作者: Davis, Amy J. ; Miller, Ryan S. ; VerCauteren, Kurt C. ; Walsh, Daniel P. ; Ivan, Jacob S. ; Newton, Erica J. ; Clune, Jeff ; Newkirk, Eric S. ; Erb, John ; Tabak, Michael A. ; Conrey, Reesa Y. ; Wolfson, David W. ; Lewis, Jesse ; Boughton, Raoul K. ; Beasley, James C. ; Norouzzadeh, Mohammad S. ; Brook, Ryan K. ; Iannarilli, Fabiola ; Stenglein, Jennifer ; Odell, Eric A.

Abstract:

Motion‐activated wildlife cameras (or “camera traps”) are frequently used to remotely and noninvasively observe animals. The vast number of images collected from camera trap projects has prompted some biologists to employ machine learning algorithms to automatically recognize species in these images, or at least filter‐out images that do not contain animals. These approaches are often limited by model transferability, as a model trained to recognize species from one location might not work as well for the same species in different locations. Furthermore, these methods often require advanced computational skills, making them inaccessible to many biologists. We used 3 million camera trap images from 18 studies in 10 states across the United States of America to train two deep neural networks, one that recognizes 58 species, the “species model,” and one that determines if an image is empty or if it contains an animal, the “empty‐animal model.” Our species model and empty‐animal model had accuracies of 96.8% and 97.3%, respectively. Furthermore, the models performed well on some out‐of‐sample datasets, as the species model had 91% accuracy on species from Canada (accuracy range 36%–91% across all out‐of‐sample datasets) and the empty‐animal model achieved an accuracy of 91%–94% on out‐of‐sample datasets from different continents. Our software addresses some of the limitations of using machine learning to classify images from camera traps. By including many species from several locations, our species model is potentially applicable to many camera trap studies in North America. We also found that our empty‐animal model can facilitate removal of images without animals globally. We provide the trained models in an R package (MLWIC2: Machine Learning for Wildlife Image Classification in R), which contains Shiny Applications that allow scientists with minimal programming experience to use trained models and train new models in six neural network architectures with varying depths.

2020-09-01·IEEE transactions on neural networks and learning systems1区 · 计算机科学

Teacher–Student Curriculum Learning

1区 · 计算机科学

Article

作者: Taco Cohen ; John Schulman ; Tambet Matiisen ; Avital Oliver

We propose Teacher-Student Curriculum Learning (TSCL), a framework for automatic curriculum learning, where the Student tries to learn a complex task, and the Teacher automatically chooses subtasks from a given set for the Student to train on. We describe a family of Teacher algorithms that rely on the intuition that the Student should practice more those tasks on which it makes the fastest progress, i.e., where the slope of the learning curve is highest. In addition, the Teacher algorithms address the problem of forgetting by also choosing tasks where the Student's performance is getting worse. We demonstrate that TSCL matches or surpasses the results of carefully hand-crafted curricula in two tasks: addition of decimal numbers with long short-term memory (LSTM) and navigation in Minecraft. Our automatically ordered curriculum of submazes enabled to solve a Minecraft maze that could not be solved at all when training directly on that maze, and the learning was an order of magnitude faster than a uniform sampling of those submazes.

