段(fragment-based)的最新準確度(state-of-the-art accuracy)。

The important thing is these are all available as NIMs, as video inputs, microservices. Um, but in every single case, what we’ve done is we’ve taken that model architecture and we’ve learned how to make it way more efficient. Right? Uh, 1 to 1, you get exactly the same results. Um, but for example, an AlphaFold, uh, we have now a NIM that is five times faster, um, uh, for a single GPU inference, um, uh, than the standard implementation of AlphaFold that you might get publicly and DiffDock. We’ve done incredible work here. And we’ve actually calculated this into a library called Covariance to accelerate all types of equivariant model architectures. And we’re, you know, pushing 7 times faster in the DiffDock. And there we’ve actually invented new datasets, trained next-generation models, and improved the accuracy and performance of these models as well. Our diffusion were about 2 times faster in that model. Grok uh is the next-generation model. Um, and so comparing quite the same way in terms of its uh in terms of performance. Uh, but this entire paper out this is going to be featured. ICLR later this year, um, and um, and is achieving state-of-the-art accuracy for a fragment-based approach in a second.

好，我想談談三個我們非常驕傲的合作整合案例，我們與這些夥伴密切合作。首先，我們與BioVia緊密合作，部署了一個Moment 16的NIM，作為左上角的展示。這裡的重點在於，BioVia實作的是一種「自選預測(choose your own oracle)」的方法。如果你從事分子設計(molecular design)，我們有一個Moment和優化器(optimizer)，讓你可以引入外部預測(external oracle)，也就是你自己做的預測，來引導模型的演進，以及生成新的候選分子(candidate molecules)，然後將結果回饋給Moment。你可以在BioVia平台上靈活地做到這一點。我們很快將Moment NIM整合成一個端到端藍圖(end-to-end blueprint)，這不是我們公開發布的那個，而是利用相同技術客製化的藍圖，讓他們的客戶能快速應用，並在他們的平台上實現生成化學設計(generative chemistry design)。我們還與科學城(City of Sciences)合作。

Okay, um, so I wanna just talk about three partner integrations uh that are really that we’re extremely proud of and that we’re working very closely. Um, so we work very closely with BioVia um to uh deploy a um uh a Moment 16 as a NIM here on the top left. And the important thing here is that what BioVia implements um is a choose-your-own-oracle kind of approach. Um, if you’re doing molecular design, uh, so we have a Moment. We have the optimizer uh that allows you to sort of take an external oracle, so some prediction that you make that guides the evolution of the model and how it generates new candidate molecules, um, and feed that back into Moment. Um, but you can do that in a flexible way on the BioVia platform. And so they integrated very quickly, uh, our Moment NIM into an end-to-end blueprint, a customized blueprint, not the one that we published, the one that uh takes advantage of the same technologies in order to quickly stand this out for their customers and enable this generative chemistry design on their platform. Um, we’ve also worked with um uh with the City of Sciences.

這是一個非常酷、非常驚艷的實作，它將許多通常難以取得的技術整合在一起，並在一個完全無程式碼(no-code)的平台上提供給使用者。他們將Moment、AlphaFold和DiffDock整合到他們的電子實驗筆記本(electronic lab notebook)解決方案中。這樣的結果是，當你使用自己的資料時，你可以立即使用化學設計(chemical design)、結構預測(structure prediction)和其他這個藍圖(blueprint)中實現的工作流程(workflows)。最後，但絕非不重要，我們與Cadence和OpenEye密切合作，在他們的Orion分子設計平台(molecular design platform)上進行整合。他們利用了我們最近發布的一些NIMs，其中最引人注目的是Baffle和Ultima，這是AlphaFold的多重擴展(multiple extension)，加上Moment，用於他們的分子設計平台。

This is a super cool, really uh really amazing uh implementation that takes a lot of these technologies, which are typically very difficult for people to access and makes them available in a totally no-code uh platform. Uh, so they integrated Moment, AlphaFold, DiffDock um into their electronic lab notebook solution. Um, and the result of that is that as you’re working with your own data, you now have instant access to workflows for chemical design, for structure prediction and other pieces of this uh this blueprint that they’ve implemented. And last, certainly, but not least, uh we work very closely with Cadence um and OpenEye on their Orion molecular design platform uh integrating by email. They took advantage of a number of NIMs that we’ve published recently, most perhaps notably Baffle to Ultima. This is the multiple extension um of AlphaFold in addition to Moment uh for their molecular design platform as well.

