Source-linked AI summary
InstructPro2: Natural Language and Biological Grounding Guided De Novo Protein Design
Charles Novak, Zhenqiao Song, Chuan Li, Jianwen Xie, Lei Li
TL;DR
Protein generation methods often support narrow tasks or make limited use of biological constraints, despite the range of functions design could target. InstructPro2 conditions protein generation on natural-language instructions and biological context, and reports benchmark gains alongside transfer to antibody and ligand-binding design.
Problem
Method
InstructPro2 conditions a protein decoder on natural-language instructions and biological context from dedicated ligand and antigen encoders, pretrained on 205.7 million function-description and protein pairs.
Source: §1 · pp. 1–2Results
Takeaways & Limitations
The same decoder transfers to real-world antibody and ligand-binding protein design tasks.
Source: §6 · p. 9Evaluation is computational throughout; experimental validation of the designed proteins remains an important direction for future work.
Source: §6 · p. 9
Abstract
from arXiv · showhide
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.
1. Introduction
InstructPro2 combines natural-language instructions with biological context to generate functional proteins, addressing the narrow scope and limited grounding of existing methods. The introduction reports benchmark gains and transfer to antibody and ligand-binding design.
1. Introduction
- Existing protein generation methods have added structural, functional-tag, and natural-language conditioning, but remain limited in scope or biological grounding.The introduction notes that computational design can avoid reliance on slow evolution and labor-intensive high-throughput screening.Source: §1 · p. 1
- The framework pairs a natural-language text encoder and dedicated ligand and antigen encoders with a protein decoder conditioned on both modalities.Its two-stage training first pretrains on instruction-following tasks, then incorporates biological context for downstream tasks.Source: §1 · p. 1
- The model is designed to generate diverse functional proteins directly from human natural-language instructions.Source: §1 · p. 1
- InstructPro2 achieves 3.04 Å global RMSD versus 5.33 Å and functional recovery F1 of 74.8 versus 62.2, outperforming strong baselines including Pinal.The model also reaches a DockQ score of 0.39 on antibody design and a 72.6% docking success rate on ligand-binding design.Source: §1 · p. 2
- InstructPro2Bench provides 205.7 million function-description and protein pairs across diverse families for model pretraining.The dataset is assembled from UniProtKB.Source: §1 · p. 2
2. Related Work
Protein generation has progressed from unconditioned sequence models to methods guided by annotations, structures, binding partners, and natural-language instructions. Generalizable design for arbitrary user-defined functions remains an open challenge.
Functional Protein Design
- Early sequence models generated diverse, well-folded proteins without explicit conditioning, while later work steered generation using attributes, structures, or binding targets.Source: §2 · p. 2
Natural Language-Guided Protein Design
- Natural-language design approaches include mapping text to intermediate structures, learning joint protein–text embeddings, and generating from open-ended instructions.Source: §2 · p. 2
- Existing natural-language methods remain restricted to narrow tasks and limited real-world scenarios, leaving arbitrary user-defined, biologically grounded design an open challenge.Source: §2 · p. 2
3. Proposed Method: InstructPro2
InstructPro2 models protein design as generating a sequence from a natural-language function description and biological context. Its method combines text and modality conditioning with a large instruction-following pretraining corpus and datasets for antibody and ligand-binding design.
3.1. Problem Definition
- The model learns to generate protein sequences from natural-language descriptions while satisfying biological context C, such as a target ligand or antigen.Its objective is to model pθ(r | w, C), where r is a protein sequence and w is the functional description.Source: §3.1 · p. 2
3.2. Model Architecture
- A text encoder supplies instruction representations to the protein decoder, while fine-tuned ligand or antigen encoders provide biological context through cross-attention adapters.The adapters are zero-initialized; the text and modality encoders are frozen during fine-tuning while the other listed components remain trainable.Source: §3.2 · pp. 2–3
3.3. InstructPro2Bench Construction
- InstructPro2Bench contains 205,727,986 protein-text pairs, with descriptions assembled from InterPro annotations and selected to capture distinct protein functions.Descriptions use up to three entries, prioritizing specific annotations and removing redundant selections; training combines autoregressive generation and masked infilling.Source: §3.3 · p. 4
3.4. Real-World Task Data Curation
- Antibody data combine solved antibody-antigen complexes with additional antibody examples lacking paired antigen information, and splits cluster examples by CDR-H3 identity.Ligand design pairs protein sequences with ligand SMILES and function descriptions; its held-out test sets include seen and unseen ligands.Source: §3.4 · pp. 4–5
4. Experiments
The experiments compare InstructPro2 with language-guided, unconditional, and task-specific baselines across pretrained protein generation, antibody design, and ligand-binding design. Results cover structural accuracy, function, and binding, including evaluation on ligands unseen during training.
