Valhalla Technology英灵殿科技Valhalla
英灵殿科技 · VALHALLAVALHALLA TECHNOLOGY

给生物分子一个
世界模型。A world model
for biomolecules.

AlloDesign 是全模态生成式分子世界模型——在统一物理框架下实现蛋白质、DNA、RNA、小分子四大模态任意组合的从头设计。湿实验之前先问它一遍。AlloDesign is an all-modality generative molecular world model — de novo design of any combination across proteins, DNA, RNA, small molecules under a unified physics framework. Ask it before you run the wet lab.

累计融资近 5000 万美元~$50M raised across multiple rounds*AlloDesign · 环肽设计突破AlloDesign · cyclic peptide breakthrough*
分子设计Molecular Design·蛋白设计Protein Design·配体生成Ligand Generation·核酸设计Nucleic Acid Design·环肽设计Cyclic Peptide Design·全原子生成ALL-ATOM GENERATION·相互作用设计Interaction Design·部分扩散Partial Diffusion·位点指定Site Specification·基序脚手架Motif Scaffolding·分子设计Molecular Design·蛋白设计Protein Design·配体生成Ligand Generation·核酸设计Nucleic Acid Design·环肽设计Cyclic Peptide Design·全原子生成ALL-ATOM GENERATION·相互作用设计Interaction Design·部分扩散Partial Diffusion·位点指定Site Specification·基序脚手架Motif Scaffolding

为什么做这个Why we build this

做新药，慢。New drugs are slow.

一药十年，十亿美金。药物发现卡在反复的实验循环里——合成、测试、看结果、改设计、循环。One drug, ten years, a billion dollars. Drug discovery is stuck in a loop of repeated experiments — synthesize, test, read results, redesign, repeat.

我们觉得这条路径对绝大多数疾病来说太慢了。所以做了一个世界模型，让化学家在合成之前能先在模型里跑一遍。We think that path is too slow for most diseases. So we built a world model that lets chemists run the experiment in silico before ever touching a flask.

它不靠猜——它学的是规则。It doesn't guess — it learns the rules.

science AlloDesign Workspace
Confidence
0.94
ΔG (kcal/mol)
-9.4
BINDING AFFINITY TREND
ATOMS ANALYZED
24
+12% vs prior run
ALLODESIGN · 全模态分子世界模型ALLODESIGN · ALL-MODALITY WORLD MODEL

学的是分子相互作用的规则。Learning the rules of molecular interaction.

AlloDesign 是全模态生成式分子世界模型——在统一物理框架下实现蛋白质、DNA、RNA、小分子四大模态任意组合的从头设计与结构优化，覆盖全链条分子研发需求。AlloDesign is an all-modality generative molecular world model — de novo design and structural optimization of any combination across proteins, DNA, RNA, small molecules under a unified physics framework, covering the full molecular R&D pipeline.

online_prediction

蛋白设计Protein Design

基于 Pairformer 的几何感知编码器与扩散解码器，面向任意物种与靶点类型实现从头设计。支持柔性受体模式，突破传统刚性受体假设，在真实蛋白构象系综上完成高精度序列与结构联合优化。Built on Pairformer's geometry-aware encoder and diffusion decoder for de novo design across any species and target type. Supports flexible-receptor mode, breaking through traditional rigid-receptor assumptions to achieve high-precision joint sequence-structure optimization on real protein conformational ensembles.

auto_awesome

配体设计Ligand Design

基于 Modality Token 条件注入与自回归–扩散协同采样，在靶点口袋内直接生成高可合成性、高选择性小分子。支持口袋条件化生成与类药性联合优化，显著压缩从命中到先导的周期。Leveraging Modality Token conditioning and autoregressive-diffusion collaborative sampling, directly generates highly synthesizable, selective small molecules within target pockets. Supports pocket-conditioned generation and druggability joint optimization, significantly compressing the hit-to-lead cycle.

manage_search

核酸设计Nucleic Acid Design

将 RNA 适配体、siRNA 与质粒等核酸模态统一纳入同一生成框架，利用碱基配对与二级结构先验实现可控生成。面向递送、编辑与调控场景，提供与蛋白/小分子互补的第三类设计通道。Unifies RNA aptamers, siRNA, plasmids and other nucleic acid modalities into a single generative framework, leveraging base-pairing and secondary structure priors for controllable generation. Addresses delivery, editing, and regulation scenarios, providing a third design channel complementary to proteins and small molecules.

