智东西（公众号：zhidxcom）
编译 | 黄菲
编辑 | 江宇
智东西10月30日消息，昨日，国内AI for Science新锐企业扰动之心（PerturbMind AI）宣布完成数千万元天使轮及天使+轮融资，由清松资本领投，同杨基金及多家药企跟投，正式走出隐匿模式。
该公司聚焦几何世界模型驱动的下一代AIDD（AI药物研发）创新范式，探索以生命系统状态推演为核心的研发模式。
此次官宣，扰动之心还发布了数字生命世界模型生态系统PerturbSeek旗下的全球首个虚拟细胞状态推演模型AlphaCell 1.0 Pro。
该模型以几何世界模型为底层架构，标准参数量为1B，可根据应用需求做调整，向下可裁剪至 300M的轻量化模型AlphaCell 1.0 Flash，向上可扩展至50B参数的超大规模模型AlphaCell 1.0 Mega，适配科研机构、中小型药企及大型跨国药企的不同部署需求。
与传统AI药物研发不同，扰动之心希望通过几何世界模型学习生命系统在不同干预条件下的动态变化，推演疾病状态迁移及药物干预后的系统反应，进而为复杂疾病机制解析、新靶点发现和新型模态药物研发提供技术支持。
一、从靶点筛选转向状态推演，发布虚拟细胞模型
扰动之心此次发布的AlphaCell 1.0 Pro，是该公司虚拟细胞模型AlphaCell 1.0 Preview的升级版本。
今年3月5日，扰动之心创始人兼CEO啜国晖带领的扰动之心团队发布在bioRxiv发布AlphaCell相关预印本论文，介绍了该模型基于几何世界模型架构模拟扰动引起的细胞动态变化的方法。
▲扰动之心创始团队发布AlphaCell 1.0 Preview论文，来源bioRxiv
此次发布的AlphaCell 1.0 Pro标准参数量为1B，并具备参数量弹性调整能力，可根据客户算力条件和研发需求动态调整模型规模。
其中，模型向下可裁剪至300M参数，形成轻量化版本AlphaCell 1.0 Flash，满足医院、科研机构及中小型药企的轻量化和私有化部署需求；向上可扩展至50B参数的AlphaCell 1.0 Mega，为大型跨国药企（MNC）提供复杂场景下的系统级扰动轨迹推演能力。
在商业化方面，扰动之心计划围绕不同规模的模型提供端侧服务订阅、私有化部署及药物研发管线共建等服务。
在技术路线方面，扰动之心希望改变传统AI药物研发（AI-driven Drug Discovery，AIDD）以靶点为中心的研发模式。
传统AIDD通常围绕已知靶点寻找能够与之结合的候选药物分子，类似于根据一把已知的锁去找到合适的钥匙。但对于复杂疾病而言，单一靶点未必能够解释疾病产生的全过程，药物干预也可能引发其他生物系统层面的变化。
扰动之心谈道，疾病本质上是生命系统原有稳态遭到破坏后产生的状态偏移。因此，药物研发不仅需要寻找能够作用于特定靶点的分子，还需要理解疾病状态如何形成，以及不同干预手段可能带来怎样的系统性变化。
基于这一思路，该公司将研发重点从单个靶点转向整个生命系统的状态变化，希望通过模型推演识别能够使疾病状态恢复正常的干预方式。
二、几何世界模型构建技术底座，探索生命系统推演
围绕上述技术路线，扰动之心打造了PerturbSeek数字生命世界模型生态系统，以扰动组学、几何深度学习和几何世界模型理论为三大技术支柱。
其中，扰动组学主要研究基因、蛋白质等生物对象受到不同干预后产生的变化，为理解生命系统的动态反应提供数据基础。
几何深度学习则用于学习生物数据背后的复杂结构。扰动之心称，其通过构建几何深度学习算法，对生物隐空间流形（Biological Latent Manifold）进行学习和建模，将复杂的生物数据映射到潜在空间中，以表征不同细胞状态及其内在关系。
几何世界模型则在此基础上，进一步学习生命系统的状态及其变化规律。它将细胞状态和外部干预纳入统一的模型框架，模拟基因、药物等不同因素作用下细胞状态的变化，进而推演生命系统的动态演化过程。
简单来说，该技术路线希望其模型不仅能够识别某种干预与生物变化之间的关联，还能够学习生命系统在不同条件下如何变化，并推演特定干预可能产生的结果。
在此基础上，PerturbSeek希望结合具体疾病和临床约束条件，识别具有潜在治疗价值的干预方式，而不只是围绕已知靶点进行候选分子筛选。
针对模型训练，扰动之心并不主张单纯扩大参数规模和数据量。该公司提到，生命科学领域的模型训练未必能够直接套用通用AI领域的规模定律（Scaling Law），模型参数量需要与实际任务需求、预测性能及计算成本相匹配。
因此，扰动之心希望寻找模型性能与研发成本之间的平衡点，即参数量甜点（Sweet Spot），通过优化底层架构和训练效率，在有限样本条件下提升模型的泛化能力。
三、同济科研团队创业，携手多家药企推动技术落地
扰动之心创始人兼CEO啜国晖博士毕业于同济大学数字生命智能体实验室。啜国晖称，该公司的底层技术创新一定程度上受到谷歌推出的AI科研模型DeepMind Alpha系列工作的启发，希望将前沿世界模型技术转化为药企和临床能够使用的平台解决方案。
▲扰动之心创始人兼CEO啜国晖，来源：同济大学官网
在他看来，生命系统具有复杂的动态变化特征，模型不仅需要理解不同生物对象之间的关系，还需要刻画系统在受到外部干预后的状态迁移。扰动之心希望通过几何世界模型建立数字生命系统的状态表征和状态转移模型，并将推演结果转化为药物研发线索。
该公司首席科学家、同济大学数字生命智能体实验室负责人刘琦教授称，新型模态药物研发需要对复杂生命系统进行整体理解和建模。传统AIDD在已知靶点的分子设计方面已经积累了较多技术，但在复杂疾病机制解析和新靶点发现方面仍面临挑战。
▲扰动之心首席科学家、同济大学数字生命智能体实验室负责人刘琦，来源：同济大学官网
刘琦谈道，扰动之心希望将药物研发从单纯的分子层面建模，推进至生命系统状态层面的理解和推演，再通过实验验证和数据反馈形成闭环，让数字模型与真实生命系统之间建立联系。
在科研积累方面，扰动之心团队曾与罗氏、阿斯利康、药明康德等国内外药企及合同研究组织（CRO）展开合作。该团队相关研究成果发表于Nature Methods、Nature Machine Intelligence、Nature Computational Science、Cancer Cell等期刊，并获得药明康德生命化学研究奖、吴文俊人工智能自然科学奖等。
产业合作方面，扰动之心称，该公司已与国内一家头部国产芯片企业签署战略合作协议，PerturbSeek引擎已完成对国产GPU芯片的全栈适配。
此外，该公司还与国内特定领域的药物研发企业签署合作协议，正在开发和部署多个原型系统，并围绕具体应用场景建设数据及实验反馈闭环。
结语：生命世界模型走向产业，实际研发价值何时落地？
从靶点筛选到生命系统状态推演，AI药物研发正在探索新的技术路径。
不过，世界模型能否真正理解复杂生命系统，仍需要通过持续的实验验证和实际药物研发项目检验。
对于扰动之心而言，如何将模型预测转化为可验证的药物研发线索，并进一步推动新靶点发现和新型药物研发，是其后续商业化落地需要回答的问题。
来源：扰动之心
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
Compiled.
Editor Ziang Jiang Woo