好，我們只剩幾分鐘，我想告訴你們我們正在進行的下一代工作，讓你們了解接下來會有什麼。我們一直在努力研究，並將這些成果應用到我今天提到的產品中。其中一個是名為Grok的模型，這是下一代的通用分子設計平台(general molecular design platform)，基於片段(fragment-based)。這有很多優勢。它們使用稍微不同的語法，但結果是你可以真正將新化學物質(novel chemistry)的生成根植於現實，而現實可以被定義為，例如是否可合成(synthesizable)。你也可以從你想填入的大量元素(mass elements)開始，讓生成過程根植於個別使用者能理解的東西，並以模型範例為基礎。現在很多人正在利用這一點，來擴展我們在Moment上的工作。我們這週還發布了新的NIMs，包括MMSeqs和OpenFold 2。這些模型就像我們處理其他模型一樣，我們讓它們的速度大幅提升。例如，我們的MSA搜尋NIM(MSA search NIM)，在底部的那個，比標準的AlphaFold 2實作快了23倍，而OpenFold 2現在比之前提到的AlphaFold 2快了7.5倍。端到端(end-to-end)，你在結構預測(structure prediction)上獲得了真正領先的性能，這些NIMs促成了這一點。我們這裡還有一些小藍圖供你試用，還有如何將這些東西串聯起來，用於結構預測和下一代虛擬篩選藍圖(virtual screen blueprint)的範例，也就是所謂的V2.0版，它包含了更快的對齊(alignment)、MMSeqs、更快的結構預測和OpenFold，以及下一代基於片段的設計，在Grok中實現，並在V2.0中提升了性能和準確度。這是我們在虛擬篩選藍圖構建方式上的一個重要世代改進。

Okay, so we’ve got just a few minutes left and I want to tell you uh the next-generation sort of work that we’re doing and give you a sense of what’s coming next. Um, we’ve been working very hard on research work uh that we can expose in the products that I told you about today. Uh, one of them is a model called Grok. Uh, this is the next generation, a general molecular design platform um that is fragment-based. Um, there are a lot of advantages to this. Um, they use a slightly different grammar. But what the result of this is that you can really ground um the generation of novel chemistry um in reality and reality can be defined as, is it synthesizable for an example. Uh, but you can also start with, you know, mass elements that you want to fill in. So you can ground the generation in something that um is understandable by the individual user uh with a model example. And so a lot of people are taking advantage of this now uh to extend the work that we’ve been doing in Moment. Uh, we’re also releasing new NIMs this week. Those include MMSeqs and OpenFold 2. Um, these are models that actually we worked just like we’ve worked with the other models and made them a lot faster. So our MSA search NIM, um which is there at the bottom um is, I think, twenty-three times faster than the standard uh implementation in AlphaFold 2 and our OpenFold 2 uh is now 7.5 times faster than the AlphaFold 2 uh that would be previously discussed. And so end-to-end you have really state-of-the-art uh performance in structure prediction uh facilitated by these and we have uh little blueprints here for you to try out um and examples of how you might stitch these things together for uh the purpose of structure prediction end-to-end as well as the next generation of our virtual screen blueprint. So-called V 2.0, um which incorporates much faster uh alignment and MMSeqs, much faster structure prediction and OpenFold. Next-generation fragment-based design, um in Grok, and improved performance and accuracy in V2.0. So a really important uh generational improvement uh in the way that we built blueprints for virtual screening.

好，最後但同樣重要，我還有兩分鐘。這真的很重要。你們在這場演講中聽到我多次提到我們如何進行加速，也就是我們所說的「向下推展加速(drive them down the stack)」。這意味著我們找到產業中的關鍵瓶頸(critical bottlenecks)，並想出如何以對開發者友善的方式解決這些問題。我們與所有開發生物學基礎模型(foundation modeling in biology)的團隊合作。這些只是最近發布的一些重要模型的範例。我們想做的是找到方法，讓我們能在這個領域中，對各種不同類型的模型實現共同加速。最近我們與MIT以及Genesis Therapeutics的合作者合作，加速了一個名為Bolt 1的模型。Bolt 1是一個完全開源(open source)的架構。我們透過與TensorRT的合作實現了這一點，我已經多次提到這個非常重要的函式庫(library)，還有我們與cuDNN團隊打造的自訂核心(custom kernel)，也就是早年在AI模型性能上定義標準的那個NIM團隊。基於這兩項技術，我們在Bolt 1架構上實現了近五倍的推論性能提升。現在我們在思考，我們知道如何將這些整合起來，也了解了一些早期的加速成果。那麼，我們如何將這些應用到正在發生的下一代工作中，讓它變得更廣泛適用呢？