302

项与 OpenAI OpCo LLC 相关的新闻（医药）

2025-09-23

·英矽智能

Observer｜Zhavoronkov博士入围AI Power Index榜单

近期，专注记录具有全球影响力人物的纽约《观察家报》（Observer）发布AI Power Index榜单，遴选全球100位推动人工智能跨行业发展的全球领袖。英矽智能创始人、首席执行官兼首席商务官Alex Zhavoronkov博士入选，其在生物技术与AI交叉领域的独特贡献获得表彰。 Observer AI Power Index遴选百位AI行业领袖点击文末【阅读原文】了解详情经过针对候选人在技术、市场和政策方面的多元影响力评估，Observer AI Power Index在全球范围内表彰众多知名行业领袖，如OpenAI联合创始人兼首席执行官Sam Altman、英伟达联合创始人、总裁兼首席执行官Jensen Huang、微软首席执行官Satya Nadella、谷歌和Alphabet首席执行官Sundar Pichai，以及2024年诺贝尔化学奖获得者、Google DeepMind首席执行官兼联合创始人Demis Hassabis。作为英矽智能创始人和多次科睿唯安“高被引科学家”入选者，Alex Zhavoronkov博士带领团队搭建并验证横跨生物学、化学和临床医学、科学的生成式AI解决方案Pharma.AI。自2021年来，英矽智能在Pharma.AI赋能下建立30余条多元化自研创新药物管线，其中10个已获得临床试验许可。2025年6月，英矽智能领先管线Rentosertib（ISM001-055）登上《自然-医学》杂志，其IIa期临床试验数据表明特发性肺纤维化（IPF）患者用力肺活量（FVC）得到改善，有望实现肺功能恢复，完成首个AI药物研发临床概念验证。 Alex Zhavoronkov博士在采访中向《观察家报》强调，“我们正迈向制药超级智能时代，这是一个智能体不仅仅简化工作流程，还会实际做出决策和设计实验的时代。相关的讨论还没有兴起，但当AI开始管理AI，一切都将会改变。” 通过整合先进的AI和自动化技术，英矽智能在实际应用案例中展现出效率提升，为AI驱动的药物研发树立了标杆。与传统药物研发通常需要2.5-4年的时间周期相比，英矽智能在2021至2024年间的自研项目，从立项到提名临床前候选药物（PCC）的平均耗时为12-18个月之间，每个项目仅需合成和测试约60-200 个分子。关于英矽智能英矽智能是一家由生成式人工智能驱动的全球先锋生物科技公司，利用其专有的 Pharma.AI 平台和先进的自动化实验室，加速药物发现并推动生命科学领域的创新。通过整合人工智能和自动化技术，英矽智能正在为纤维化、肿瘤学、免疫学、疼痛、肥胖和代谢紊乱等未满足的疾病领域提供创新药物解决方案。英矽智能由人工智能驱动发现的产品管线中，已有10个分子获得临床试验许可，其中进展最为领先的Rentosertib（原名ISM001-055）是一种潜在全球首创用于治疗特发性肺纤维化的候选药物，已完成2a期临床研究并获得令人鼓舞的积极结果。此外，英矽智能还在探索前沿领域，包括老龄化研究、可持续化学和农业创新方面的突破。更多信息，请访问网站www.insilico.com 商务合作，请联系 bd@insilico.ai 媒体垂询，请联系 pr@insilico.ai

