Iambic Therapeutics, Inc.

Iambic Therapeutics has used its AI models to build a pipeline and put its first drug candidate in the clinic earlier this year. Now, the San Diego company is debuting a model with even bolder hopes for what AI can tackle in biotech. On Tuesday, Iambic announced Enchant, a transformer-based model to make clinical predictions. The model is trained on a wide array of preclinical and research data, along with a small amount of clinical data. In a white paper , the biotech said the model exceeded state-of-the-art performance in predicting certain properties like the half-life of a compound in the human body. Its leaders believe the model will excel at making many more predictions that ultimately determine how good an experimental drug will be in humans. “We’ve arrived at a really exciting point with this technology,” Iambic’s co-founder and chief technology officer Fred Manby told Endpoints News . “This demonstration that, yes, you get better at clinical predictions by training on more preclinical data really feels like a moment where something qualitative has changed about the way you can leverage AI in drug discovery.” Iambic’s leaders said Enchant aims to “break the wall” between preclinical and clinical research. The biotech already uses the model’s predictions for properties like half-life or solubility of compounds when selecting preclinical drug candidates. Other early efforts in this space include Google DeepMind’s Tx-LLM , Insilico’s nach0 , and OpenAI’s ChatGPT . Iambic CEO Tom Miller said Enchant’s performance outranks these other models, in a different way than how DeepMind led the competitive space in protein-structure prediction with AlphaFold. He described Enchant’s performance in making clinical predictions as a “zero-to-one situation” for the AI bio field. “The state-of-the-art has nothing to do with DeepMind,” Miller said. “This is a situation where that is just one among a bunch of other methods that aren’t very good and aren’t surpassing the state-of-the-art.” Specifically, Enchant beat the publicly available state-of-the-art performance for predicting the half-life of compounds. On a zero-to-one scale measuring how its predictions match reality, Iambic’s model scored 0.73, showing a moderate-to-strong correlation. The previous record was a 0.58 from a model developed by a research team at a small San Diego biotech called Expert Systems. Iambic started working on Enchant about a year ago, feeding it a wide variety of data, from PubMed and bioRxiv to the Protein Data Bank and Iambic’s own laboratory data. Enchant’s design — a multimodal transformer — mirrors ChatGPT in relying on the AI to make connections across seemingly disparate datasets. Iambic’s leaders believe doing so could better connect lab research on drugs to how they actually perform in the human body — a critical challenge for boosting clinical success rates. Iambic described parts of its performance, like the half-life metric, in a blog post, which was not published in a peer-reviewed journal. Manby said he expects future publications on Enchant and referenced a paper presented in July at a machine learning conference describing an earlier version of the model. While the new white paper provides some examples, it’s hard to tell if those are hand-picked measures of where it shines, or if Enchant is on track to predicting a wide range of clinical properties. Its leaders said they are confident it can beat state-of-the-art performance on many measures beyond half-life. “We’re optimistic we’ll be able to reproduce that across a whole slew of different clinical endpoints,” Manby said.

