acceptodds
Under review as a conference paper at ICLR 2027

FBindGen: Function-Conditioned Protein Binder Sequence Generation with Target-Aware Pair Modeling

Abstract

Existing protein binder design methods primarily focus on structural compatibility with a specified target, while direct control over binder function remains less explored. We introduce FBindGen, a diffusion protein language model that generates binder sequences from a target protein and functional annotations at the residue level. FBindGen combines functional conditioning with target-aware pair modeling by allowing sequence and pair representations to update each other throughout denoising. To evaluate how well generated binders follow the specified function, we introduce EnzBind, a benchmark of protein complexes involving enzymes with functional annotations and substrate information. On EnzBind and DB5.5, FBindGen shows improved agreement between generated sequences and the specified functions while maintaining competitive predicted interface confidence. On EnzBind, it also achieves the highest predicted catalytic efficiency among the evaluated methods. Evaluation on Heterodimer99 further shows that the learned pair representations capture interchain contact information. Together, these results show that FBindGen provides functional control over binder sequence generation while preserving compatibility with the target.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein binder design, Function-conditioned protein generation, Diffusion language models, Protein–protein interactions

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jyjbOpenReview ↗
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
  1. The authors construct the EnzBind benchmark to evaluate function-conditioned binder design, providing a useful testbed for this task.
  2. 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.
  3. FBindGen achieves competitive or superior performance across several binding- and function-related metrics compared with existing binder generation methods.
Critical weaknesses
  1. 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?
  2. 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.
  3. 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?
  4. 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

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

MethodTime (s/binder) ↓Time vs. FBindGen
BoltzGen8.85016.8×
Proteina-Complexa17.20132.6×
FBindGen0.5281.0×

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.

MetricFull InterProNo annotations
Catalytic residue accuracy61.8%4.3%
All catalytic residues correct35.0%0%
Catalytic geometry pass rate31.4%0%

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.

InputInterPro F1ProTrekipTM
Full InterPro0.49600.31410.2511
No annotations0.02280.19630.1504

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.

SourceHelix (%)Beta sheet (%)Coil/loop (%)
Native40.54%16.70%42.76%
FBindGen48.33%13.67%38.00%
BoltzGen47.54%21.57%30.88%
Proteina-Complexa64.51%7.27%28.21%

6. Do the interfaces form hydrogen bonds?

We used PyRosetta 2025.33 with Rosetta's HBondSet to 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.

Source≥1 H-bond (%)Mean H-bonds
FBindGen79.47%3.06
BoltzGen74.08%2.72
Proteina-Complexa64.62%1.71
Native64.44%3.71

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.

SourceMean sc_quick
FBindGen0.5056
BoltzGen0.4987
Proteina-Complexa0.4805
Native0.6115

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 acidNative (%)FBindGen (%)BoltzGen (%)Proteina-Complexa (%)
Alanine (A)7.7510.8315.0011.77
Cysteine (C)2.030.650.710.53
Aspartate (D)5.036.084.274.44
Glutamate (E)5.938.819.278.03
Phenylalanine (F)4.293.262.273.28
Glycine (G)6.926.008.596.50
Histidine (H)2.541.110.701.40
Isoleucine (I)6.065.015.297.02
Lysine (K)5.495.605.987.70
Leucine (L)9.4712.2413.2814.18
Methionine (M)2.891.390.750.79
Asparagine (N)4.023.242.154.23
Proline (P)4.736.674.133.74
Glutamine (Q)3.662.641.672.96
Arginine (R)5.396.084.372.68
Serine (S)6.155.075.205.37
Threonine (T)5.235.034.974.57
Valine (V)7.076.538.487.43
Tryptophan (W)1.551.040.530.38
Tyrosine (Y)3.822.722.373.00

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.

MethodUnique sequences / 100Pairwise diversity (mean ± SD)
FBindGen99.890.7890 ± 0.1689
BoltzGen99.950.7493 ± 0.2009
Proteina-Complexa95.240.8779 ± 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.