Yesterday, according to the news of October 30th, the country ' s AI for Science’s Nueva Nueva Disturbation (PerturbMind AI) announced the completion of tens of millions of dollars of Angel Wheels and Angel + Finance, with investments from Qing Song Capital, the Yang Foundation and a number of pharmacists, to officially move out of the hidden mode.
The company focuses on the next generation of AIDD (AI Drug Research and Development) innovative paradigms driven by geometry world models and explores research and development models with life-system status driven at their core.
This time, the Disturbing Heart also announced publish's first virtual cell status model, AlphaCell 1.0 Pro, in a digital life world model ecosystem, PerturbSeek.
The model is based on a geometric world model with standard parameter count as 1B, which can be adapted to the application needs, and the light quantization model, AlphaCell 1.0 Flash, which can be cut down to 300M, the super-scale model, AlphaCell 1.0 Mega, which can be extended upward to 50B parameters, and the different deployment requirements of appropriate scientific institutions, small and medium-sized pharmacists and large transnational pharmacists.
Unlike traditional AI drug development, the disturbing heart hopes to learn about the dynamics of life systems under different intervention conditions through geometric world models, and to promote the transport of disease and the systemic response after drug intervention, thus providing technical support for the analysis of complex disease mechanisms, the discovery of new target points and the development of new model drugs.
publish virtual cell model
AlphaCell 1.0 Pro of publish is an upgraded version of the company ' s virtual cell model AlphaCell 1.0 Preview.
On March 5, this year, publish, the disturbing team led by the CEO, the founder of the disturbing heart, presented a preprint paper in BioRxivpublish AlphaCell on the model ' s approach to modelling cell dynamics caused by disturbances based on geometric world model structures.
publish AlphaCell 1.0 Preview paper from BioRxiv
AlphaCell 1.0 Pro Standard parameter count for publish this time is 1B with parameter count flexibility to adjust the size of the model according to the client ' s ability to calculate and develop dynamic demand.
Among these, the model can be reduced downward to 300M parameters, forming AlphaCell 1.0 Flash, version of Light quantization, to meet the light quantization and privatization deployment needs of hospitals, scientific institutions and small and medium-sized pharmacists; AlphaCell 1.0 Mega, which can be extended upward to 50B parameters, to provide a system-level disturbance trajectory in a complex scenario for large transnational pharmacies (MNC).
In terms of commercialization, the Disturbing Heart plans to provide end-side service subscriptions, privatization deployments and drug R & D pipelines around models of different sizes.
In terms of technical routes, the disturbing heart hopes to change the target-centred R & D model of traditional AI drug development (AI-driving Drug Discovery, AIDSD).
