acceptodds
Under review as a conference paper at ICLR 2027

NLPro: A Protein Foundation Model to Follow Human Instructions and Biological Context

Abstract

Designing proteins with specified functions remains a fundamental challenge in computational biology. Motivated by the remarkable instruction-following capabilities of large language models (LLMs), natural language-guided protein generation offers a promising avenue for controllable, on-demand protein design. To this end, we present NLPro, a foundation model that integrates human instructions and biological context for flexible functional protein design. NLPro leverages a text encoder and specialized biological encoders to effectively capture functional descriptions alongside biological contextual constraints. To train a foundation model, we curate NLProBench from UniProtKB, containing 205 million <description, protein> pairs, and subsequently fine-tune the model on two important real-world tasks: ligand-binding protein design and antibody design. NLPro achieves state-of-the-art performance across all benchmarks, attaining a functional F1 score of 74.8%, a DockQ score of 0.39 in antibody design, and an ipTM score of 0.90 in ligand-binding protein design. These results demonstrate NLPro's capability to generate functional proteins for arbitrary goals, highlighting its strong potential for real-world biomedical applications.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
instruction-following capability, functional protein design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_y12JOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents NLPro, a protein generation framework conditioned on natural-language functional descriptions and biological context such as ligands and antigens. It first constructs NLProBench and pretrain a protein decoder using autoregressive generation and span-infilling objectives. The pretrained model is subsequently adapted to antibody and ligand-binding protein design through gated cross-attention to specialized modality encoders. Computational evaluations report improvements in structural accuracy, InterPro recovery, antibody DockQ, and ligand pocket geometry over several existing baselines.

Critical strengths

1. The paper presents a large-scale and practically relevant protein design framework. NLPro unifies natural-language functional descriptions, ligand information, and antigen context within a single generative architecture, and evaluates the framework across various tasks. This broad scope makes the work relevant to both protein language modeling and conditional biological design.

2. The dataset construction and experimental documentation are relatively comprehensive. The authors construct a large protein-text corpus, apply cluster-level train-test splitting, compare against several categories of baselines, and provide detailed descriptions of the architecture, training configuration, evaluation metrics, and baseline settings. These resources and implementation details could support subsequent research on instruction-guided protein generation.

Critical weaknesses
  1. The benefit of the biological-context pathway appears limited and task-dependent. Figure 3 shows that ligand conditioning improves pocket RMSD and docking success, but has little effect on pLDDT, ipTM, or predicted binding affinity. More notably, antigen conditioning does not improve antibody design, with the text-only variant achieving a slightly higher DockQ than the text-plus-modality variant. Since integrating biological context is a central component of NLPro, the inconsistent gains make its contribution less convincing.
  2. The statistical reliability of several reported improvements is insufficiently established. Some performance differences are small, such as 0.39 versus 0.38 DockQ in antibody design and 81.8% versus 80.8% success on unseen ligands, while the paper does not report variation across training runs or confidence intervals. Without uncertainty estimates or significance tests, it is difficult to determine whether these improvements are stable.
Questions for authors
  1. Could the authors further analyze why antigen conditioning does not improve antibody design? In addition, could they compare correct and mismatched ligand or antigen inputs to determine whether NLPro is sensitive to the specific biological context provided?
  2. Could the authors repeat the main downstream experiments across multiple training seeds and report confidence intervals, together with paired bootstrap significance tests over evaluation targets, particularly for the antibody and unseen-ligand results?
  3. Could the authors report the sequence and structural similarity between generated proteins and their nearest neighbors in the training set, together with diversity across repeated generations for the same instruction? This would help clarify whether the model generates novel candidates rather than close variants of known proteins.
2 comments

Q1. Why antigen conditioning does not improve antibody design, and whether NLPro is sensitive to the specific context provided

Figure 3's text-only DockQ was computed on an incomplete target subset. Rescored on a common 433 targets the two variants are indistinguishable, 0.383 against 0.381, and the heavy-antigen interface is unchanged (−0.003, 95% CI [−0.017, +0.012], p=0.75).

Antigen conditioning changes whether AlphaFold3 places an antibody-antigen interface, not the quality of the interfaces it places. DockQ is defined only where an interface exists, so we use ipSAE (Dunbrack, 2025), which scores a chain pair from the predicted aligned error on every prediction.

