InstructBioMol

The token C is a letter in English, a carbon atom in a chemical string, and cysteine in a protein sequence. That collision is a small symptom of a larger mismatch: the instructions a researcher wants to give — "design an enzyme that acts on this substrate" — are prose, while the things being designed are intrinsically multimodal. A molecule is a 2D bond graph and a 3D conformer as well as a string; a protein is a sequence and a fold. InstructBioMol, from a team at Zhejiang University, addresses both halves at once: it expands a language model's vocabulary so molecules and proteins are first-class token types rather than ordinary text, and feeds the model fused structural features alongside the words.

What distinguishes it from earlier instruction-tuned biomolecular language models is the shape of its alignment. Prior work paired natural language with molecules or with proteins, leaving the molecule–protein axis uncovered. InstructBioMol trains all six directed pairings across the three modalities — text to molecule and back, text to protein and back, and protein to molecule and back — which is what makes the applied tasks expressible as single instructions. Naming a target protein and asking for binders is a protein-to-molecule query; naming a substrate and asking for a catalyst is a molecule-to-protein query.

The work appeared as a preprint in October 2024 and was published in Nature Machine Intelligence, with code, checkpoints and training data released alongside it.

#Key Features

• Any-to-any modality alignment: A single set of weights handles molecule captioning, description-based molecule and protein generation, protein property answering, target-based drug design and substrate-based enzyme design without per-task refitting.
• Motif-guided feature extraction: Frozen encoders — a five-layer GIN over 2D molecular graphs, Geoformer over 3D conformers, ESM-2 (35M) over protein sequence and SaProt (35M) over backbone structure — feed a transformer encoder–decoder whose decoder is primed with a motif prompt and learnable queries, so fusion is anchored to functional substructures rather than raw geometry.
• Expanded biomolecular vocabulary: SELFIES bracket groups become molecule tokens and prefixed amino acids become protein tokens, delimited by explicit modality markers, which prevents the lexical ambiguity between English, chemistry and sequence alphabets.
• Staged instruction tuning: A first stage learns from tens of millions of coarse pairs, a second narrows to curated data, trading breadth for quality rather than mixing the two.
• Joint protein–ligand output: A LoRA adapter extends the model to emit a protein and its binding molecule from one textual description.

#Technical Details

The backbone is Llama-2-7B, vocabulary-expanded to 32,985 tokens with a 4,096-token context; the released checkpoint is roughly 6.9B parameters. Continual pretraining covers about 100M molecules from PubChem, 59M protein sequences from UniProt and 6M scientific texts from PubMed, bioRxiv and ChemRxiv. Instruction tuning stage 1 uses 60M molecule–text pairs (PubChem, ChEBI), 35M protein–text pairs (Swiss-Prot, TrEMBL) and 1M molecule–protein pairs (BindingDB, Rhea); stage 2 narrows to 52K ChEBI, 2M Swiss-Prot and 1M BindingDB/Rhea examples. Training ran on eight 80GB H800 GPUs.

On molecule captioning it reaches BLEU-4 59.3 and METEOR 69.1; description-based molecule generation reaches 52.9 exact match at 99.0% validity. Protein property answering improves on the previous best by 13.5% on average, and generated proteins reach 21.9% identity to the reference. For target-based drug design over 100 test proteins it averages a Vina score of −8.27 with 68.9% high-affinity hits and a 49.1% success rate. Designed enzymes score 70.4 on the ESP enzyme–substrate interaction predictor, above the 60.0 threshold its authors recommend as evidence of a genuine pairing. Docked poses in both design evaluations come from DiffDock.

#Applications

The intended use is as a research copilot for medicinal chemistry and biocatalysis: proposing hit molecules against a named target, proposing enzyme sequences for a named substrate, answering questions about a protein's family, localization and function, editing a molecule toward a stated property, and drafting a protein–ligand pair from a description. All reported design results are computational — binding scored by docking, enzyme–substrate compatibility by a learned predictor, with no wet-lab validation — so the outputs feed an experimental triage step rather than standing on their own.

#Impact

InstructBioMol makes a case that the interface to biomolecular design can be ordinary language without giving up structural input, and it does so with an unusually complete release: MIT code and weights for all three training stages, plus the instruction corpora and auxiliary encoder checkpoints. Its limits are stated plainly by its authors. Compute constraints kept DNA and RNA out of the vocabulary, and tasks outside its instruction mix — chemical reaction prediction, for instance — are not supported. The architecture is built for extension: additional modality encoders and vocabulary entries are the mechanism by which those gaps would close.