Okay, last, but not least I have 2 minutes left. Uh, this is really important. Um, we’ve also been… You’ve heard a lot from me over this talk about the uh the way that we take acceleration that we so-called drive them down the stack, which means uh we find things that are critical bottlenecks for the industry. And we figure out how to close those in ways that are really developer-friendly. And so we have been working with all of the groups that are developing foundation modeling in biology. Um, and uh these are just examples of important uh models that have been uh published in the recent past. Um, and what we want to do is find ways where we can develop common acceleration across a lot of different types of models in this space. And so one of the things that we did recently is partner uh with MIT uh and with the collaborators at Genesis Therapeutics uh to accelerate a model called Bolt 1. Um, and Bolt 1 is a fully open-source uh architecture um and uh what we’re able to achieve uh through work that we did with TensorRT. So I’ve mentioned that library a few times, very very important for us, as well as custom kernel that we’ve built with the cuDNN team, um yes that same team that sort of defined uh early performance in AI models um to get almost a five X speed up in inference performance based on those two technologies in the Bolt 1 architecture. And so what we’re doing now, um, is we’re thinking about okay so we know how to put these things together um and we know about some of the early acceleration that we’re getting here. How do we actually take this and make it general across all of the next generation of work that’s happening?

這有點像是對未來的預告。請在其他會議上回來找我們，我們會在那裡分享如何確保這些技術能應用於下一代生物基礎模型(next generation biological foundation models)的整個生態系統。我很高興向大家介紹我們與Fancia和Dino Therapeutics的合作。他們是我們很棒的合作夥伴。我相信明天Fancia將會宣布Neo 1，這是他們的原子級結構(atomistic structure)和全新設計模型(de novo design model)，是首個同時具備這兩種功能的模型。而Dino Therapeutics的DinnerFolds則與我們在這個領域合作，旨在理解並從動態模擬(dynamics)中快速取樣，比實際運行基於物理的計算(physics-based calculation)快得多。這些都是我們技術合作夥伴帶來了不起的創新。這不是故事的終點。我們有一個完整的生態系統，許多人在一起努力，真正站在設計分子(molecular design)、標準設計工作流程(standard design workflows)和結構預測工作流程(structure prediction workflows)能力的最前沿。我們非常驕傲能與你們在螢幕上看到的每個人，以及我們整個技術生物製藥(tech bio pharma)和數位生物學(digital biology)生態系統中的夥伴合作。最後，感謝你們的參與，感謝你們來到這裡。我們學到了一些東西，可惜沒有時間提問。但我們會在後續努力中補上。

And so this is a little bit of a teaser for the future. Uh, please come back and see us at other conferences uh where we’ll find ways of making sure that this is available to the entire ecosystem of next-generation uh biological foundation models. I’m really pleased to tell everyone about the collaborations that we have with Fancia and Dino Therapeutics. Uh, these are great partners of ours. Uh, then I believe tomorrow is going to announce Neo 1, uh which is their uh atomistic structure and de novo design model, is the first of its kind that both. Um, and then DinnerFolds from Dino Therapeutics, which we’ve been collaborating in this space um to understand and to be able to sample from dynamics much much faster than actually running a physics-based calculation. Um, so really incredible innovations from our technical partners here. Um, and this is not the end of the story. Uh, we have an entire ecosystem of people that are putting all this together um and really innovating at the bleeding edge of our ability to design molecules. Standard design workflows, structure prediction workflows. Uh, we’re extremely proud of our collaborations with everyone you see here uh on this screen and in our entire tech bio pharma and digital biology ecosystem. So without closing, thank you for paying and thank you for being here. And we learned something and unfortunately, we have time for questions. But we’ll uh well make effort in the um uh yeah.