LOGOS

A protein's ligand-binding pocket is made of residues that sit close together in space but can be hundreds of positions apart in the chain, which is why nearly every model reasoning about pockets takes 3D coordinates and processes them with a geometric network. LOGOS (Language Of Generative Objects in Science) takes a different route: it marks pocket residues in place inside the linear amino acid sequence with boundary tokens, then expands each one into the SMILES string of its side chain. Interface and ligand end up in one chemical alphabet and one token stream, and pocket-conditioned ligand design becomes next-token prediction.

That trick generalizes into a unified "scientific grammar": paired boundary tokens for each entity type plus relational tokens for reactions, transformations and search, nesting inside one another so that a metal-organic framework is a metal cluster wrapped around reused small-molecule spans. Pre-training and the downstream tasks then share one space, leaving no representational gap at fine-tuning time.

LOGOS came out of Alibaba Group and the Gaoling School of Artificial Intelligence at Renmin University of China in June 2026. It is a natural-sciences model rather than a purely biological one: two of its seven pre-training modalities (reactions, metal-organic frameworks) and two of its six tasks (retrosynthesis, material generation) sit outside biology, with proteins, antibodies, small molecules, binding sites and protein-ligand complexes making up the rest. It positions itself against NatureLM, arguing that natural language as the cross-domain interface spends capacity on linguistic ability domain-native tasks do not reward.

#Key Features

• Tokenized spatial interactions: Pocket-ligand contacts are discretized into tokens rather than modeled as continuous geometry, so a plain autoregressive transformer captures structural interactions with no coordinates and no equivariant layers.
• Sequence-only pocket search: Binding sites are proposed from primary sequence alone, extending the task from proteins with solved structures to the far larger set without one.
• Transfer to unseen task formats: Protein fitness editing and antibody CDR design appear nowhere in the pre-training grammar, yet both are reached through a task token and light fine-tuning.
• Three scales on stock LLM backbones: The 8B model continues pre-training from Qwen3-8B and the 1B and 3B from Llama-3.2, keeping each backbone's vocabulary and adding only the grammar's special tokens.

#Technical Details

The corpus totals 44.87 billion tokens across seven modalities, drawn from UniRef90 proteins, Observed Antibody Space sequences clustered at 70% identity, PubChemQC SMILES, Open Reaction Database and ECReact reactions, a hypothetical metal-organic framework library, P2Rank-predicted pockets on PDB structures, and Q-BioLiP protein-ligand complexes sampled at 5-10 Å radii. Training runs on 32 NVIDIA A800 GPUs in two stages: continued pre-training for 100,000 steps, then supervised fine-tuning mixed across four tasks. Backbone and embedding weights are inherited from the source LLM and only the new special-token embeddings are randomly initialized, which the ablations show converges faster than training from scratch.

On the PDBBind v2016 core set, LOGOS-8B reaches a mean Vina score of −7.76 with QED 0.57, ahead of coordinate-based generators including TargetDiff (−7.13) and TamGen (−7.38); LOGOS-1B reaches −7.64 with roughly 1/56 the parameters of NatureLM 8×7B (−6.91). Pocket identification hits Top-n 66.5 on COACH420 and 58.5 on HOLO4K, second only to P2Rank, whose predictions supplied the pre-training annotations and so bound the score. On GFP and AAV fitness editing it reaches 0.69–0.93 normalized fitness against 0.29–0.35 for the strongest prior method. Against inverse-folding baselines including ProteinMPNN and ESM-IF1, LOGOS-8B leads on CDR-H1/H2 and CDR-L1/L2 amino acid recovery (79.8% on CDR-H1) but trails structure-conditioned RADAb on CDR-H3 (47.0% versus 57.0%), the loop least predictable from framework sequence.

#Applications

The pocket-conditioned generator suits drug design teams working from a sequence and a predicted pocket rather than a co-crystal, and since the same checkpoint also proposes binding sites, hit generation and site selection need not be two tools. Protein engineers can use the editing interface to push fitness landscapes toward higher-activity variants, and antibody groups can fill framework regions with CDR1/CDR2 sequences at recovery competitive with structure-based inverse folding. Retrosynthesis and MOF generation come from the same weights, making it one dependency for projects crossing the biology-chemistry line.

#Impact

LOGOS's broader argument is infrastructural: scientific foundation models should be continued pre-trains of general LLMs, sharing their architectures, training stacks and inference servers rather than a parallel stack — with a 1B model beating a 46.7B mixture-of-experts on ligand design as the evidence. Several caveats travel with it. Nucleic acids are absent, no genomic or transcriptomic modality being in the grammar. Evaluation is entirely in silico, docking scores and refolded-structure RMSD standing in for experiment. The work is an arXiv technical report, not peer reviewed. Weights for all three scales are on HuggingFace under CC BY 4.0, but the Apache-2.0 repository carries task inference scripts rather than the training or data-processing code the paper describes.