What is LOGOS
LOGOS (Language Of Generative Objects in Science) is the first open-source unified scientific grammar multi-domain scientific generative foundation model, jointly developed by Alibaba ATH-Token Foundry and Renmin University of China's Gaoling School of Artificial Intelligence. The model encodes heterogeneous scientific objects such as proteins, small molecules, materials, and antibodies into a unified Token sequence. Based on a pure sequence modeling paradigm, it consistently matches or surpasses domain-specific methods across six major scientific tasks, including pocket-ligand generation, retrosynthesis prediction, protein editing, and material generation. With 1B parameters, it surpasses 56B-parameter models, validating the feasibility of one model solving thousands of scientific tasks.
LOGOS's Main Features
- Pocket-conditioned ligand generation: Given a protein pocket structure, generate small-molecule drugs that bind precisely, while satisfying the triple requirements of binding affinity, drug-likeness, and synthetic feasibility.
- Retrosynthesis prediction: Given a target molecule, predict the precursor molecules required for synthesis, with Top-1 accuracy reaching 74.8%, assisting chemists in designing synthetic routes.
- Pocket site identification: Predict protein binding sites using only amino acid sequences, without requiring 3D structural data, achieving Top-n accuracy of 58.5% on the HOLO4K dataset.
- MOF material generation: Generate novel metal-organic framework materials, with the proportion of novel building blocks (NBB) reaching 17.78%, a 76% improvement over the baseline.
- Protein editing: Achieves a score of 0.93 on the Hard-difficulty Fitness task, a 174% improvement over the baseline of 0.34.
- Antibody CDR design: Antibody complementarity-determining region design achieves AAR of 79.82%, surpassing structure-dependent inverse folding methods.
LOGOS's Technical Principles
- Unified "scientific grammar" + spatial interaction discretization: Design a shared vocabulary to uniformly encode heterogeneous objects such as proteins, small molecules, and materials into discrete Token sequences; invent a textual description method to grammaticalize 3D spatial contact patterns, enabling understanding of complex spatial interaction rules without explicit 3D coordinates.
- Full alignment of form and objective: The sequence form of pretraining data equals the input-output form of downstream tasks; pretraining's next-token prediction equals downstream conditional generation objectives, eliminating the gap between pretraining and downstream applications.
- Positive cross-domain knowledge transfer: The unified grammar enables data from biology, chemistry, materials, and other domains to flow back for training within the same model framework. Experiments confirm that multi-task joint training outperforms independent training, achieving a "1+1>2" synergistic effect.
Follow WeChat and reply "Open Source" to joinAI Open Source Project Exchange Group
How to use LOGOS
- Obtain open-source resources: Visit HuggingFace (LOGOS-Hub) to download model weights, or obtain inference code from GitHub (LOGOS-Hub/LOGOS).
- Deploy based on the LLM ecosystem: Directly reuse mature engineering infrastructure such as vLLM inference acceleration and model quantization, without building an independent technology stack.
- Downstream task invocation: Format scientific objects into Token sequences according to the unified grammar, and complete tasks such as pocket-ligand design and retrosynthesis prediction through autoregressive generation.
LOGOS's Core Advantages
- Pure sequence paradigm surpasses 3D models: On pocket ligand generation tasks, a pure sequence method defeats diffusion models that rely on 3D coordinates for the first time, proving that spatial structure can be fully captured through sequence prediction.
- Extreme Parameter Efficiency: LOGOS-1B surpasses NatureLM on multiple tasks with 1/56 of the parameter count (1B vs 8×7B), demonstrating extremely high parameter efficiency.
- Zero 3D Coordinate Dependency: Pocket site identification requires only amino acid sequences, without the need for expensive and scarce 3D structural data, greatly lowering the barrier to drug discovery.
- Unified Architecture Reduces Engineering Costs: Shares architecture, training paradigm, and inference infrastructure with LLMs, allowing direct reuse of the existing mature engineering ecosystem without the need for an independent tech stack.
LOGOS Project Address
- GitHub Repository：https://github.com/LOGOS-Hub/LOGOS
- HuggingFace Model Hub：https://huggingface.co/LOGOS-Hub
- arXiv Technical Paper：https://arxiv.org/pdf/2606.16905
Comparison of LOGOS with Similar Competitors
| Comparison Dimension | LOGOS | NatureLM |
|---|---|---|
| Core Positioning | Multi-domain generative foundation model with unified scientific grammar | Domain-specific language models for molecules and proteins |
| Parameter Scale | 1B / 3B / 8B | 8×7B（56B） |
| Modeling Paradigm | Pure sequence autoregression, no 3D coordinates required | Relies on 3D coordinates and specialized geometric networks |
| Task Coverage | Proteins, molecules, materials, reactions, antibodies, pocket identification | Molecule generation, protein structure |
| Cross-domain Transfer | Unified grammar enables positive knowledge transfer | Independent modeling per domain, difficult to transfer |
| Parameter Efficiency | 1B parameters surpass 56B models on multiple tasks | Requires larger parameter counts to achieve equivalent performance |
| Engineering Ecosystem | Reuses LLM infrastructure (vLLM, quantization, etc.) | Builds a tech stack independent of LLMs |
LOGOS Application Scenarios
- AI-Assisted Drug Design: Generates candidate drug molecules targeting specific protein target pockets, while optimizing binding affinity, drug-likeness, and synthetic feasibility.
- Chemical Synthesis Route Planning: Provides retrosynthetic pathway prediction for organic synthesis, with Top-1 accuracy reaching 74.8%, reducing experimental trial-and-error costs.
- Protein Engineering Modification: Directionally edits protein sequences to enhance stability or function, with Hard Fitness score reaching 0.93, used for enzyme engineering and biocatalyst development.
- New Material Discovery: Generates novel MOF materials, with the proportion of novel building units increased by 76%, suitable for gas storage, separation, and energy conversion.
- Antibody Drug Development: Designs high-affinity antibody CDR regions, with AAR reaching 79.82%, accelerating therapeutic antibody R&D.