Does AlphaFold3 place an antigen interface?

conditioningninterface placedvs text only95% CIp
text only44128.8%reference
text + correct antigen44135.1%+6.4 points[+2.7, +10.0]0.0009
text + mismatched antigen44132.4%+3.6 points[+0.2, +7.3]0.056

A mismatched antigen recovers about half the gain, so part of what the pathway supplies is a generic prior that a partner is present. The correct antigen beats the mismatched one on interface confidence (+0.012, p=0.028), so the rest is partner-specific. On the same 433 targets ipTM rises from 0.517 to 0.530 (p=0.008) and pLDDT from 85.68 to 86.11 (p=0.002).

Q2. Confidence intervals and paired bootstrap significance tests

Paired bootstrap over targets, 10,000 resamples, both models moving together, with exact McNemar tests for binary outcomes.

The averaged antibody DockQ does not separate from zero, which is the 0.39 against 0.38 cited. Decomposing by interface shows why: we lead on the antigen interface and trail on heavy-light, which the supplied framework fixes rather than the designed loops.

Antibody, NLPro minus RefineGNN

metricndifference95% CIp
heavy chain to antigen DockQ433+0.0266[+0.0066, +0.0469]0.012
all interfaces averaged DockQ433+0.0112[−0.0044, +0.0269]0.16
heavy to light DockQ433−0.0192[−0.0286, −0.0095]<0.0001
ipTM436+0.0239[+0.0121, +0.0362]<0.0001
mean pLDDT436+0.80[+0.47, +1.13]<0.0001

Docking success, pocket RMSD below 2 Å. NLPro minus baseline, percentage points

splitbaselinendifference95% CIp
seenProGen21293+6.2[+3.1, +9.2]0.0001
seenPinal1293+18.8[+15.9, +21.8]<0.0001
seenInstructPro-3B1293+4.8[+1.9, +7.7]0.0014
unseenProGen2120+0.8[−7.5, +9.2]1.00
unseenPinal121+3.3[−5.0, +11.6]0.54
unseenInstructPro-3B121+11.6[+4.1, +19.0]0.0066

The 81.8% against 80.8% cited is the ProGen2 unseen comparison. Retraining from several seeds does not fit the discussion period, so the uncertainty is over targets rather than training runs.

Q3. Similarity to training-set nearest neighbours, and diversity across repeated generations

All designs use nucleus sampling, p=0.95 and τ=0.8.

Ligand designs against the 205.7M-sequence pretraining corpus

mean identityat 90% or above
Ground truth63.9%3.7%
NLPro61.5%1.2%
Pinal63.9%5.5%
ProGen267.7%4.3%

Against the stage-2 fine-tuning corpus, 39,488 sequences, no design reproduces a sequence exactly, and our mean identity of 54.9% against the ground truth's 52.4% is matched by Pinal at 54.2%, which never saw this corpus.

Ligand designs against their closest match in AFDB/UniProt50, 54M structures

TM-score, seenidentity of match, seenidentity of match, unseen
Ground truth0.9250.6460.724
NLPro0.8930.4860.568
Pinal0.8740.4810.661
ProGen20.8160.4920.661
InstructPro-3B0.9280.7090.833

Antibody. The framework is supplied and only the CDR loops are generated, so we measured the loops. CDR-H3 recovery against the target loop is 36.8% for NLPro and 35.8% for RefineGNN, with 2 of 438 designs reproducing the target exactly. Against the fine-tuning corpus, 13.6% of our loops of length 12 or more occur verbatim in training, against 0 of 349 for both the ground truth loops and RefineGNN. The ground truth rate of zero shows the split held, so this is memorisation rather than leakage, and we report it as a limitation. It appears specific to loop infilling, where the output space is short and the corpus redundant. On the ligand task no design in any arm reproduces a fine-tuning sequence verbatim.

Diversity. Three decoding seeds per instruction. No ligand design repeats between seeds, 0 of 126 unseen. On antibodies 87.4% of targets yield at least two distinct CDR-H3 loops and 63.5% yield three.

Dear Reviewer,

We have added experiments and addressed all your concerns. As the rebuttal deadline is approaching, would you please take a look at our response? If you have any additional questions, we are more than happy to discuss.

Best,

Authors of paper 21634

Reviewer_Df9aOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces NLPro, a natural-language-guided protein design model that combines text encoder with an autoregressive protein decoder and other biological context encoders. The authors curated NLProBench, a large dataset of 205.7M protein–text pairs constructed with InterPro annotations. It is further finetuned for antibody CDR design and ligand-binding protein design by incorporating antigen or ligand representations through cross-attention.

The experiments evaluate sequence designs using structural quality, recovery of InterPro annotations, and metrics against specific tasks such as DockQ and RMSD. NLPro reports strong performance compared with several existing protein language models and text-guided protein design methods.