  1. 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.
  2. 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.
  3. 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

Reviewer_QuGxOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In this paper authors propose FBindGen, a diffusion protein language model conditioned on a target sequence, structure, and InterPro annotations. Authors propose EnzBind, a binder design benchmark of enzyme targets (with InterPro annotations), and evaluate their method.

Critical strengths

Propose the task and benchmark of enzyme binder generation

Critical weaknesses

Authors argue FBindGen maintains competitive binding performance, but they are not using standard metrics used to evaluate binding quality in silico. They generate one sequence per EnzBind entry and look at mean ipTM across 275 entries. Mean ipTM is not the metric used in the field; people generate 100+ sequences and look at unique success rate. (where success is conjunction of self-consistency and interface confidence, where the exact cutoff might differ by papers).

InterPro alignment/recovery evaluation in current form is flawed. There is no sequence/structure novelty or diversity analysis. In this evaluation, literally "Native reference" is SOTA. Because FBindGen is directly trained on InterPro annotation, we cannot say this method can generate novel, diverse binders until authors show the sequence/structure novelty or diversity analysis. In other words, FBindGen can copy the native binder (part or whole of fold or sequence) and still look like it is "outperforming" other models.

Questions for authors

What is sequence/structure novelty and diversity of FBindGen binders? Can you generate 100 sequences per EnzBind entry and report unique successes following Proteina-Complexa's cutoff? Can you also report their max sequence identity, similarity, max TM-score to existing InterPro families?

5 comments

Thank you for suggesting that we report pass rates and the diversity and novelty of the generated designs.

1. Can you generate 100 sequences per EnzBind entry and report unique successes following Proteina-Complexa's cutoff?

For each of the 275 EnzBind cases, we generated 100 sequences per method. FBindGen does not generate a backbone, so scRMSD cannot be evaluated. We required ipAE < 7 Å, pLDDT > 90, and ProTrek > 0.3 for all methods. We clustered passing binder structures within each case using Foldseek (TM-score threshold 0.5), then summed the cluster counts across cases.

Table Q1. Joint screening with ipAE, pLDDT, and ProTrek

MethodPassing designsStructural clusters
BoltzGen16/27,500 (0.058%)3
Proteina-Complexa153/27,500 (0.556%)5
FBindGen159/27,500 (0.578%)14

2. What is sequence/structure novelty and diversity of FBindGen binders?

We measured diversity as the average percentage of positions that differed between sequences, using all 100 designs per case. FBindGen showed higher sequence diversity than BoltzGen and lower diversity than Proteina-Complexa.

Table Q2. Diversity among generated sequences and similarity to the natural reference binder

MethodPairwise diversity (%) ↑Unique sequences / 100 ↑Native identity (%) ↓
BoltzGen74.93%99.958.56%
Proteina-Complexa87.79%95.247.29%
FBindGen78.90%99.8917.96%

3. Can you also report their max sequence identity, similarity, max TM-score to existing InterPro families?

For each design, we found the closest sequence in UniRef50 and the closest structure in PDB. The table reports average sequence identity and TM-score across all designs.

Table Q3. Novelty

MethodUniRef50 identity (%) ↓PDB TM-score ↓
BoltzGen37.99%0.6308
Proteina-Complexa39.76%0.6020
FBindGen29.56%0.4884

We understand your concerns about FBindGen's ability to design binders and hope these results address them. We would be happy to answer any further questions.

  1. Thank you for generating 100 sequences per case. However, the joint results do not directly answer my question about the claim that FBindGen "maintains competitive binding performance." Please report binding performance separately from ProTrek. Joint with ProTrek filter makes it difficult to determine whether differences in pass rates reflect binding quality or the functional proxy. Specifically, please report the percentage of designs satisfying ipAE < 7 Å and pLDDT > 90, without ProTrek filtering. Ideally the distribution of per-case pass rates across the 275 cases too.

  2. Authors only really answer about sequence, when I asked structural diversity as well. Please also report structural diversity as the number of Foldseek clusters at TM=0.6 among designs satisfying ipAE < 7 Å and pLDDT > 90, without ProTrek filtering. This is the standard in the field. Clustering should be performed within each case.

