project
LOGOS - Alibaba's first open-source unified scientific model
LOGOS (Language Of Generative Objects in Science) is the first unified scientific grammar and multi-domain scientific generation foundation open-sourced by Alibaba's ATH-Token Foundry in collaboration with Renmin University of China's Hillhouse Institute of Artificial Intelligence...
What is LOGOS?
LOGOS (Language Of Generative Objects in Science) is the first open-source, unified scientific grammar-based multi-domain scientific generation model developed by Alibaba's ATH-Token Foundry in collaboration with Renmin University of China's Hillhouse Institute of Artificial Intelligence. The model encodes heterogeneous scientific objects such as proteins, small molecules, materials, and antibodies into unified token sequences. Based on a pure sequence modeling paradigm, it consistently matches or surpasses domain-specific methods in six major scientific tasks, including pocket ligand generation, retrosynthesis prediction, protein editing, and materials generation. It outperforms 56-parameter models with only 1 parameter, validating the feasibility of a single model solving numerous scientific tasks.
Main functions of LOGOS
- Pocket conditional ligand generationGiven a protein pocket structure, generate small molecule drugs that can bind precisely, while simultaneously meeting the triple requirements of binding affinity, drug similarity, and synthetic feasibility.
- Inverse Synthesis PredictionGiven a target molecule, it predicts the precursor molecules required for synthesis, with a Top-1 accuracy of 74.8%, assisting chemists in designing synthetic routes.
- Pocket site identificationIt predicts protein binding sites based solely on amino acid sequences, without requiring 3D structural data, and achieves a Top-n accuracy of 58.5% on the HOLO4K dataset.
- MOF material generationNovel metal-organic framework materials were generated, with a novel building block (NBB) ratio of 17.78%, an improvement of 76% compared to the baseline.
- Protein editingThe score on the Hard Fitness task was 0.93, an improvement of 174% from the baseline of 0.34.
- Antibody CDR DesignThe antibody complementarity-determining region design achieved an AAR of 79.82%, surpassing structure-dependent reverse folding methods.
Technical Principles of LOGOS
- Unified "scientific grammar" + spatial interaction discretizationDesign a shared vocabulary to 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 the understanding of complex spatial interaction rules without explicit 3D coordinates.
- Form and goal perfectly alignedThe sequence form of the pre-trained data is equal to the input and output form of the downstream task, and the pre-trained next-token prediction is equal to the downstream conditional generation target, thus eliminating the gap between pre-training and downstream application.
- Cross-domain knowledge positive transferThe unified syntax enables data from fields such as biology, chemistry, and materials to be fed back into the same model framework for training. Experiments have shown that multi-task joint training is more effective than independent training, achieving a synergistic effect of "1+1>2".
Follow us on WeChat and reply with "open source",join inAI open source project discussion group
How to use LOGOS
- Obtain open source resourcesVisit HuggingFace (LOGOS-Hub) to download the model weights, or obtain the inference code from GitHub (LOGOS-Hub/LOGOS).
- LLM-based deploymentIt directly reuses mature engineering infrastructure such as vLLM inference acceleration and model quantization, without the need to build an independent technology stack.
- Downstream task invocationThe system formats scientific objects into token sequences using a unified syntax, and generates them through autoregression to perform tasks such as pocket ligand design and retrosynthetic prediction.
LOGOS's core advantages
- Pure sequence paradigm surpasses 3D modelIn the task of pocket ligand generation, the pure sequence method has for the first time outperformed diffusion models that rely on 3D coordinates, proving that spatial structure can be fully captured through sequence prediction.
- Extreme parameter efficiencyLOGOS-1B outperforms NatureLM on multiple tasks with 1/56 of the parameter count (1B vs 8×7B), demonstrating extremely high parameter efficiency.
- Zero 3D coordinate dependencyPocket site identification only requires amino acid sequences, eliminating the need for expensive and scarce 3D structural data, significantly lowering the barrier to drug discovery.
- Unified architecture reduces engineering costsIt shares architecture, training paradigm and inference infrastructure with LLM, and can directly reuse existing mature engineering ecosystems without the need for an independent technology stack.
LOGOS project address
- GitHub repository: https://github.com/LOGOS-Hub/LOGOS
- HuggingFace model libraryhttps://huggingface.co/LOGOS-Hub
- arXiv technical paper: https://arxiv.org/pdf/2606.16905
Comparison of LOGOS with similar competing products
| Comparison Dimensions | LOGOS | NatureLM |
|---|---|---|
| Core positioning | A multi-domain generative foundation model for a unified scientific grammar | Specialized language models for the molecular and protein fields |
| Parameter size | 1B / 3B / 8B | 8×7B (56B) |
| Modeling paradigm | Pure sequence autoregression, no 3D coordinates required. | Relying on 3D coordinates and specialized geometric networks |
| Task Coverage | Proteins, molecules, materials, reactions, antibodies, pocket recognition | Molecular generation, protein structure |
| Cross-domain migration | Unified syntax enables positive knowledge transfer | Independent modeling across different domains makes transfer difficult. |
| Parameter efficiency | The 1B parameter outperforms the 56B model on multiple tasks. | A larger number of parameters is needed to achieve the same performance. |
| Engineering Ecology | Reuse LLM infrastructure (vLLM, quantization, etc.) | Building a technology stack independent of LLM |
Application scenarios of LOGOS
- AI-assisted drug design: Generate candidate drug molecules targeting specific protein targets while optimizing binding affinity, drug similarity, and synthetic feasibility.
- Chemical synthesis route planningIt provides retrosynthetic route prediction for organic synthesis, with a Top-1 accuracy of 74.8%, reducing the cost of experimental trial and error.
- Protein engineeringTargeted editing of protein sequences to improve stability or function, achieving a Hard Fitness score of 0.93, used in enzyme engineering and biocatalyst development.
- New material discoveryThe invention generates novel MOF materials with a 76% increase in the proportion of novel building blocks, which are suitable for gas storage, separation and energy conversion.
- Antibody drug developmentDesigning high-affinity antibody CDR regions with an AAR of 79.82% accelerates the development of therapeutic antibodies.