挑战The Challenge

分子 AI，长期困在模态孤岛里Molecular AI has been trapped in modality silos

AlphaFold 让 AI “看见”了蛋白质结构——但真实的生物靶点是多组分系统：蛋白、核酸、小分子药物协同作用。此前的模型按模态各自训练，彼此无法理解，每一个真实任务都要拼接互不兼容的工具。AlphaFold let AI “see” protein structure — but real biological targets are multi-component systems: proteins, nucleic acids, and small-molecule drugs acting together. Previous models were trained per modality, unable to understand one another, so every real task required stitching together incompatible tools.

但生命运行在统一的物理之上：在原子层面，所有分子相互作用都是电磁的。按模态切分模型是人为的分类，不是物理的分类——统一的物理需要统一的模型。AlloDesign 建立在三个想法之上：Yet life runs on unified physics: at the atomic level, all molecular interactions are electromagnetic. Slicing models by modality is a human classification, not a physical one — unified physics demands a unified model. AlloDesign is built on three ideas:

表示Representation

最小化学生成单元（MCGU）Minimum Chemical Generation Units (MCGU)

所有分子分解为共享的化学基元——跨模态共用的最小单元。蛋白、RNA、DNA、配体与肽，共用一套表示。All molecules decomposed into shared chemical primitives — the smallest units common across modalities. One representation for proteins, RNA, DNA, ligands, and peptides.

Token 化Tokenization

模态与单元 TokenModality & Unit Tokens

在共享基元之上，模态 token 与单元 token 编码类型特异信息——一套共享语法、带有专有词汇的分子“语言”。On top of shared primitives, modality and unit tokens encode type-specific information — a molecular “language” with shared grammar and specialized vocabulary.

架构Architecture

Pairformer + 全原子扩散Pairformer + Full-Atom Diffusion

Pairformer 模块学习成对相互作用信号；全原子扩散模块在空间约束下生成三维结构——先预测结构，再面向功能重设计。A Pairformer module learns pairwise interaction signals; a full-atom diffusion module generates 3D structures under spatial constraints — predict structure, then redesign toward function.

三种粒度，从整体设计到原子级控制Three granularities — from whole molecules to single atoms

给定任意一组分子伙伴作为上下文，AlloDesign 设计与之相互作用的新分子。Given any set of molecular partners as context, AlloDesign designs a new molecule that interacts with them.

G1 · 整个分子Whole Molecule

设计一个能结合靶点的完整新蛋白、RNA、DNA、小分子或环肽。对设计对象拥有完整的生成控制权。Design a complete new protein, RNA, DNA, small molecule, or cyclic peptide that binds a target. Full generative control over the designed entity.

G2 · 功能基序Functional Motif

在已有分子内设计特定的结合基序或活性位点——功能区域级别的定向编辑。Design a specific binding motif or active site within an existing molecule — targeted editing at the functional region level.

G3 · 指定原子Prompted Atoms

约束特定原子（如催化残基、官能团），围绕它们进行设计——细粒度、约束引导的生成。Constrain specific atoms (e.g., catalytic residues, functional groups) and design around them — fine-grained, constraint-guided generation.