Traditional AIDSD usually looks around known targets for candidate drug molecules that can be integrated, similar to finding the right key under a known lock. For complex diseases, however, a single target may not necessarily explain the entire process of disease generation, and drug interventions may also trigger changes at the level of other biological systems.
Disturbingly speaking, disease is essentially a state of deviation resulting from the destruction of the original stability of the life system. Drug research and development therefore requires not only the search for molecules that can work on specific targets, but also an understanding of how the state of disease is formed and what systemic changes may be brought about by different interventions.
In line with this approach, the company has shifted the focus of its research and development from individual target points to changes in the state of the entire life system, with the hope that modeling will lead to a return to a normal state of intervention.
II. Geometric world model construction technology base to explore life system development
The PerturbSeek digital life world model ecosystem was built around the above-mentioned technical route, with the three main technical pillars of the disturbance grouping, geometric depth learning and geometric world model theory.
Among them, disturbances contortology focuses on changes resulting from different interventions in biological objects such as genes and proteins, providing a data base for understanding the dynamic responses of life systems.
Geometric depth learning is used to learn the complex structures behind biological data. The disturbing heart claims that by building geometric depth learning algorithms, it learns and models bio-inhibit space currents (Biologic Latet Manifold) to map complex biological data into potential space to characterize different cell states and their intrinsic relationships.
On that basis, geometry models further learn about the state of life systems and their patterns of change. It incorporates cell state and external intervention into a unified model framework that simulates changes in cell state by different factors, such as genes, drugs and so forth, and thus triggers the dynamic evolution of the life system.
In short, the technology route hopes that its model will not only identify the link between a particular intervention and biological change, but will also be able to learn how life systems change under different conditions and to extrapolate the possible outcomes of a particular intervention.
On this basis, PerturbSeek wishes to identify interventions of potential therapeutic value, in combination with specific diseases and clinical constraints, rather than merely screening candidates around known targets.
For model training, the heart of disturbance does not advocate simply increasing the size of the parameters and the amount of data. The company mentioned that model training in the field of life sciences does not necessarily directly apply the Scaling Law in the generic AI area, and that model parameter count needs to match actual mission needs, forecasting performance and costing.
Therefore, the disturbing heart seeks to find a balance between model performance and R & D costs, namely, parameter count Sweet Spot, which enhances the generalization of models under limited sample conditions by optimizing the bottom structure and training efficiency.
iii. Start-up of a team of scientists to work with a number of drug companies to promote technology Land
The founder of the Disturbation Heart and Dr. CEO Quixian, graduated from the Digital Life AI agent Laboratory at the University of Toji. The company ' s bottom-up technological innovation, according to Zhong Guo Yi, was to some extent inspired by Google ' s work on the DeepMind Alpha series of the AI scientific model, in the hope of transforming forward-world modelling techniques into platform solutions that could be used by pharmacists and clinically.
Founder of the Disturbing Heart and CEO of the Nation, Source: Tutor University Network
In his view, life systems had complex dynamic dynamics, and models needed to understand not only the relationships between different biological objects, but also the condition of the painting system following external intervention. The Disturbing Heart hopes to develop a digital life system's state representation and state transfer model through geometric world models and to translate the results into a drug development trail.
Professor Liu Qi, Chief Scientist of the company and Head of the Digital Life AI agent Laboratory at Toji University, stated that the development of new model drugs required a holistic understanding and modelling of complex life systems. Traditional AIDSD has accumulated more technology in molecular design of known target points, but challenges remain in the analysis of complex disease mechanisms and the detection of new target points.
Liu Qi, Chief Scientist of Disturbation and Head of the Digital Life AI agent Laboratory at Toji University, Source: Toji University Officer Network
Liu Qi spoke of the desire to model the development of drugs from a purely molecular level to an understanding and evolution at the life system state level, and to create a closed circle through experimental validation and data feedback, linking digital models to real life systems.
In terms of scientific research, the Disturbing Heart Team has worked with national and international pharmacies and contract research organizations (CROs) such as Roth, Aslicon and Medicaid. The team ' s research results were published in the periodicals of Nature Methods, Nature Machine Intelligence, Nature Computational Science, Cancer Cell and received the Medicinal Conservatory Award for Life Chemistry, the Wu Wenjun Award for Artificial Intelligence Natural Science, etc.
With regard to industrial cooperation, it is disturbing to note that the company has signed a strategic cooperation agreement with a domestic head-chip company and that the PerturbSeek engine has completed its full match-up of the national GPS chip.
In addition, the company has signed a cooperation agreement with drug research and development enterprises in specific areas of the country, and is developing and deploying prototype systems and building data and laboratory feedback loops around specific applications.
Concluding remarks: When does the real R&D value land when life world models go to industry?
From target screening to life-system state, new technological pathways are being explored in AI drug development.
However, whether or not the world model truly understands complex life systems still needs to be tested through ongoing experimental testing and actual drug development projects.
The question of how to translate model predictions into verifiable drug research and development leads and to further promote the discovery of new targets and the development of new drugs for disturbance is a question to be answered in their subsequent commercialization.
Source: Disturbing Heart
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.