Terray Therapeutics has raised a $120 million Series B round, aiming to get its first drug candidate into the clinic while advancing in its AI-fueled approach to making molecules. Founding CEO Jacob Berlin exclusively told Endpoints News that the 110-employee startup is generating the “key dataset missing for AI-driven drug discovery” in studying billions of protein-small molecule interactions. Terray has developed tiny chips that can generate more biological data than previously possible. Berlin has spent the past 12 years developing these chips, or microarrays, which are about the size of a nickel and contain millions of wells to hold molecules in place. He left a tenured professorship at City of Hope to start Terray in 2018, which now operates out of a 52,000-square-foot laboratory in the foothills of Los Angeles. The biotech uses these chips to see how tens of millions of molecules bind — or don’t — to a protein of interest. That echoes a common refrain that AI breakthroughs only happen with enough of the right data. And getting those biological data require the right hardware. In the case of the recent Nobel Prize-winning protein model, AlphaFold, it was built off the Protein Data Bank, a painstakingly curated public resource that itself was accelerated by then-new crystallography and sequencing machines. Terray has generated its own proprietary dataset of protein-molecule interactions 50 times larger than publicly available data. The biotech has used that data to train AI models to discover and design small molecules in a new way for a biopharma company, placing the AI’s ideas at the center of its R&D process. The question is if these chips, data and models will ultimately make better drugs. Terray’s Series B will provide multiple years of runway to start finding an answer, which includes aiming to enter the clinic with its lead drug in 2026, Berlin said. Terray is keeping its pipeline under wraps, beyond disclosing a general focus on immunology. Terray is one of several AI bios taking that chance to build its own pipeline. AI-focused startups also working on small molecules include Iambic Therapeutics, Exscientia and Atomwise. Terray CFO and COO Eli Berlin, the brother of the CEO, said the biotech’s hardware has been key in attracting investors in this market. “People recognize the difference in our story,” he said. “That hardware innovation really resonated, because it lets us build these scaled datasets that are precise and iterate rapidly.” Terray’s chips helped design its first molecules, which Jacob Berlin said are often in new chemical space. “Every single molecule we bring forward has a scaffold that has been de novo discovered and optimized at Terray from the ground up,” Berlin said. “That’s totally