引进/卖出临床2期临床终止临床1期临床结果

2025-09-22

·药物分子设计

「逆龄大脑药」首次人体试验！奥特曼押注RTR242，返老还童将要成真？

定了！北京又一所985，去雄安！刚刚！美国政府制裁中国科技大学等37家实体央企首次发布！材料领域“十大基础科学问题” 第一性原理计算解决50年悬而未决难题：半导体中铜为何扩散更快？ Ab initio及第一性原理入门参考书介绍 985博导亲测：用DeepSeek写国自然本子，3天完成30天工作量 Nature：博士太多，高校已经装不下了！我的博士生已经半个月没主动联系我了, 不知道他是不是在做科研, 怎么能让他更主动一点？固体物理与DFT模拟【4】：好书推荐——谢希德、陆栋《固体能带理论》来自公众号：新智元本文以传播知识为目的，如有侵权请后台联系我们，我们将在第一时间删除。新智元报道编辑：倾倾【新智元导读】长寿正成为硅谷巨头的新赛道！奥特曼投资的Retro要延长人类10年健康寿命，Bezos砸下30亿美元成立Altos Labs，背后是AI、资本与生物科技的全面对撞。问题是，这场豪赌，究竟能换来「逆龄十年」，还是又一场虚幻？ AI教父奥特曼，本该忙着把OpenAI打造成算力帝国。可就在外界以为他只会谈模型、芯片和数据的时候，他却转身丢出一颗「年轻大脑药丸」。这家公司叫Retro Biosciences。奥特曼首投1.8亿美元，如今宣布要在年底前，让第一位临床受试者吃下实验药 RTR242。它的使命，是清理大脑里的「垃圾」，重启衰老细胞的自噬系统，让记忆重新闪亮。如果 ChatGPT颠覆了人与机器的关系，那么奥特曼正试图用另一种方式，颠覆我们和时间的关系。 Retro 的十年豪赌：押注「年轻大脑」 Retro Biosciences从2021年创立之初，就打出一个足够疯狂的目标：不是拖慢衰老，而是要真正逆转，让人类多10年健康寿命。奥特曼为它砸下了1.8亿美元作为启动资金，如今这家公司宣布年底前将在澳大利亚启动首个临床试验，让第一位受试者吃下名为RTR242的药丸。这颗药的机制不是传统制药巨头那种「清除淀粉样斑块」，而是重启细胞的自噬系统。把随着年龄堆积的变异蛋白、损坏分子通通清掉，从根源解决脑部「垃圾堆积」的问题，试图让阿尔茨海默症患者的记忆重新点亮。更大胆的是，Retro并不满足于一条管线。除了脑部药丸，他们还在研发年轻化血干细胞疗法（RTR890），以及中枢神经系统修复项目（RTR888），目标是替换衰老的免疫细胞，减轻炎症性损伤。 Retro的CEO Joe Betts-LaCroix形容我们要像空调改变人类对「热与能量」的认知一样，把「衰老」与「衰退」从此分开。奥特曼的野心不止如此。他让OpenAI的GPT-4b micro直接加入研发，帮助Retro改造重编程因子，在实验里相关指标提升了50倍以上。为了支撑这场豪赌，Retro已经喊出10亿美元融资目标，希望从一家「实验室里的梦想公司」，变成能挑战Bezos支持的Altos Labs的对手传统药只会拖慢？Retro想按下逆龄键过去两年，阿尔茨海默症领域最大的突破，是百健/Eisai的Leqembi（lecanemab）和礼来的Kisunla（donanemab）先后获批。它们的原理都是清除大脑里的β-淀粉样蛋白，临床数据显示能在早期阶段放慢认知衰退的速度。但问题是：药物需要静脉输注，还伴随脑部影像监测，价格高昂，患者体验并不轻松。传统阿尔茨海默症药物的靶点——通过清除大脑中的β-淀粉样蛋白（Aβ）斑块，来延缓认知衰退。 Retro的RTR242则选择了一条完全不同的路。它直接进入细胞内部，唤醒被年龄拖垮的自噬系统。系统一旦重新启动，就能把堆积的错误折叠和损坏蛋白清理掉，让神经细胞像电脑重启后那样恢复活力。 CEO Joe Betts-LaCroix直言：这些只能延缓的药，简直「不够劲儿」。 Retro的赌注是更激进的叙事：衰退不是必然，只要找到办法，衰老和疾病可以被切开。相比之下，传统药企还在「延缓」，Retro已经在谈「逆转」。当GPT走进实验室：AI下场炼青春 Retro并不是孤军奋战，他们背后还有奥特曼的另一张王牌——OpenAI。就在今年8月，OpenAI公布了一项和Retro的合作成果：利用定制版GPT-4b micro 来辅助蛋白质与重编程因子的设计。在实验室里，这些AI生成的方案，让关键的干细胞标记表达量提升了50倍——这是传统方法难以想象的跃迁。这正是Retro与传统药企最大的不同：前者试图让AI直接介入生命科学的根本环节，成为研发引擎，而后者依然依靠传统的药物靶点与临床试错。 Retro与OpenAI合作改造重编程因子实验。右图的细胞（第10天）相比传统方法，重编程效率显著提升，验证了AI在长寿药物研发中的加速潜力。这意味着什么？意味着 GPT 不再只是陪你聊天、写代码，它已经走进了显微镜下，开始直接改写生命科学的实验结果。AI从内容工厂，变成了蛋白质设计工厂。如果说Leqembi和Kisunla代表的是制药行业的安全牌，那么Retro正在用AI打开另一条路：一个跨界的未来——AI不只生成语言，还能生成青春。长寿赛道豪赌：Altman对撞Bezos 长寿已经成了硅谷资本的新战场，动辄上十亿美元的豪赌不断涌现。Retro最强对手是Jeff Bezos押注的Altos Labs。这家公司从2021年成立，就一次性筹到超过30亿美元，投资阵容包括亿万富豪Yuri Milner、Palantir联合创始人Joe Lonsdale，以及Arch Venture的Robert Nelson。相比之下，Retro目前只拿到奥特曼的1.8亿美元首投，但计划在接下来募资10亿美元，才能真正进入与Altos同级别的战场。这就是硅谷的长寿格局：一个是稳扎稳打的「国家队」，一个是激进冒险的「AI先锋」。未来谁能真正跑通，不仅决定哪家公司能收割市场，也关系着人类能否第一次在医学史上，把衰老当成可治疗的「病」。从ChatGPT到算力帝国，奥特曼已经在科技史上留下浓墨重彩的一笔。可他显然不满足于此，又在长寿赛道丢下了一枚药丸赌注——要把「衰老」和「衰退」切开。一边是OpenAI用GPT重塑知识与产业，一边是Retro用AI做实验，试图让大脑逆龄十年。这两条线看似毫无关系，却可能在未来交汇成一个全新故事：AI不仅改变工作方式，还可能改写我们的生命轨迹。问题是，这颗药丸会不会成为第二个 ChatGPT，真的颠覆医学格局？还是只是一场被资本和想象力裹挟的硅谷泡影？答案，还得等第一位病人吞下那颗药丸之后。参考资料： https://x.com/kimmonismus/status/1967508290775973965 https://www.businessinsider.com/retro-biosciences-sam-altman-antiaging-brain-pill-longevity-healthspan-2025-9 https://openai.com/index/accelerating-life-sciences-research-with-retro-biosciences/?utm_source=chatgpt.com

