"BioGeometry" Completes Hundreds of Millions of Yuan Strategic Financing to Build Life Sciences "Micro-World Model" | 36Kr Exclusive 「百奥几何」完成数亿元战略融资，打造生命科学"微观世界模型"｜36氪首发

First, let’s talk about the technology. GeoFlow, the “microscopic world model,” has an ambitious name—it aims to model molecular interactions at the atomic scale. The iterations from V1 to V3 have indeed caught up with AlphaFold 3 in protein structure prediction and even started delving into de novo design. The progress is visible, especially in antibody design, where a 20% hit rate and a three-week turnaround represent a revolutionary leap compared to the traditional, laborious approach of screening hundreds of millions of molecules. But let’s be frank: a 20% hit rate sounds impressive, but what about the 80% failed molecules? Drug development isn’t a lottery—every failure burns through capital. AI may lower the barrier, but it doesn’t eliminate the risk. More critically, the binding affinity of model-generated molecules still leaves “room for improvement.” This is like a chef claiming the recipe is perfect, but the taste is left for the customer to adjust. Between “understanding life” and “designing life,” there might be an entire Pacific Ocean of wet-lab data.

Capital is hypes that “life AI will become the next most imaginative frontier.” This isn’t wrong—the Nobel Prize has been awarded, and global pharmaceutical companies are also engaging in frenzied BD transactions. But in BioGeometry’s story, what strikes me as most alarming is its selling point of “amplifying the value of early-stage molecules.” Traditional pharma companies buy pipelines based on clinical data, valuing later stages more because risks have been partially validated. Now, AI companies claim early-stage molecules can be sold at high prices—this either disrupts the logic or is another bubble. Think about it: if a de novo-designed antibody has impressive preclinical data but flops in human trials, isn’t this “high value” just a castle in the air? BioGeometry touts its zero-shot design and one-round optimization to achieve multiple indicators—certainly a show of prowess. But drug development is a marathon, not a sprint; speed doesn’t always equal quality. Particularly with its adoption of TTS (time-for-quality) technology, trading inference time for quality in protein design is essentially about throwing computational power at the problem. How is this different from large language models burning through GPUs to train parameters? What life sciences demand is precision and reproducibility, not the brute-force aesthetics of “hard work yields miracles.”

Now, let’s look at business collaborations. Over 20 BD deals sound impressive, but the source text lacks detail, focusing only on the “twin-target” case in tumor immunology. AI design can solve难题 that traditional methods struggle with—this is absolutely praiseworthy, such as developing highly specific antibodies to avoid damaging healthy cells. The direction is right. However, the collaboration details are vague—are these successes isolated cases or the norm? Pharma companies’ money doesn’t come from the wind; their willingness to bet on early-stage AI molecules is either because they’re genuinely convinced by the technology or because they’re making desperate bets amid sector anxiety. Backed by Yoshua Bengio and a team with a strong academic background, BioGeometry has made solid contributions with open-source tools—these are tangible advantages. But between academic brilliance and industrialization lies a whole jungle of clinical trials, regulatory approvals, and market dynamics. No matter how impressive AI design is, it ultimately has to prove itself in human bodies. And the uncertainty of biology has never been fully controlled by algorithms.

The narrative of life AI always carries a techno-optimism, as if bigger models and more data can crack the code of life. BioGeometry’s GeoFlow is iterating toward “molecular systems design,” showing great ambition, but I can’t help but pour some cold water on it: life isn’t LEGO blocks. Molecular interactions are highly context-dependent and dynamic. Atomic-level precision modeling sounds beautiful, but with the chaotic environmental factors and individual variations inside cells, can a model really capture it all? Looking back over the past few years, the number of failed AI drug development companies rivals the successes—from Insilico Medicine to Recursion, each once shone brightly, but now everything hinges on clinical results. BioGeometry’s synthetic biology pipeline, involving technology licensing, is a pragmatic step—after all, AI companies need to survive first. But the grand vision of “designing life” will likely take many more years of trial and error in the lab.

At the end of the day, this financing round for BioGeometry is another top-up of capital’s faith in life AI. There are real technological advancements, commercial progress is underway, and the team boasts big names—all worthy of recognition. But I’d rather view this as a high-risk, high-reward adventure: AI is reshaping the rules of the drug development game, moving from “serendipitous discovery” to “rational design.” Yet the road ahead is littered not with flowers, but with failed molecules and burned capital. As observers, we should applaud innovation but also maintain a cold gaze—after all, life sciences never lack stories; what’s lacking is the patience to turn stories into drugs. BioGeometry’s GeoFlow might indeed be the next breakthrough, but until it truly rewrites treatment standards, all praise is still premature. Capital can chase dreams, but reality only pays for results.