different than what you see across the ecosystem, where AI models have to stay close to historical sets of data because they don’t have our hardware innovation.” Terray’s process starts by studying a protein target of interest against 80 million molecules, synthesized on nanoscale beads and placed on its chips. After first looking for binding molecules, they then zoom in on the hits to run more screens. As Terray brings compounds into lead optimization or the stage of refining a lead into a drug candidate, AI models suggest what to do next. Humans still make the decisions, Berlin said, but AI increasingly shapes Terray’s experiments. For instance, the company built a diffusion model to generate small molecules. In an August preprint , Terray’s team put their own twist on the diffusion architecture. They require it not only to diffuse out from a starting molecule to generate new compounds but also to diffuse backward to that starting point. That acts like guard rails in restricting the model’s creations, discouraging it from generating radically different molecules. “Computationally, it basically inherently bounds just how wild it gets,” Berlin said. For the molecules at the end of this process, Terray is keeping them under wraps, but Berlin said they have completely new designs that may raise some industry eyebrows. “In every case, med chemists would be surprised. They would say, ‘Huh,’” he said. “I’m surprised by what we find and what we move, so I think other people will be too.” Editor’s note: This story has been corrected in referring to Terray’s chip as microarrays, not microassays.

近日，瑞典皇家科学院宣布，将2024年诺贝尔化学奖授予蛋白质设计先驱David Baker教授与AlphaFold开发者Demis Hassabis、John M.Jumpe。 继前几日AI教父John J.Hopfield和Geoffrey E.Hinton获得诺贝尔物理学奖后，AI+制药已成为诺奖的焦点。 如果以诺贝尔奖作为人类历年科学发展的缩影，那么对制药业而言，诺贝尔生理学或医学奖提供的是作用机制、诺贝尔化学奖提供的是构建方法、诺贝尔物理学奖提供的是处理效率。 技术的不断突破，让AI制药站上风口。 百舸争流 AI技术在制药领域的应用已经引起了广泛的关注，有数据显示，截至2023年底，全球共有897家AI制药企业。其中国内AI制药企业就达到了93家，而2020年底时仅为16家。 从主要布局来看，大多数AI制药公司还是将重心放在了早期药物开发这一起始阶段上。 图1 AI制药公司布局 图片来源：参考来源3 从AI制药参与者类型来看，主要分为三大类，IT巨头、AI制药企业和大型药企。三类企业依托各自在研发、生态、算法的优势切入行业。 表1 AI制药参与企业分类 数据来源：公开资料整理 为了降本增效，MNC们对AI充满着无限热情，年初的JPM大会上，谷歌母公司Alphabet旗下的AI生物技术公司Isomorphic Labs宣布与礼来和诺华签署了两笔价值近30亿美元的大额交易；赛诺菲更是在去年宣布了“All in AI”战略；近日礼来也任命Thomas J.Fuchs为公司第一任首席人工智能官，各大跨国制药企业动作不断。 经统计，仅2023年MNC在AI+药物研发领域就达成了超30项合作，已披露的总价值约100亿美元，单笔总额最高达27亿美元，其热情可见一斑。而在国内，恒瑞、上海医药等传统企业也早于21年与XtalPi等人工智能企业达成合作。 图2 部分药企与人工智能企业合作概况 图片来源：参考来源5 对于IT巨头而言，依托自身强大的算法与科技实力，作为“卖铲人”获利也是一条一本万利的道路。 根据媒体报道，软银集团宣布计划在2025年前共投资1500亿日元（约9.6亿美元）来增强其人工智能(AI)数据中心的运算能力。 而腾讯早已参与其中，其投资的中国AI药物研究公司晶泰科技(QuantumPharm)6月13日在香港上市，成为国产AI制药第一股，并且在不久前结束的腾讯全球数字生态大会上，腾讯健康总裁吴文达在发言中指出，腾讯健康依托于大模型打造的解决方案已经在1300多家企业、医院等机构落地。 英伟达的参与更加令人注目，这家全球知名的科技公司，通过旗下的风险投资基金NVentures，积极拓展其在AI制药领域的投资版图。