InstructPro

Designing a protein that binds a chosen small molecule is bottlenecked by data: experimentally solved protein–ligand complexes number in the tens of thousands, far too few to train a large structure-conditioned generator. InstructPro, from Carnegie Mellon University and Lambda, sidesteps that bottleneck by conditioning on a different and far more abundant signal — the free-text functional annotations that UniProt already attaches to millions of proteins. Instead of specifying a binding pocket geometry, the user writes what the protein should do and supplies the ligand as a SMILES string, and the model emits an amino acid sequence.

To make that training regime possible the authors assembled InstructProBench, a corpus of 9,592,829 (function description, ligand, protein) triples drawn from UniProtKB entries that carry both a function annotation and a bound ligand, with ligand structures resolved through ChEBI. Sequences were clustered at 30% identity with MMseqs2 and split by cluster, and a separate 127-example test set holds ligands the model never saw during training, which is what makes the zero-shot claims meaningful rather than a memorization check.

InstructPro was posted as a preprint in June 2025 by Zhenqiao Song, Ramith Hettiarachchi, Chuan Li, Jianwen Xie, and Lei Li. Two variants were trained and both sets of weights, plus the dataset, are distributed from the project's GitHub repository.

#Key Features

• Language-conditioned design: A protein is specified by a sentence describing its intended function rather than by a target backbone or pocket, which lets the model draw on annotation text that vastly outnumbers solved complexes.
• Explicit ligand channel: The target ligand enters through a dedicated SMILES encoder, so the same functional instruction can be steered toward different chemical partners.
• Generalization to unseen ligands: On ligands absent from training the 3B model still reaches an AlphaFold3 ipTM of 0.882, with sequence novelty near 69%, indicating designs are not retrieved training neighbors.
• A reusable benchmark: InstructProBench supplies 9.6M triples with cluster-based splits and a held-out unseen-ligand set, giving the instruction-guided design task a standard evaluation.

#Technical Details

InstructPro is a four-part encoder–decoder. A 12-layer, 12-head, 768-dimensional text encoder initialized from PubMedBERT reads the function description; a 6-layer SMILES encoder initialized from a chemical RoBERTa reads the ligand; a 64-element memory module distills the text representation to a fixed-size summary; and an autoregressive protein decoder generates the sequence. The decoder is initialized from ProGen2 — ProGen2-base for the 27-layer, 1,536-dimensional InstructPro-1B (964M parameters total) and ProGen2-BFD90 for the 32-layer, 2,560-dimensional InstructPro-3B (2.98B total). Training ran for 1,000,000 steps on eight NVIDIA H100 GPUs. Designs are scored by folding the complex with AlphaFold 3 and measuring ipTM, plus Gnina binding affinity, MM/GBSA binding free energy, hydrogen bond counts, pLDDT, and novelty. On seen ligands InstructPro-1B reaches ipTM 0.918 and affinity -8.764; on unseen ligands the 1B and 3B models reach ipTM 0.869 and 0.882 with binding free energies of -20.9 and -25.8 kcal/mol. Baselines include Pinal, ESM-3, ProGen2, and an AlphaFold 3 plus LigandMPNN pipeline.

#Applications

The intended user is a designer who knows the chemistry of the molecule they want bound and can describe the desired activity in words — a biosensor for a metabolite, a scaffold for a cofactor-dependent enzyme, or a sequestration protein for a specific ligand — but has no experimentally solved complex to work from. Because generation is conditioned on text and SMILES rather than on a structure, the workflow starts before any structural hypothesis exists, and candidate sequences can then be folded and docked in silico before ordering genes.

#Impact

InstructPro's contribution is the reframing: it treats the abundant natural-language record of protein function as the training signal for ligand-binding design, and shows that scaling from 1B to 3B parameters improves every reported metric, including on unseen ligands. The paired InstructProBench dataset lowers the barrier for others working on instruction-guided design. The results are entirely computational — ipTM, docking scores, and MM/GBSA energies are proxies, not measured binding — and no wet-lab validation is reported, so the practical hit rate remains open. The GitHub repository carries no license file, and the weights and data are hosted on Google Drive rather than a versioned model hub, which limits reuse.