百奥几何又融到钱了，数亿元，领投方一堆国资和头部基金。这消息本身不稀奇，2024年AI制药赛道冷得像冰窖，到了2025年，突然又热火朝天，钱像潮水一样涌进来，仿佛大家一夜之间都看懂了生命科学的底层密码。但仔细想想，这轮热闹背后，藏着多少真金白银的信仰，又有多少只是资本在恐慌性地追逐下一个“iPhone时刻”？百奥几何的故事讲得漂亮，从数字AI、物理AI一路讲到生命AI，号称要“设计生命”，听起来就像科幻片开场白。可问题是，生命科学这潭水，深得连诺奖得主都未必能完全趟明白，一个AI模型真能成了新的“上帝之手”？

先说技术。GeoFlow这个“微观世界模型”，名字起得很有野心，想在原子尺度上建模分子相互作用。从V1到V3的迭代，确实在蛋白结构预测上追平了AlphaFold 3，甚至开始搞从头设计。这进步肉眼可见，尤其是在抗体设计上，命中率20%、三周出结果，比起传统“手搓”上亿分子库的笨办法，简直是革命性的。但吐槽一句：20%的命中率，听起来不错，可剩下80%的失败分子呢？药物研发不是买彩票，每一次失败都是烧钱，AI降低了门槛，可没消除风险。更关键的是，模型生成的分子结合亲和力还得“提升空间”，这就像厨师说菜谱完美，但味道还得客人自己调——从“理解生命”到“设计生命”，中间隔着的可能是一整个太平洋的湿实验数据。

资本狂吹“生命AI将成为下一个最具想象力疆域”，这话没错，诺贝尔奖都发了，全球药企也在疯狂BD交易。但百奥几何的故事里，最让我警觉的是它把“早期分子价值放大”当卖点。传统药企买管线，看临床数据，阶段越后越值钱，因为风险已部分验证。现在AI公司宣称早期分子就能高价出手，这要么是颠覆逻辑，要么是另一种泡沫。想想看，如果一个从头设计的抗体在临床前数据光鲜，到了人体试验却趴窝，那这“高价值”岂不是空中楼阁？百奥几何吹嘘的零样本设计、一轮优化搞定多指标，确实炫技，但药物研发是马拉松，不是短跑冲刺，快不一定等于好。尤其是它把TTS技术——用推理时间换质量——引入蛋白质设计，本质上还是堆算力，这跟大语言模型烧卡训参数有何区别？生命科学需要的是精准和可重复，不是“大力出奇迹”的暴力美学。

再看商业合作。20多项BD听起来战果累累，但原文语焉不详，只挑了肿瘤免疫的“孪生靶点”案例猛夸。AI设计能解决传统方法搞不定的难题，这绝对值得赞赏，比如高特异性抗体避免误伤正常细胞，这方向对路。可问题是，合作细节模糊，成功案例是孤证还是常态？药企的钱也不是大风刮来的，他们愿意押注早期AI分子，要么是真被技术说服，要么是赛道焦虑下的赌注。百奥几何背后有Yoshua Bengio站台，团队学术背景硬核，开源工具也贡献不少，这加分项实实在在。但学术光环和产业化之间，隔着临床、审批、市场一整套丛林法则。AI设计再牛，最终还得在人体里说话，而生物学的不确定性，从来不是算法能完全驾驭的。

生命AI的叙事里，总有一种技术乐观主义，仿佛只要模型够大、数据够多，就能破解生命之谜。百奥几何的GeoFlow往“分子系统设计”迭代，野心不小，可我忍不住想泼点冷水：生命不是乐高积木，分子互动有高度的上下文依赖性和动态性，原子级精度建模听起来很美，但细胞内那堆乱七八糟的环境因素、个体差异，模型真能一网打尽？看看过去几年，AI制药领域倒下的公司不比成功案例少，从Insilico Medicine到Recursion，个个都曾风光无限，现在还得看临床结果说话。百奥几何的合成生物学管线搞技术转让，算是务实一步，毕竟AI公司得先活下来，但“设计生命”的宏愿，恐怕还得在实验室里摔打多年。

说到底，百奥几何这轮融资，是资本对生命AI信仰的又一次充值。技术上有真进展，商业化在推进，团队有大牛，这都值得肯定。但我更愿意把这看作一场高风险高回报的冒险：AI正在重塑药物研发的游戏规则，从“偶然发现”走向“理性设计”，可这条路上，堆满的不是鲜花，而是失败的分子和烧掉的钱。作为旁观者，我们既要为创新鼓掌，也得保持冷眼——毕竟，生命科学从不缺故事，缺的是能把故事变成药的耐心。百奥几何的GeoFlow或许真是下一个突破口，但在它真正改写治疗标准之前，一切赞美都还太早。资本可以追逐梦想，但现实只会为结果买单。