- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper introduces FBindGen, a function-conditioned protein binder sequence generation model that aims to jointly control binding and functional properties. Given a target protein’s sequence and structure, binder length, and residue-level functional annotations, FBindGen uses discrete diffusion to generate binder sequences. Experiments show that FBindGen can generate binders with stronger alignment to the desired functional annotations while maintaining competitive binding/interface quality, suggesting that incorporating functional information directly into the generation process can enable more controllable protein binder design.
Critical strengths
- The authors construct the EnzBind benchmark to evaluate function-conditioned binder design, providing a useful testbed for this task.
- The paired generation experiments show that changing the functional conditioning can lead to systematic changes in the generated binder sequences, supporting the controllability of the model.
- FBindGen achieves competitive or superior performance across several binding- and function-related metrics compared with existing binder generation methods.
Critical weaknesses
- The roles of S2P, P2S, and PairFormer need to be ablated. In particular, as a sequence generation model, it needs to explain why pair + structural distance loss are introduced. What is the difference between doing this and structure generation?
- The evaluation of Sequence–text alignment is not fair to the baselines. None of them are conditioned on Binder InterPro IDs for function. Being more consistent with the annotation does not necessarily mean having the corresponding function. The authors need to conduct more specific sequence/structure-level analyses on the generated samples to support the role of per-residue InterPro annotations in function.
- As a sequence design model, although it cannot report self-consistent metrics such as RMSD based on folding models, it could still report the structural distribution of the generated sequences predicted by folding models. Are they all helical bundles? Do the interfaces form hydrogen bonds? How is the shape complementarity?
- What is the distribution of the designed sequences? How diverse are they? What is the distribution of the 20 amino acid types?
Questions for authors
How was the “predicted AlphaFold heterodimers” dataset constructed?
4 comments
7. How is the shape complementarity?
Shape complementarity describes how closely the target and binder surfaces fit together; higher scores indicate better geometric fit. We used PyRosetta 2025.33 with Rosetta's ShapeComplementarityFilter to calculate the approximate score sc_quick.
| Source | Mean sc_quick |
|---|---|
| FBindGen | 0.5056 |
| BoltzGen | 0.4987 |
| Proteina-Complexa | 0.4805 |
| Native | 0.6115 |
FBindGen scored slightly above both baselines and below the native reference.
8. What is the distribution of the designed sequences? What is the distribution of the 20 amino acid types?
We counted amino acids across all 27,500 designs per method and compared their frequencies with the 275 corresponding native binder sequences.
Amino acid composition, as a percentage of all counted residues.
| Amino acid | Native (%) | FBindGen (%) | BoltzGen (%) | Proteina-Complexa (%) |
|---|---|---|---|---|
| Alanine (A) | 7.75 | 10.83 | 15.00 | 11.77 |
| Cysteine (C) | 2.03 | 0.65 | 0.71 | 0.53 |
| Aspartate (D) | 5.03 | 6.08 | 4.27 | 4.44 |
| Glutamate (E) | 5.93 | 8.81 | 9.27 | 8.03 |
| Phenylalanine (F) | 4.29 | 3.26 | 2.27 | 3.28 |
| Glycine (G) | 6.92 | 6.00 | 8.59 | 6.50 |
| Histidine (H) | 2.54 | 1.11 | 0.70 | 1.40 |
| Isoleucine (I) | 6.06 | 5.01 | 5.29 | 7.02 |
| Lysine (K) | 5.49 | 5.60 | 5.98 | 7.70 |
| Leucine (L) | 9.47 | 12.24 | 13.28 | 14.18 |
| Methionine (M) | 2.89 | 1.39 | 0.75 | 0.79 |
| Asparagine (N) | 4.02 | 3.24 | 2.15 | 4.23 |
| Proline (P) | 4.73 | 6.67 | 4.13 | 3.74 |
| Glutamine (Q) | 3.66 | 2.64 | 1.67 | 2.96 |
| Arginine (R) | 5.39 | 6.08 | 4.37 | 2.68 |
| Serine (S) | 6.15 | 5.07 | 5.20 | 5.37 |
| Threonine (T) | 5.23 | 5.03 | 4.97 | 4.57 |
| Valine (V) | 7.07 | 6.53 | 8.48 | 7.43 |
| Tryptophan (W) | 1.55 | 1.04 | 0.53 | 0.38 |
| Tyrosine (Y) | 3.82 | 2.72 | 2.37 | 3.00 |
FBindGen's Jensen–Shannon distance to the native composition was 0.1305, compared with 0.1931 for BoltzGen and 0.1629 for Proteina-Complexa. Smaller distances indicate more similar amino acid compositions.