4.1. Experimental Setup
- The evaluation compares language-guided, unconditional, and task-specific baselines using functional, binding, reliability, and structural-accuracy metrics.The ligand comparison scores methods on targets completed by every method.Source: §4.1 · pp. 5–6
4.2. How good is the pretrained InstructPro2?
- 3.04 Å global RMSD and 74.8 functional recovery F1: InstructPro2 outperforms the strongest baseline, which scores 5.33 Å and 62.2, respectively.It also achieves the highest sub-2 Å accuracy, 63.5% versus 43.5% for the next-best model.Source: Table 1 · p. 6, §4.2 · p. 6
4.3. Can InstructPro2 design high-quality antibodies?
- 0.39 DockQ: InstructPro2-Ab leads the compared models, narrowly ahead of antibody-specific RefineGNN at 0.38.The fine-tuned model generates all three heavy-chain CDR loops conditioned on an antigen and a natural-language blueprint.Source: §4.3 · p. 6
4.4. Can InstructPro2 design ligand binding proteins?
- 2.24 Å pocket RMSD and 72.6% docking success: InstructPro2-Lb places pockets most accurately and succeeds on the most seen-ligand targets.InstructPro-3B has lower GNINA-predicted affinity (−7.64 against −7.05), but less accurate pocket placement and lower success (67.9% against 72.6%).Source: Table 3 · p. 7, §4.4 · p. 7
- 1.81 Å pocket RMSD: InstructPro2-Lb retains its pocket-accuracy and success-rate lead on ligands unseen during training.Pinal, ProGen2, and InstructPro-3B have pocket RMSDs of 2.56 Å, 2.47 Å, and 3.05 Å, respectively.Source: §4.4 · p. 7
5. Analysis
The analysis finds a trade-off between decoding strategies and a limited pocket-geometry benefit from explicit biological grounding. Case studies illustrate individual antibody and ligand-binding designs, with the authors noting that the examples were selected rather than random.
5.1. Effect of Different Sampling Strategies
- Greedy decoding reaches 63.5% global sub-2 Å accuracy, while nucleus sampling reaches 87.5 InterProScan F1; sampling also raises recall and precision.Sampling yields more varied sequences that match a single reference less closely but express requested function more often.Source: §5.1 · p. 7
5.2. Influence of Biological Grounding
- 2.53 Å to 2.24 Å on seen ligands and 2.41 Å to 1.81 Å on unseen ligands: explicit ligand information reduces pocket RMSD, with a larger reduction on the small unseen set (N = 124).On antibodies, neither text-only nor text-plus-modality conditioning holds a consistent advantage.Source: §5.2 · p. 7
5.3. Case Study
- In selected examples, InstructPro2-Ab scores 0.883 DockQ on COVOX-269, while InstructPro2-Lb reaches 0.09 Å pocket RMSD on unseen THB HUMAN.The authors caution that these are selected examples, not a random sample.Source: §5.3 · p. 7
6. Conclusion
InstructPro2 combines natural-language instructions with biological grounding and applies the same decoder to antibody and ligand-binding design. It reports benchmark gains, but its evaluation is computational and lacks wet-lab validation.
6. Conclusion
- Table 4 evaluates unseen-ligand design on 126 shared targets; InstructPro2-Lb places pockets most accurately while matching the best baseline on success rate.The evaluation covers targets completed by every compared method.Source: Table 4 · p. 8
- Figure 3 compares greedy decoding with nucleus sampling and evaluates antibody and ligand-binding models with and without auxiliary modality encoders.The caption specifies nucleus sampling at p = 0.95 and τ = 0.8 for the pretraining benchmark panel.Source: Figure 3 · p. 8
- Figure 4 presents antibody and ligand-binding design cases, with designed structures in blue, ground truth in orange, and antigen in gray.The caption reports strong binding to the partner in both cases.Source: Figure 4 · p. 8
- 3.04 Å global RMSD versus 5.33 Å and 74.8 functional recovery F1 versus 62.2 accompany reported antibody and ligand-binding design results.The conclusion reports 0.39 DockQ for antibody design and a 72.6% docking success rate for ligand-binding design.Source: §6 · p. 9
- The evaluation is computational throughout, and experimental validation of designed proteins remains future work.Source: §6 · p. 9
AI use statement
Generative AI assisted in producing synthetic dataset descriptions and polishing the writing; the authors state that they performed the other work themselves.
