LOGOS

LOGOS (Language of Generative Objects in Science) is a multi-domain generative framework from Alibaba Tongyi Lab. It encodes proteins, antibodies, small molecules, chemical reactions, materials and their spatial interactions as token sequences over a shared vocabulary, so a single autoregressive model can generate, predict and design across several natural-science domains.

It works on domain-native representations rather than natural language as an intermediary. Spatial relationships such as protein pocket–ligand contacts are discretized into tokens, so the model does not need geometric neural networks or explicit 3D coordinates. According to a YouTube creator summary, it was released in sizes from 1B to 8B under Apache 2.0; this is not confirmed in the provided README text.

The GitHub repository provides inference scripts for four tasks: retrosynthesis prediction, protein ligand-binding site identification, interaction-aware ligand design for binding pockets, and unconditional material generation. Running them needs a CUDA GPU, the NVIDIA PyTorch Docker image and a model checkpoint from Hugging Face. The README's quick-start still uses a placeholder model path, and no Hugging Face or paper link was provided, so checkpoint availability should be verified.