9. How diverse are they?
Sequence diversity across 275 EnzBind cases, using 100 designs per case.
| Method | Unique sequences / 100 | Pairwise diversity (mean ± SD) |
|---|---|---|
| FBindGen | 99.89 | 0.7890 ± 0.1689 |
| BoltzGen | 99.95 | 0.7493 ± 0.2009 |
| Proteina-Complexa | 95.24 | 0.8779 ± 0.1403 |
10. How was the “predicted AlphaFold heterodimers” dataset constructed?
We used published predictions from the AlphaFold Database complex-structure release by Han, Tsenkov, Venanzi et al.. The release used AlphaFold-Multimer to model candidate protein pairs from STRING physical interactions.
Thank you again for your constructive feedback. The individual S2P, P2S, and Pairformer ablations are still running, and we will update this response as soon as the results are available. We would be happy to answer any further questions.
Thank you for your response and for providing the additional results. Based on my background, I see generating an enzyme that actually performs a specified function from a label alone as a very difficult problem. Many established enzyme design workflows start from a catalytic motif taken from a naturally occurring structure to help preserve the desired activity. The reason is that many catalytic mechanisms are still not fully understood. For an entirely new structure, we cannot tell whether it has the intended catalytic activity without experimental validation, which may be beyond the scope of a computer science conference paper.
From this perspective, I do not see how one can verify in silico that an enzyme generated from a functional label actually performs the specified function. I am not convinced that either the proposed “alignment” metric or the newly reported catalytic residue accuracy establishes this. If the goal is simply to recover the catalytic residues, why not extract them and use them as a catalytic motif? In that setting, conditioning sequence generation on an unindexed catalytic sequence motif makes more sense to me than conditioning it on a functional label alone.
My main concern is how the problem is defined. If the authors wish to claim that the generated enzymes perform the intended function, I would expect experimental validation, and that work might be better suited to a journal. For a computer science conference, I would favor a problem formulation whose central claim can be tested in silico using criteria that researchers in the field would find convincing.
Thank you for your feedback. We agree that designing active enzymes is challenging and that computational metrics cannot prove activity. Our goal is to study whether functional annotations can guide binder sequence generation. We would like to clarify three points.
- Functional inputs provide another way to design enzymes. ZymCTRL [1] and Pinal [2] have experimentally shown that functional inputs can guide the design of active enzymes. This supports our goal of using functional annotations to design binders.
- Functional inputs can guide enzyme design when catalytic mechanisms are unclear. Catalytic motifs are useful, but functional inputs offer another way to design enzymes. Some aspects of LPMO catalysis remain unclear [3]. We tested an AA9 LPMO using its InterPro annotations. Across 18 designs, the highest ProTrek score was 0.202, compared with 0.058 without functional annotations. These results suggest that functional inputs can improve sequence–function alignment scores even when catalytic mechanisms are unclear.