性能Performance

性能基准与湿实验验证Benchmarks & Wet Lab Validation

计算指标在真实生物实验中得到验证——8 个靶点上实现 nM 至 pM 级结合亲和力。*Computational metrics confirmed in real biological experiments — nM to pM binding affinities across 8 validated targets.*

任务Task指标Metric基线BaselineAlloDesign提升Improvement
蛋白设计通量Protein design throughput候选分子数量（AME 测试集）Candidate volume (AME test set)RFDiffusion2—候选数多 20 倍20× more candidates*
蛋白设计通量Protein design throughput计算速度Computational speedRFDiffusion—快约 10 倍~10× faster*
RNA 设计RNA design设计成功率Design success rateRNAFrameFlow—高约 2 倍~2× higher*
蛋白–RNA 复合物（零样本）Protein–RNA complex (zero-shot)平均成功率Avg. success rate—77.9%*零样本迁移Zero-shot transfer
蛋白结合小分子Protein-binding small molecule设计成功率Design success rateSurfGen—提升超 40 倍>40× improvement*
湿实验验证Wet lab validation结合亲和力（8 个靶点）Binding affinity (8 targets)RFDiffusion / BindCraftnM – pM*提升数倍至 100 倍以上Several× to 100×+ improvement*

跨模态迁移——核心成果Cross-modal transfer — the key result

AlloDesign 在 RNA 和 DNA 上——训练数据远少于蛋白质的模态——的表现证明它学到的是可泛化的分子相互作用规则，而非蛋白质特有的模式。这种迁移能力是世界模型假说的核心论据。AlloDesign's performance on RNA and DNA — modalities with far less training data than proteins — demonstrates that it learned generalizable molecular interaction rules, not protein-specific patterns. This transfer is the core claim of the world model hypothesis.

完整论文Full paperarrow_forward

进展Progress

几个数，看进度。Progress, in numbers.

1.2B*
模型参数Model parameters
1018
可搜索结构Searchable structures
47×*
对比 DFT 基线vs DFT baseline
4
分子模态统一Unified modalities

时间线Timeline

十二个月，从论文到私测。Twelve months — from paper to private beta.

2025 · 09
跟某肿瘤药企的第一次 pilot 跑通。First pilot completed with an oncology pharma partner.*
2025 · 10
第一篇分子世界模型论文挂 arXiv。First molecular world model paper on arXiv.*
2026 · 01
三个月内连续完成多轮融资，累计近 5000 万美元。Multiple rounds closed in 3 months, ~$50M total.*
2026 · Q2
AlloDesign 环肽设计取得阶段性突破，商业化落地启动。AlloDesign cyclic peptide breakthrough; commercialization begins.*

我们的故事Our Story

师承 2024 年诺贝尔化学奖得主 David Baker 教授，创始人张昊天出身于被誉为全球 AI 蛋白质设计领域“黄埔军校”的 Baker Lab。在 AIDD 行业长期专注于单一模态研发的惯性之下，张昊天率先打破边界，提出并构建了“全模态分子设计”的全新范式——从分子物理底层规律出发，在统一物理框架下描述所有分子间的相互作用。Trained under 2024 Nobel Laureate Prof. David Baker at the Baker Lab — the “West Point” of AI protein design — founder Haotian Zhang broke the single-modality paradigm that dominated AIDD. He proposed and built the “all-modality molecular design” framework: describing all intermolecular interactions from first-principles physics under a unified framework.

基于这一原创思路，公司正在迭代研发新一代多模态生成式分子世界模型 AlloDesign，实现蛋白质、DNA、RNA、小分子四大模态任意组合的从头设计。配套自研高通量实验平台，打通分子设计、自动化合成与亲和力验证的完整链路，形成数据驱动的闭环研发体系。Building on this original insight, the company is iterating on AlloDesign — a next-generation multi-modal generative molecular world model enabling de novo design across any combination of proteins, DNA, RNA, small molecules. A proprietary high-throughput experimental platform closes the loop from design to automated synthesis to affinity validation, forming a data-driven R&D cycle.