专利侵权

2025-09-19

·EndPoints

The Endpoints 11: OpenAI-backed Chai Discovery’s brainiac quartet aims to design drugs on the computer

Joshua Meier found his way onto Zavain Dar’s list when he was still a senior at Harvard University. At a New York networking event for the tech company Hugging Face, Meier engaged in what can only be described as small talk for a select group of biotechies: a conversation about building DNA-encoded libraries to represent molecules in latent space with SMILES strings, Dar recalled. Meier left Dar, then a partner at Lux Capital, with an impression. “He jumped off the page in a room full of hardcore ML devs, before any of this stuff was in vogue,” Dar said in an interview. Dar added Meier to a running list of promising talent whom he saw as “future entrepreneurs.” And by 2024, Dar was running his own VC firm, Dimension Capital, and Meier had just left his role as Absci’s chief AI officer to launch a startup. Dar wrote a $400,000 check before the new company had a name or a fully fleshed-out idea of what it would do. Less than two years in, Meier, 29, is one of four young and ambitious co-founders of Chai Discovery, an AI biotech startup. In June, they debuted their Chai-2 suite of models , producing de novo proteins with surprisingly high success rates. Work is underway on Chai-3. Since then, Chai has reeled in more than Dar’s first check. The startup closed a $30 million seed round last year, which included OpenAI and Thrive Capital, and a $70 million Series A round this year, led by Menlo Ventures. The latest raise valued Chai at $550 million, according to a source familiar with the matter. Chai is the latest in a long line of ventures reflecting Silicon Valley’s increasing influence on drug discovery and its bet that AI will reverse biopharma’s high frequency of clinical failure and low rates of R&D productivity. “We want to make biology look more like engineering than it looks like science,” Meier said in an interview. “There’s no reason to believe the capabilities should plateau here. I think they’re only going to keep getting better.” Chai-2 is an early proof point. Mikael Dolsten, the recently retired chief scientific officer of Pfizer, said he was blown away by the model’s results. He joined the company’s board, believing Chai-2 will soon be part of every biopharma’s R&D toolkit, and the company’s research can unlock new functions for antibodies. Chai is built more like a software startup than a biotech. With about 12 employees today, it doesn’t have its own laboratory or racks of computing chips. They occupy just 5,000 square feet in the San Francisco Bay Area. The company is not yet ready to discuss its business model or whether it will build its own pipeline. It’s a radically different approach to AI in biotech, compared with the hiring sprees and megaround raises of rivals like Xaira Therapeutics , Generate:Biomedicines and Isomorphic Labs . “I’m generally skeptical of large amalgamations of capital with scatterplots of big egos put into singular offices together and asked to figure it out,” Dar said. “I don’t think that’s a good way for efficient value creation, and value capture in any kind of entrepreneurial ecosystem.” Chai believes its small team will outcompete larger groups. They’re moving at what Dar called a “full-speed sprint,” as shown by releasing Chai-2 less than a year after Chai-1. That places a premium on the talent they bring in. Co-founders include Matthew McPartlon, 31, formerly at VantAI and one of the first builders of deep learning models in protein design; and Jacques Boitreaud, 27, previously a machine learning lead at the French biotech Aqemia. Chai has recruited some leading AI bio researchers, including Zhuoran Qiao , who developed several of Iambic Therapeutics’ top AI models, and Nathan Rollins, previously a machine learning scientist at Seismic Therapeutic. Chai’s holy grail is zero-shot drug candidates, or AI-generated molecules that are already optimized and could be ready for human testing. Co-founder and president Jack Dent said he believes achieving that is “a lot sooner” than five years out. For what’s next, Dar said he sees a long-term future for AI in predicting more drug properties and generating more types of drugs beyond protein-based therapeutics. “Right now, it’s high-affinity binders. Tomorrow, it’s PK/PD and ADME, all done in silico ,” Dar said, referencing a drug’s characteristics of pharmacokinetics, pharmacodynamics, absorption, distribution, metabolism, and excretion. “And then, the expansion across modalities. Today, it’s minibinders and antibodies. Tomorrow, it’s multispecifics and the next day it’s macrocycles and glues.”