- Functional alignment metrics can help select active enzyme candidates. Pinal [2] and ProDVa [4] use ProTrek to evaluate sequence–function alignment. Pinal used pLDDT, pAE, and ProTrek to select ADH candidates and later confirmed their activity experimentally. Similarly, ipTM has experimental support for evaluating protein interactions. A recent LUCIA study [5] found that 77% of predicted protein pairs with ipTM ≥ 0.80 showed detectable interactions in wet-lab experiments, supporting the use of ipTM for candidate selection. This shows that computational metrics can help select candidates for experimental testing.
We will clarify that our computational evaluation focuses on functional annotation agreement, responses to changes in functional conditions, and predicted interface quality. Experimental validation will be pursued in future work.
We appreciate your feedback and hope these clarifications address your concerns.
References
[1] Munsamy et al. (2024). Conditional language models enable the efficient design of proficient enzymes. bioRxiv.
[2] Dai et al. (2026). Programmable design of functional proteins from natural language. Nature.
[3] Munzone et al. (2024). Expanding the catalytic landscape of metalloenzymes with lytic polysaccharide monooxygenases. Nature Reviews Chemistry.
[4] Liu et al. (2025). Protein Design with Dynamic Protein Vocabulary. NeurIPS.
[5] Zhang et al. (2026). Systematic Validation of AlphaFold-Predicted Interactomes with LUCIA. bioRxiv. https://doi.org/10.64898/2026.04.02.716051
Thank you for recognizing the value of EnzBind for evaluating function-conditioned binder design. We tested the effect of the annotations and analyzed the generated sequences, folds, and interfaces.
1. In particular, as a sequence generation model, it needs to explain why pair + structural distance loss are introduced. What is the difference between doing this and structure generation?
Pretrained protein language models are already strong at generating individual sequences. Conditioning on a target sequence alone, however, does not explicitly teach the model how a binder should interact with that target. We therefore supervise pair representations using distances within the binder and across the target–binder interface. This spatial information guides sequence generation.
FBindGen builds on the pretrained model's sequence knowledge and generates binder sequences directly. Direct sequence generation is faster than 3D structure generation in our timing comparison below.
All three methods used an NVIDIA H100. Each method used its official default number of decoding steps.
Mean generation time
FBindGen was 16.8× faster than BoltzGen and 32.6× faster than Proteina-Complexa.
2. The evaluation of Sequence–text alignment is not fair to the baselines. None of them are conditioned on Binder InterPro IDs for function.
The lack of functional conditioning is precisely the problem we aim to address. Without a way to specify the desired function as input, it is difficult to design binders with that function.
3. Being more consistent with the annotation does not necessarily mean having the corresponding function.
We checked whether the generated proteins reproduced the catalytic residues of the natural reference enzyme. Using M-CSA records, we first checked whether the same amino acids appeared at the known catalytic positions. We then compared their spatial arrangement in the predicted structures with the reference.
We generated five designs per case under each annotation condition.
The first row reports the percentage of catalytic positions with the correct amino acid. The other rows report the percentage of designs passing each check. Percentages were calculated within each case and then averaged equally across cases.
With full annotations, 31.4% of designs passed both the catalytic residue and geometry checks, compared with 0% without annotations.
4. The authors need to conduct more specific sequence/structure-level analyses on the generated samples to support the role of per-residue InterPro annotations in function.
We compared samples generated with full InterPro annotations and without annotations.
With InterPro annotations, the generated sequences better matched the requested IDs and function descriptions and had higher predicted interface confidence.
5. As a sequence design model, although it cannot report self-consistent metrics such as RMSD based on folding models, it could still report the structural distribution of the generated sequences predicted by folding models. Are they all helical bundles?
We analyzed predicted structures per method with native structures for comparison. The designs contained beta-sheet and coil/loop residues as well as helices.
Secondary-structure composition of binder residues.
6. Do the interfaces form hydrogen bonds?
We used PyRosetta 2025.33 with Rosetta's
HBondSetto count interface hydrogen bonds in the generated structures and native references. At least one interface hydrogen bond was predicted in most FBindGen designs.Interface hydrogen bonds inferred from each structure.