愿景Vision

短期以药物发现作为首个落地场景，面向全球 MNC、Biotech 提供定制靶点分子研发与工业级模型授权。中长期搭建可自我迭代的 Scientific AGI 基础设施，落地自驱动智能实验室，打造可独立完成科学推演、实验迭代的虚拟科学家体系。Near-term: drug discovery as the first commercialization scenario, offering custom target molecular R&D and industrial-grade model licensing to global MNCs and Biotechs. Long-term: building self-iterating Scientific AGI infrastructure with autonomous labs — a virtual scientist system capable of independent scientific reasoning and experimental iteration.

“我们始终相信，AGI in Life Science 的整体技术路线一定是从单模态到多模态，从单任务到多任务。全模态模型并不是简单地把不同分子类型放进同一个模型，而是从底层物理规律出发，形成能够跨越分子模态、连接计算与实验的统一技术体系。”“We firmly believe the technical roadmap for AGI in Life Science must go from single-modality to multi-modality, from single-task to multi-task. An all-modality model is not simply putting different molecule types into one model — it is building a unified technical system that bridges molecular modalities and connects computation with experiment, grounded in first-principles physics.”

— 张昊天，创始人 & CEO— Haotian Zhang, Founder & CEO

机构支持Institutional Foundations

临港实验室Lingang Laboratory

上海 AI 创新策源地Shanghai AI Innovation Hub

浙江大学Zhejiang University

药学 / AI 联合培养博士Pharmacy / AI Joint PhD Program

香港中文大学CUHK

CS 联合培养博士 · 战略顾问 王平安教授CS Joint PhD · Strategic Advisor Prof. Ping'an Wang

投资方Backing

三个月内连续完成多轮融资，累计近 5000 万美元。投资方包括五源资本、真格基金、中科创星、襄禾资本、国方创投、陆石投资、奇绩创坛、L2F 光源创业者基金及知名产业方，光源资本担任独家财务顾问。Multiple rounds closed within 3 months, totaling ~$50M. Investors include 5Y Capital, ZhenFund, CAS Star, Xianghe Capital, Guofang Ventures, Lushi Investment, Qiji Chuangtan, L2F Capital, and strategic industry partners. Lighthouse Capital served as exclusive financial advisor.

路线图Roadmap

从原型到生产，再到自主进化From prototype to production to autonomy

AlloDesign 是地基；路线图在其上逐层构建，把全模态分子世界模型变成药物发现引擎与 Scientific AGI 基础设施。AlloDesign is the foundation; the roadmap builds the layers that turn an all-modality molecular world model into a drug discovery engine and Scientific AGI infrastructure.

01
现阶段Now

AI4Bio

AlloDesign 作为面向全球 MNC 与 Biotech 的服务：定制靶点分子研发、工业级模型授权与高通量湿实验协同。AlloDesign as a service for global MNCs and Biotechs: custom target molecular R&D, industrial-grade model licensing, and high-throughput wet lab coordination.

02
开发中In Development

AI4AI

自我迭代的科学智能层：从文献与实验中提取结构化知识，让模型能解释推理、归因失败、并提出下一步实验。A self-iterating scientific intelligence layer: structured knowledge from literature and experiments, so the model explains reasoning, attributes failures, and proposes next experiments.

03
愿景Vision

AI4Phy

自主实验室：设计 → 合成 → 测试 → 更新，在闭环中从真实实验反馈中持续进化。The autonomous laboratory: design → synthesize → test → update, improving from real experimental feedback in a closed loop.

招聘Careers

我们正在招人We're hiring

AI 算法、系统工程、分子设计、产品管理等方向均有开放职位。年轻团队，高影响力，快速成长。We have open roles in AI algorithms, systems engineering, molecular design, product management, and more. Young team, high impact, fast growth.

加入我们

这世纪的医学，
从分子开始写。This century's medicine
is written in molecules.

我们在融资、在招人。投资人、合作伙伴、想搞硬核问题的科学家——都欢迎发邮件。1-2 个工作日内回。We're fundraising and hiring. Investors, partners, scientists who want to tackle hard problems — all welcome to email us. We reply within 1-2 business days.