AI use statement
- Large language models helped produce natural-language descriptions for InstructPro2Bench and the Stage 2 antibody corpus, and also helped polish and paraphrase writing.Source: AI use statement · p. 9
- The authors state that all other work was done by them and that they take responsibility for the final content.Source: AI use statement · p. 9
A. Dataset Construction
The dataset pipeline composes protein descriptions from InterPro annotations rather than writing one description per protein. Pretraining examples also use span infilling alongside autoregressive generation, with the same localized-generation procedure used for antibody HCDR generation.
A.1. Stage 1: InterPro-Grounded Description Pipeline
- InterPro descriptions are assembled from summaries of vocabulary entries, so annotation effort scales with vocabulary size rather than the 205.7 million-protein corpus.The pipeline summarizes each of 49,677 entries once, selects up to three entries per protein across specificity tiers, and removes near-duplicates.Source: §A.1 · p. 12
- Selected InterPro summaries are joined with randomized templates for annotations, connectors, organism and keywords, then wrapped in a randomized instruction prompt.Source: §A.1 · p. 12
A.2. Span Masking and Infilling Construction
- Masked regions are divided into spans and replaced by distinct sentinel tokens, with the target recording the masked sequence and each sentinel–span pair.Span lengths are sampled from a Poisson distribution with λ = 20 and clipped to 3–32 amino acids.Source: §A.2 · p. 12
- Infilling prompts ask the model to recover missing spans using biological context, placing the functional description at varying positions in the instruction.Source: §A.2 · p. 12
- 50.0% of training examples use infilling; these have a mean corruption ratio of 0.339 and average 6.23 masked spans per sequence.Sequences shorter than 10 amino acids remain autoregressive; eligible sequences receive target masking lengths of 20%–40%.Source: §A.2 · p. 12
- The same span-infilling procedure is used for HCDR generation in Stage 2, providing a consistent mechanism for localized sequence generation across training stages.Source: §A.2 · pp. 12–13
A.3. Antibody–Antigen Dataset
The antibody corpus combines antigen-paired blueprints with supplemental heavy-chain examples, using curated source fields and clustered splits. The ligand dataset inherits benchmark pairs and splits, with function descriptions standardized across training stages.
A.3. Antibody–Antigen Dataset
- The Stage 2 antibody corpus contains 21,601 examples from antigen-paired design blueprints, a heavy-chain-only supplement, and an additional antigen-paired supplement.Table 5 gives the corpus composition.Source: §A.3 · p. 13, Table 5 · p. 13
- The primary pipeline enriches SAbDab complexes with antigen metadata and literature, then cleans chains, scaffolds and domains before generating structured descriptions.Domain labels require at least 80% overlap with a UniProt Domain or Region feature; engineered-construct status is flagged from purification-tag motifs.Source: §A.3 · p. 13
- Supplemental heavy-chain attributes come from verified fields rather than inference; only 1,535 of 15,124 processed rows have an explicit target.The pipeline manually reviewed a 400-example sample.Source: §A.3 · p. 13
- Antibody training targets CDR regions for infilling and assigns 40%-identity CDR-H3 clusters to train, validation and test in an 8:1:1 ratio.The resulting split sizes are 4,030, 433 and 444 examples.Source: §A.3 · p. 13
A.4. Ligand–Protein Dataset
- The ligand-conditioned dataset inherits InstructProBench pairs and splits, including 1,500 seen-ligand and 127 unseen-ligand test examples.Source: §A.4 · p. 13
- Ligand-design descriptions use the InterPro-grounded pipeline, holding the text pathway fixed across stages and making ligand cross-attention the new Stage 2 conditioning signal.Source: §A.4 · p. 14
- Ligand results use common scoreable targets across compared models, covering 126 of 127 unseen-ligand examples and 1,288 of 1,500 seen-ligand examples.The common-set size is limited by the baseline with the fewest completed generations.Source: §A.4 · p. 14
B.1. Parameter Accounting and Training Configuration
Stage 1 trains the text-conditioned protein-generation model, while Stage 2 adds biological-context pathways and updates most model parameters with pretrained encoders frozen.
B.1. Parameter Accounting and Training Configuration
- Stage 2 adds gated cross-attention adapters and modality projections, bringing the model to 3.20B parameters and making 3.04B trainable.The Stage 1 model has 2,990,576,029 parameters, of which 2,873,132,124 are trainable.Source: §B.1 · p. 14, Table 6 · p. 14
- The text encoder and pretrained ligand and antigen encoders remain frozen during Stage 2, keeping their conditioning representations fixed.The paper states this preserves pretrained domain knowledge and a uniform interface across modalities.Source: §B.1 · p. 14
- Eight adapters are placed every fourth transformer block and initialized with zero-scaled modality output, preserving the Stage 1 computation at the start of Stage 2.Each adapter applies layer normalization followed by multi-head cross-attention into the multimodal context.Source: §B.1 · p. 14
- ChemBERTa encodes ligand SMILES and ESM-2 encodes antigen sequences, with both outputs projected to the 2560-wide backbone.Learned modality-specific scales and type embeddings are applied to the projected outputs.Source: §B.1 · p. 14
B.2. Optimization and Compute
Pretraining and finetuning use different learning rates, batch limits, and clipping settings; pretraining also excludes sequences longer than 1,024 positions.