Critical strengths
  1. The authors curated a large scale dataset NLProBench with more than 205M protein–text pairs. They summarizes the InterPro vocabulary once and composing descriptions for individual proteins is a scalable way to obtain a large scale of protein-text training corpus.
  2. The framework is unified and supports multiple modes of generation including full-sequence generation and span infilling, making it able to perform different design tasks all in one model.
Critical weaknesses
  1. The benchmark construction and evaluation are tightly coupled. First, the authors use InterPro annotations to construct the training prompts. While during the evaluation, the main metric of evaluating text-protein alignment is checking whether the InterPro entries are covered, the reported F1 may largely measure how well the model maps InterPro annotations back to InterPro-recognizable sequences. I will suggest the authors evaluated the ProTrek[1] score. Also, it makes me wonder what if we perform retrieval and combine. For each of the InterPro annotation, we select a fragment from the training corpus that matches the annotation, and combines those fragments. That is similar to ProDVa[2].
  2. The current train/test split does not demonstrate generalization to unseen functions or instructions. The authors split the dataset based on sequence identity, which is commonly used in protein work. However, the authors claimed that the model can "following human instructions for arbitrary goals." In order to cliam that, they should also split the dataset based on InterPro annotation similarities and test on the hold-out InterPro annotation.

[1] Su, J., He, Y., You, S. et al. A trimodal protein language model enables advanced protein searches. Nat Biotechnol 44, 1310–1316 (2026). https://doi.org/10.1038/s41587-025-02836-0 [2] Nuowei Liu, Jiahao Kuang, Yanting Liu, Tao Ji, Changzhi Sun, Man Lan, and Yuanbin Wu. 2026. Protein design with dynamic protein vocabulary. In Proceedings of the 39th International Conference on Neural Information Processing Systems (NIPS '25), Vol. 38. Curran Associates Inc., Red Hook, NY, USA, Article 599, 19939–19970. https://doi.org/10.52202/085713-0672

Questions for authors

I wonder if we can use ProGen only in the antibody design task rather than introducing an additional protein langauge model?

2 comments

Q1. ProTrek evaluation

We ran ProTrek with the official ProTrek_650M checkpoint on the 490 stage-1 test targets every model shares. Under nucleus sampling NLPro is indistinguishable from Pinal, well ahead of ProGen2 and ESM3 at both sizes, and about one point below the ground truth. The shuffled control confirms the metric reads text alignment rather than amino-acid composition.

ProTrek, 490 shared targetsmean scoreNLPro sampling minus this model95% CIp
Ground truth11.32−1.07[−1.42, −0.75]<0.0001
Pinal10.20−0.17[−0.64, +0.30]0.49
NLPro, nucleus sampling10.14reference
NLPro, greedy8.46+1.68[+1.21, +2.16]<0.0001
ESM3 7B5.94+3.90[+3.35, +4.46]<0.0001
ProGen25.92+4.26[+3.71, +4.82]<0.0001
Shuffled ground truth2.28+7.86[+7.35, +8.37]<0.0001

Greedy's lower mean comes from output length rather than weaker alignment: it produces sequences under 50 residues on 125 of 490 targets, against 7 for every other model including the ground truth. On its longer outputs greedy averages 10.37 and is also indistinguishable from Pinal (−0.21, 95% CI [−0.72, +0.30], p=0.43).

We report ProTrek alongside InterProScan F1 rather than in place of it. The two metrics agree that NLPro is far ahead of ProGen2 and ESM3, and differ on the size of the margin over Pinal.

Q2. A split based on InterPro annotation similarity

The test set contains no unseen InterPro entries. We streamed all 205,727,986 training records and checked every entry the stage-1 test set asks for: all 956 appear in training. Testing on held-out entries would mean re-splitting the corpus by annotation and pretraining again.

The split does contain unseen combinations of entries. 11 of the 1,000 test examples ask for a combination that never appears in training, and the rest occur from a few times to more than 10,000. Recall rises with frequency for every model, which is what the reviewer's concern predicts.

InterProScan recall by how often the requested combination appears in training

occurrences in trainingnNLPro samplingNLPro greedyPinalESM3 7BProGen2
0, never seen110.3630.3140.1630.1130.100
1 to 970.5020.3500.2710.2830.000
10 to 99250.5640.4120.4070.3390.128
100 to 999770.5930.5550.3710.2150.253
1,000 to 9,9992790.7620.6250.3690.2580.216
10,000 or more6010.9170.8160.7610.5830.603

The decline is not specific to our model. Comparing the 120 rare examples, those under 1,000 occurrences, against the 880 common ones, recall drops by 0.27 for NLPro greedy, 0.31 for NLPro sampling, 0.28 for Pinal, 0.24 for ESM3 7B and 0.28 for ProGen2, all p<0.0001. ESM3 and ProGen2 were trained on different corpora, so their recall should not depend on how often a combination appears in ours. Their declines are the same size as ours, so the rare combinations are harder for any model rather than less familiar to ours. NLPro leads every bucket, and on the eleven never-seen combinations it more than doubles the best baseline, 0.363 against 0.163. We acknowledge that 11 examples is not enough to draw conclusions on its own, but it is consistent with the margin we hold across the larger buckets.