  3. For the novelty analysis, I was asking specifically about similarity to members of the InterPro family associated with each EnzBind case, rather than entire UniRef50 or PDB.

Thank you for your follow-up. We have completed the additional analyses.

1. Binding performance

We evaluated 100 designs per case using ipAE < 7 Å and complex pLDDT > 90.

MetricFBindGenBoltzGenProteina-Complexa
Passing designs / 27,500366159269
Pass rate (%)1.3310.5780.978
Successful cases / 275182620

2. Structural diversity

We clustered the passing binder structures within each case using Foldseek with a TM-score threshold of 0.6.

MetricFBindGenBoltzGenProteina-Complexa
Successful cases / 275182620
Structural clusters349348

FBindGen shows lower structural diversity, possibly because designs with the same functional domains share similar folds. However, it offers a distinct capability by allowing users to specify functional domains during binder design.

3. Sequence identity to InterPro families

We downloaded the protein2ipr.dat.gz file from InterPro to identify the families and domains associated with each EnzBind case. We then collected their member sequences and searched each design against its case reference set using MMseqs2.

For each design, we recorded the maximum sequence identity and averaged these values over 100 designs per case. The table shows the distribution of these averages results.

Mean maximum identityFBindGenBoltzGenProteina-Complexa
0.0–0.194836
0.1–0.2111153181
0.2–0.3863639
0.3–0.438197
0.4–0.518100
0.5–0.6811
0.6–0.7021
0.7–0.8000
0.8–0.9112
0.9–1.0014

We hope these results address your concerns. Thank you for your time and consideration.

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

FBindGen generates protein binder sequences while trying to satisfy both a desired function and compatibility with a target protein. Its strongest result is improved recovery of requested functional annotations. Evidence that the designs actually bind and perform those functions remains computational.

Critical strengths

The paper additionally use the functional annotations to guide binder design except the binding target, which is useful and effective.

Critical weaknesses

Evaluation:

Benchmarks provide native binder lengths and residue-level annotation layouts. In practical design, users may know the desired function without knowing its residue boundaries. Generalization to freely specified functions, new domain combinations, or flexible lengths remains uncertain.

EnzBind’s mean ipTM advantage is only 0.07. On DB5.5, 21.3% of designs reach the reported 0.8 threshold. Separate aggregate function and interface metrics also leave unclear how often the same design satisfies both objectives. The annotation-switch experiment evaluates text alignment without reporting whether binding compatibility survives the switch.

Experimental setting:

Main generation evaluations use only one sequence per method per case. Training exclusions use dates and benchmark PDB identifiers, but the described procedure does not establish that all homologous training examples are excluded. This leaves generalization to genuinely unfamiliar protein families unresolved. Additionally, the model also needs to show its diversity and novelty.

Questions for authors

See above limitations

2 comments

Thank you for recognizing the value of functional conditioning in binder design. We developed a workflow that recommends binder length and residue boundaries for the requested functions. We also added evaluations of ipTM, diversity, novelty, and generalization.

1. In practical design, users may know the desired function without knowing its residue boundaries. Generalization to freely specified functions, new domain combinations, or flexible lengths remains uncertain.

We developed a workflow so users can specify desired functions without defining residue boundaries. Given a target and the requested functions, it uses templates to recommend binder length and assign functional annotations to residue positions.

Experiment 1. Testing whether an added request produces a second ubiquitin domain

We tested whether requesting a second ubiquitin (Ub) domain would produce it in the generated binder. Ub is a small protein used by cells as a regulatory tag. We chose HOIL-1L NZF because this target naturally binds two linked Ub domains (PDB 3B08).

In the first condition, we requested Ub only at the N-terminus (sequence beginning). In the second, we also requested Ub at the C-terminus (sequence end). Both conditions used the same target, binder length (152 residues), generation settings, and four random seeds.

MetricN-terminal Ub onlyUb at both ends
Generated designs44
Ub detected at the N-terminus ↑4/44/4
Ub detected at the C-terminus ↑0/44/4

With a request only at the beginning, all four designs had Ub detected only there. Adding the request at the end produced detectable Ub at both ends in all four designs.