B.2. Optimization and Compute
- Stage 1 pretraining uses Adam with a peak learning rate of 5 × 10^-4, 10,000 warmup updates, and gradient clipping at 1.0.Training uses 120 NVIDIA B200 GPUs, bfloat16, and an effective batch of up to 1.92M tokens per update.Source: §B.2 · p. 15
- Stage 1 excludes sequences exceeding 1,024 positions and weights the protein objective at 1.0 and text reconstruction at 0.5.The excluded sequences are omitted from training, validation, and test rather than truncated.Source: §B.2 · p. 15
- Stage 2 finetuning lowers the peak learning rate to 5 × 10^-5, uses 2,000 warmup updates and clipping at 0.5, and trains adapters and modality projections at twice the base rate.Batches are capped at 1,100 tokens and 20 sequences per GPU, without gradient accumulation.Source: §B.2 · p. 15
C. Evaluation Details
This section defines evaluation metrics, specifies baseline and case-study settings, and reproduces the full functional prompts used for selected designs.
C.1.1. STAGE 1: MONOMER DESIGN
- Stage 1 monomers are evaluated on predicted ESMFold structures, using pLDDT confidence, global and core RMSD, RMSD-threshold success rates, and InterProScan functional-recovery F1.Validation and test clusters use a reservoir-sampled representative sequence for the prompt, while each design is compared against up to five available cluster structures and the minimum RMSD is reported.Source: §C.1 · p. 15, §C.1.1 · p. 15
C.1.2. STAGE 2: COMPLEX DESIGN
- Stage 2 complex metrics use AlphaFold3 predictions and include ipTM, designed-chain pLDDT, pocket RMSD, docking success, GNINA affinity, and antibody DockQ.Pocket RMSD compares ligand poses after superposition on residues within 8 Å of each structure's ligand; docking success is the fraction below 2 Å.Source: §C.1 · p. 15, §C.1.2 · pp. 15–16
C.2. Baseline Configurations
- The baselines include Pinal, ESM3, ProGen2, RefineGNN, and Proteina-Complexa, with distinct input and generation settings for each.ESM3 receives structured function annotations, target length, and residue positions; ProGen2 receives a target-sequence prefix rather than a functional specification, while RefineGNN redesigns CDRs in separate passes and Proteina-Complexa is run unguided.Source: §C.2 · p. 16
- ESM3 functional-recovery F1 is reported only for examples receiving every requested annotation, a subset the paper describes as most favorable to ESM3.The subset contains 278 examples for the 1.4B model and 406 for the 7B model; other table cells use all generations.Source: §C.2 · p. 16
C.3. Additional Case Studies
- Figure 5 presents two antibody and two ligand-binding designs, showing prompts, designed structures overlaid on ground truth, and close-ups of binding regions.Ligand examples are marked as drawn from seen or unseen evaluation splits.Source: §C.3 · p. 17, Figure 5 · p. 18
- In antibody case studies, InstructPro2-Ab scores DockQ 0.937 versus RefineGNN's 0.748 on VRC01 and 0.967 versus 0.877 on 19B3.The reported margin over the structure-conditioned baseline varies across targets and is largest when the instruction supplies epitope information not recoverable from antigen structure alone.Source: §C.3 · p. 17
- On the unseen BLVRB HUMAN target, InstructPro2-Lb reaches a Pocket RMSD of 0.29 Å, compared with 6.80 Å for Pinal and 20.13 Å for Proteina-complexa.On the seen GHMP kinase target, Pocket RMSD is 0.39 Å for InstructPro2-Lb, 1.79 Å for Pinal, and 20.72 Å for Proteina-complexa; the BLVRB design's pLDDT is 84.08 versus 96.90 for the experimental structure.Source: §C.3 · p. 17
D. Full Functional Descriptions
- The reproduced prompts include an antibody targeting the ACE2-interacting RBD surface and protein descriptions covering GHMP kinase, nuclear hormone receptor, and polyketide oxidoreductase functions.The prompts are shown in full after the abbreviated figure versions.Source: D. Full Functional Descriptions · pp. 17–19