据智药局的数据监测显示，从2023年到2024年6月6日，英伟达共参与投资了10家AI制药公司。 此外，其自身研发的生成式AI平台NVIDIA BioNeMo，可用于蛋白质结构预测、蛋白质序列生成、分子优化、生成式化学、对接预测等。据悉，已有100多家生物制药和AI药物研发公司使用其BioNeMo平台，包括安进、安斯泰来制药、Cadence、Iambic、Insilico Medicine、Recursion和Terray Therapeutics，风光无限。 对于AI制药初创企业而言，还有很长的路要走。 从全球来看，目前海外仅有20余家AI制药企业上市，时间集中在2020年之后。 图3 海外已上市AI制药企业 图片来源：参考来源6 从公布财报看，2023年美国AI制药公司Schrodinger（薛定谔）年收入2.17亿美元，排名第一，其也是首家实现盈利的AI制药企业：其收入主要由三部分组成：软件收入、药物研发收入和其他收入，全年软件收入占了大头，收入为1.591亿美元，增长17.4%。 从研发管线来看，靶向MALT1的SGR-1505和靶向CDC7的SGR-2921均处于临床I期。另外，财报提到，未来薛定谔将会把资源越来越多地放在专利药物发现项目上，这也许意味着薛定谔会在未来锚定“卖铲人”的角色。 从国内来看，目前仅有一家上市公司晶泰科技于今年6月13日在香港上市，成为国产AI制药第一股。 晶泰科技于2014年成立，在2016年，在辉瑞举办的一场全球晶体结构预测盲测比赛上，刚创立不久的晶泰科技实现了100%准确预测，以此为契机开启了与头部跨国药企的战略合作 据招股书数据，在IPO前的融资历程中，晶泰科技共完成了8轮融资，累计金额达7.3亿美元（约合人民币52亿元），在全球AI制药企业中融资额排名第一。 目前公司拥有160+授权专利，正在进行超39项药物发现项目，已为全球300多家生物技术与制药公司及研究机构提供服务，其中包括全球前20大生物技术与制药公司（按2022年收入计）中的16家。 在国内除了晶泰科技外，目前还有进入到递交招股书环节的英汐智能、B+轮的红云生物和奕拓生物以及C轮的药物牧场、深势科技、甫康药业等企业。 AI制药公司们已经进入到上市竞速赛之中。 未来如何？ AI制药红透半边天，但是不可否认的事实是截至目前还没有完全由AI研发的新药物进入市场。 从融资的角度来看，2021年行业融资金额达历史高点，超过290亿美元，随后开始大幅下挫。 图4 历年AI制药融资情况 图片来源：参考来源7 分析原因，主要是由于当前AI制药技术不成熟，技术成果转化有障碍等原因。 比如文献显示类药物虚拟筛选库已拥有超过360亿种化合物，对于当下考虑到每个分子估计有10-50个构象异构体，总计超过5000亿个对象。虚拟筛选与实际验证的挑战都是巨大的，遑论常用的AI模型通常针对数万种或更少的化合物进行训练，这引发了人们对其在360亿种化合物的化学空间采样方面的有效性产生疑问。 预测的有效性与验证成本限制了AI制药的发展，也是目前为何没有一款上市AI药物的主要原因。 此外，另外一方面，则是由于很多公司“挂羊头卖狗肉”。 在探究AI制药行业时，波士顿咨询曾提出一个新式的词——“AI原生公司”，指的是那些真正拿AI做实事，而不是借着AI的噱头打广告的公司。 在进一步的筛选后，波士顿咨询有了惊人的发现，全世界AI制药企业高达700多家，可真正的“AI原生公司”，竟然只有114家（截至报告时间）。 技术的不成熟加上部分入局者的不靠谱，导致AI制药的融资与发展受到阻碍，甚至一些明星公司都受到极大的影响。 比如业内的明星公司Exscientia，其市值已经从上市之初的超30亿美元跌到了如今的5.97亿美元（截至2024年10月10日）。 并且由于其在管线上的接连失误，以及停止了针对A2A靶点的癌症候选药物EXS-21546的I/II期研究，进一步打击了市场信心。最终只能被英伟达投资的另一家AI制药公司Recursion并购，交易金额仅为6.88亿美元，不足上市之初的30%。 领头羊尚且如此，何论那些缺乏资金与技术实力的小型AI制药企业。 那么，未来会好吗？答案是一定的。 对于AI制药企业而言，其应用框架主要包括三个部分：数据收集和整理、使用分子描述符进行化合物表示、人工智能方法及其应用。 AI训练需要大量的数据验证与反馈，但是每一次的反馈与优化都会对下一次的优化进行拓展与提升。 这是一个逐步迭代的过程，随着数据的逐渐沉淀、模型的逐渐优化、算力的逐渐提升，当触发了爱因斯坦称为第八大奇迹的“复利”后，AI赋能制药的效率将会指数级的增长。 并且，商业的竞争本质是效率的竞争，新领域“first in class”药物所带来的优势是后来者无法比拟的。更快、更准确的药物发现、构建、验证对药企们来说具有无与伦比的吸引力，大浪淘沙之后，AI制药公司前景无限。 最后借用默沙东创始人George W.Merck讲过一句话，“我们应当永远铭记：药品是为人类健康而生产，不是为追求利润而制造的。只要我们坚守这一信念，利润必将随之而来。” 随着科技的进步，制药企业终将拿起AI这把效率之刃，劈开那些侵袭人类的病魔！ 参考来源： 1.《医药行业估值》 2.Artificial Intelligence for Drug Discovery:Are We There Yet?，https://doi-org.libproxy1.nus.edu.sg/10.1146/annurev-pharmtox-040323-040828 3.《AI制药商业模式分析：以英矽智能、晶泰科技为例》，华创证券 4.《全球AI智能制药公司布局情况现状分析（图）：早期药物开发是主要布局领域》，中研网 5.晶泰科技招股书 6.《中国AI制药企业白皮书》 7.《2024，集体逃离AI制药》，智药局 8.《50亿，AI制药最大并购诞生》，东四十条资本 9.《药物简史》 10.《英伟达、软银入局，AI制药是“王炸”还是概念？》，未来数字城市探索者 11.《AI助力新药研发破局上市公司抢滩黄金赛道》，证券时报网 12.《万字长文：AI制药的前世今生、药物发现、结构预测、从头设计、行业图谱》，BiG生物创新社 声明：本内容仅用作医药行业信息传播，为作者独立观点，不代表药智网立场。如需转载，请务必注明文章作者和来源。对本文有异议或投诉，请联系maxuelian@yaozh.com。 责任编辑 | 史蒂文 转载开白 | 马老师 18996384680（同微信） 商务合作 | 王存星 19922864877（同微信）  阅读原文，是受欢迎的文章哦