The reviewer is also right that "following human instructions for arbitrary goals" is stronger than this evidence supports, so we have narrowed the claim. The abstract and Section 4.2 now describe following natural-language instructions across diverse protein families.

Q3. Could ProGen2 alone serve the antibody task, without a separate protein language model?

We used a separate encoder mainly to avoid extending the decoder's context. Antigens run to several hundred residues, so putting one in the decoder's input would consume context we need for generation. Cross-attention lets the decoder attend to the antigen without carrying it in its own token stream. ESM-2 also gives a bidirectional representation of an input that is given rather than generated, and being frozen it adds no trainable parameters, 34M against 3.2B.

Dear Reviewer,

We have added experiments and addressed all your concerns. As the rebuttal deadline is approaching, would you please take a look at our response? If you have any additional questions, we are more than happy to discuss.

Best,

Authors of paper 21634

Reviewer_niYrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

NLPro generates protein sequences from natural-language descriptions, with optional ligand or antigen conditioning. It is pretrained on 205.7 million InterPro-derived description–sequence pairs and fine-tuned for antibody and ligand-binding design. Evaluation uses functional annotation recovery, predicted structures, and docking metrics.

Critical strengths
  1. Summarizing InterPro entries once and composing descriptions across proteins is a scalable approach to dataset construction. The resulting corpus could be a useful resource for text-conditioned protein modeling.

  2. The reported computational gains are promising. Functional F1 improves from Pinal’s 62.2 to 74.8, and unseen-ligand pocket RMSD improves from the strongest baseline’s 2.53 Å to 1.79 Å.

Critical weaknesses
  1. The closest baseline is missing. InstructPro already generates proteins conditioned on text and ligands, and supplies the ligand benchmark adopted here. Without a direct comparison and a clearer account of the methodological differences, the advance over this work is difficult to assess.

  2. The antibody results do not clearly establish the benefit of antigen conditioning. Text-only achieves DockQ 0.41 versus 0.39 with antigen input (§5.2). Moreover, the reported DockQ averages all chain-pair interfaces, including heavy–light interfaces (Appendix C.1.2), so it does not isolate antibody–antigen docking quality.

Questions for authors
  1. Why is InstructPro omitted? How does NLPro compare when evaluated with the same descriptions, targets, and sampling budget?

  2. Can you report antigen-interface-only DockQ, with uncertainty estimates, for the text-only and text-plus-antigen variants?

  3. Were downstream test proteins and close homologs excluded from the 205.7-million-pair pretraining corpus? Please quantify any overlap and the generated sequences’ similarity to training proteins.

3 comments

Q1. Why is InstructPro omitted, and how does NLPro compare under the same descriptions, targets, and sampling budget?

InstructPro-3B is now a row in both ligand tables of the updated draft and in the baselines description, as the closest prior system in objective and parameter count. We report it on the same ligands, 1,288 seen and 126 unseen inherited from InstructProBench, and the same budget of one generation per target with no best-of-N. Descriptions differ by design: each model receives the text format it was trained on.

ligand taskpocket RMSD, seensuccess, seenpocket RMSD, unseensuccess, unseen
InstructPro-3B3.40 Å67.9%3.05 Å70.0%
NLPro2.24 Å72.6%1.81 Å81.7%

InstructPro-3B leads us on pLDDT and on GNINA-predicted affinity, which the updated draft reports.

Q2. Antigen-interface-only DockQ, with uncertainty, for the text-only and text-plus-antigen variants

The reviewer is right that the published DockQ averages every chain-pair interface, including heavy to light, which the supplied framework and light chain largely fix. We recomputed it by interface role on one set of 433 antibodies scored for all arms on all interfaces, the set Table 2 now reports.

