InstructPro2: Natural Language and Biological Grounding Guided De Novo Protein Design
InstructPro2 tackles de novo protein design by translating open-ended natural language instructions into amino acid sequences while incorporating biological context such as target ligands or antigens. The framework augments a large protein language model with text and modality encoders, using lightweight gated cross-attention adapters to condition generation on both semantic and molecular signals. If the approach holds up experimentally, it could replace narrow task-specific models with a single generalist system that designers steer through plain text.
The paper presents an architecturally clean and well-motivated method for multimodal protein design, backed by strong in-silico gains on structural and domain-recovery benchmarks. Yet its claims about functional design remain largely unvalidated: every evaluation is computational, the training text is synthetic rather than human-authored, and the biological-encoder ablations show only marginal or inconsistent benefits. The work is a solid algorithmic advance but should be read as a generative modeling result, not as evidence of genuine biochemical function.
The integration of biological grounding via zero-initialized gated cross-attention adapters is a pragmatic way to preserve pretrained protein-generation behavior while learning to read new modalities. The two-stage training strategy and the large-scale pretraining corpus yield convincing quantitative improvements: InstructPro2 cuts global RMSD to 3.04Å and lifts functional-recovery F1 to 74.8, well ahead of Pinal and ESM3. The authors are also commendably transparent about their synthetic-data pipeline and the absence of wet-lab validation.
First, the lack of experimental validation is a critical gap: DockQ scores of 0.39 are merely acceptable, not high-quality, and InterProScan F1 measures only domain presence, not biochemical activity. Second, the 205.7 million pretraining instructions are synthesized by composing LLM-generated summaries of InterPro entries via randomized templates; this is not a distribution of real human design requests, and the process risks propagating LLM hallucinations into the protein-generation objective. Third, the biological-encoder ablations show that the added modality pathway provides little benefit for antibodies (DockQ 0.41 vs. 0.39 without) and only modest gains on seen ligands, suggesting the text prompt already encodes most of the necessary target information and undercutting the central claim that explicit biological grounding is essential.
The reported metrics are internally consistent and favor InstructPro2, but baseline alignment is uneven. ProGen2 is prompted with the first 32 residues of the ground-truth sequence and knows the exact target length, yet it receives no functional prompt at all, making the RMSD comparison a mismatch of conditioning paradigms. ESM3 is fed structured InterPro annotations plus residue ranges rather than free text, and its F1 is computed on a favorable subset where every requested entry is accepted. In the ligand task, the strongest competitor is the authors' own prior model InstructPro-3B, which actually attains a better GNINA-predicted affinity (-7.64 vs. -7.05) despite worse pocket placement. Direct comparisons against leading structure-based atomic design methods such as RFdiffusion3 are also absent, leaving the practical standing of language-first design unclear.
Code availability and detailed hyperparameter tables (optimizer settings, adapter placement, learning-rate schedules) support reproduction in principle, but several barriers remain formidable. Pretraining required 120 NVIDIA B200 GPUs, placing it beyond most academic labs. The training corpus is synthesized with specific closed-model versions—Gemini 3.1 Pro for InterPro summaries and Claude Haiku 4.5 for antibody blueprints—that are neither bit-reproducible nor guaranteed to remain accessible, so regenerating the exact pretraining distribution may be impossible. Evaluation further depends on a chain of external predictors (ESMFold, AlphaFold3, InterProScan, GNINA) whose versions and configurations must be exactly matched, yet random seeds and complete training logs are not provided.
The ability to generate proteins with desired functions is a central goal of computational biology. Although natural language can express diverse design objectives, existing protein generation methods often support only narrow tasks or make limited use of biological constraints. We develop InstructPro2 to translate natural-language design objectives into protein sequences while incorporating task-specific biological context. A text encoder captures the design instructions, while dedicated ligand and antigen encoders provide biological context to the sequence-generating decoder. For pretraining, we assemble InstructPro2Bench, comprising 205.7 million <function description, protein> pairs from UniProtKB. We subsequently adapt the model to ligand-binding protein and antibody design. On functional protein generation benchmarks, InstructPro2 outperforms Pinal, the strongest baseline, achieving a functional recovery F1 of 74.8 versus 62.2 and lowering global RMSD from 5.33A to 3.04A. For antibody design, the model obtains a DockQ score of 0.39, while ligand-binding protein design reaches a docking success rate of 72.6%. These findings highlight how integrating language-based instructions with biological conditioning can extend protein generation to distinct design settings.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.