Text plus antigen minus text only

interfacedifference95% CIp
heavy chain to antigen−0.0025[−0.0172, +0.0124]0.75
heavy to light−0.0004[−0.0088, +0.0082]0.92
all interfaces averaged−0.0027[−0.0150, +0.0098]0.67

NLPro minus RefineGNN, same 433 targets

interfacedifference95% CIp
heavy chain to antigen+0.0266[+0.0066, +0.0469]0.012
all interfaces averaged+0.0112[−0.0044, +0.0269]0.16
heavy to light−0.0192[−0.0286, −0.0095]<0.0001

Isolating the antigen interface shows no difference between the conditioning variants, which settles the §5.2 comparison, and it sharpens the comparison with the task-specific baseline: we lead on the antigen interface and trail on heavy to light, and averaging hides both.

On the heavy-antigen interface the means are 0.162 against RefineGNN's 0.135, a 20% relative gain. At the CAPRI high-quality threshold of 0.80 we reach 3.7% of targets (16 of 433) against 1.2% (5 of 433), +2.5 points with 95% CI [+0.7, +4.4] and McNemar p=0.013. At the medium threshold of 0.49 the rates are 12.5% and 9.5%, p=0.085, not separable from zero. Full decomposition in Appendix D.1.

DockQ cannot resolve whether antigen conditioning helps, since it is defined only where AlphaFold3 places an interface. ipSAE is defined on every prediction, and on it conditioning raises interface formation from 28.8% to 35.1% (+6.4 points, 95% CI [+2.7, +10.0], McNemar p=0.0009). A mismatched antigen recovers about half of that gain, so part of what the conditioning supplies is a generic prior that a partner is present and the rest is partner-specific.

Q3. Overlap between downstream test proteins and the pretraining corpus, and the similarity of generated sequences to training proteins

The pretraining corpus was not filtered against the downstream test sets, so we quantified the overlap directly, by BLAST against the full 205.7M-sequence corpus, taking each query's best hit at a minimum of 50% query coverage.

nmean identityat 90% or above
seen-ligand test proteins129563.9%3.7%
NLPro designs, seen129561.5%1.2%
unseen-ligand test proteins12675.6%19.0%
NLPro designs, unseen12667.9%0.0%

Overlap between the test proteins and the corpus exists, and the model does not reproduce it: the test proteins sit closer to the corpus than our designs do on both splits.

We also checked the stage-2 fine-tuning corpus, which is small enough for memorisation to be a concern. No ligand design in any arm reproduces a fine-tuning sequence exactly, and our mean identity of 54.9% against the ground truth's 52.4% is matched by Pinal at 54.2%, which never saw this corpus.

On the antibody task, where only the CDR loops are generated, 13.6% of our CDR-H3 loops of length 12 or more occur verbatim in the fine-tuning corpus, against 0 of 349 for both the ground truth loops and RefineGNN. We report that as a limitation and discuss it in our reply to Reviewer y12J.

Structurally, our ligand designs reach TM-score 0.893 against their closest match in AFDB/UniProt50 while sitting further in sequence from that match than the ground truth does, 0.486 against 0.646 on the seen split. Across three decoding seeds no design repeats, 0 of 126 on the unseen set.

Dear Reviewer,

We have added experiments and addressed all your concerns. As the rebuttal deadline is approaching, would you please take a look at our response? If you have any additional questions, we are more than happy to discuss.

Best,

Authors of paper 21634

Dear Authors,

Thank you very much for your detailed explanations, however, I think it does not isolate antibody–antigen docking quality during the experiments. Hence I prefer to have 2 on rating.Hence I remain the orginial rating.

Best Regards, Reiviewer

Comments on the paper

1 comment

We thank the reviewers for their questions and suggestions. All new analyses use matched target sets, with paired bootstrap confidence intervals and exact McNemar tests.

The following have been added to the revised PDF, with changes highlighted in blue: InstructPro-3B in both ligand tables, and for the antibody task, DockQ decomposed by interface role, ipSAE, and a mismatched-antigen control. The remainder are reported in our replies below: ProTrek across eight models, sequence and structural novelty against the 205.7M-sequence pretraining corpus, the stage-2 fine-tuning corpus and AFDB/UniProt50, performance by InterPro combination rarity, and diversity across decoding seeds.

Corrections. The text-only DockQ in Figure 3(b) came from an incomplete target subset and is fixed. On the corrected figure, antigen conditioning improves ipTM and pLDDT significantly and leaves DockQ unchanged, which is what a metric defined only where an interface has already been placed would show. Table 2 now uses one set of 433 targets with every metric present for every model, which lowers our antibody DockQ from 0.39 to 0.38 and RefineGNN's from 0.38 to 0.37. We agree with the reviewers that the averaged DockQ margin over RefineGNN is not significant, and now report DockQ by interface role, where our lead on the antibody-antigen interface is significant (0.162 against 0.135, p=0.012).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.