Experiment 2. Generating both requested domains at different binder lengths

We tested whether FBindGen could generate both ubiquitin and Kunitz domains at different binder lengths. We requested a binder with ubiquitin followed by Kunitz. Kunitz is a domain commonly found in protease inhibitors. At each of four total lengths, FBindGen generated 16 complete sequences.

LengthDesignsBoth domains detectedDetection rate
1441616/16100%
1521616/16100%
1601616/16100%
1721614/1687.5%
Total6462/6496.9%

Of the 62 designs with both domains detected, 11 had Kunitz cores with pLDDT ≥ 70 and Cα RMSD ≤ 2 Å. Seven of these 11 also had ipTM ≥ 0.60. RMSD was measured after aligning the Kunitz core alone to a natural Kunitz structure.

Experiment 3. Generating requested domains without specified boundaries

In Experiment 2, we manually set the binder length and domain positions. Here, we tested whether users could obtain the requested domains by providing only the target and domain names. The workflow used templates to recommend binder length and domain boundaries.

We generated four designs per request for the same target. For the two-domain request, the workflow specified Ub before Kunitz.

Requested domainsLengthAll domains detectedSuccess rateipTM ↑
Ubiquitin794/44/40.468
Kunitz584/43/40.176
Ubiquitin + Kunitz1533/41/40.339

Success required all requested domains to be detected, each with pLDDT ≥ 70 and Cα RMSD ≤ 3 Å. RMSD was measured after aligning each domain separately to its natural structure.

Experiment 4. Generating Kunitz at different requested positions

We evaluated FBindGen's ability to design binders when domains were requested at different sequence positions. Experiment 2 varied binder length, and Experiment 3 tested generation without user-specified lengths or boundaries.

We kept binder length at 172 residues and the Ub request at residues 1–76. We first requested Kunitz at residues 93–150. We also requested it 8 or 16 residues earlier or later. The requested Kunitz and Ub regions did not overlap. We generated eight sequences at each position. We used the success criteria from Experiment 3.

Requested Kunitz regionBoth domains detectedSuccess rateipTM ↑
77–1345/80/80.383
85–1426/80/80.426
93–150 (original)8/82/80.285
101–1588/84/80.463
109–1668/84/80.564
Total35/4010/40—

2. On DB5.5, 21.3% of designs reach the reported 0.8 threshold. Separate aggregate function and interface metrics also leave unclear how often the same design satisfies both objectives.

To assess both goals together, we counted designs with ipTM ≥ 0.8 and ProTrek > 0.25. All 27 FBindGen designs meeting the ipTM threshold also met the ProTrek threshold.

DB5.5 designs meeting both prediction thresholds

MethodPassing designsPass rate ↑
BoltzGen3/1262.38%
Proteina-Complexa7/1265.56%
FBindGen27/12621.43%

3. The annotation-switch experiment evaluates text alignment without reporting whether binding compatibility survives the switch.

In DB5.5 cases, we replaced the binder's annotations with those from a different complex while keeping the target fixed. Mean ipTM decreased. This decrease is expected, as large changes in functional annotation positions can affect the predicted binding interface.

Annotation inputipTM ↑
Original0.399
Replaced0.158

4. Main generation evaluations use only one sequence per method per case.

For each of 275 EnzBind cases, we generated 100 sequences per method. A design passed if it had ipAE < 7 Å, pLDDT > 90, and ProTrek > 0.3. Passing structures were clustered within each case using Foldseek (TM-score threshold 0.5), and the cluster counts were summed across cases.

MethodPassing designsStructural clusters
BoltzGen16/27,500 (0.058%)3
Proteina-Complexa153/27,500 (0.556%)5
FBindGen159/27,500 (0.578%)14

5. Training exclusions use dates and benchmark PDB identifiers, but the described procedure does not establish that all homologous training examples are excluded. This leaves generalization to genuinely unfamiliar protein families unresolved.

We designed binders for two human targets with less than 30% sequence similarity to the training data. HNMT is histamine N-methyltransferase (UniProt P50135; PDB 1JQD). IDI2 is isopentenyl-diphosphate delta-isomerase 2 (UniProt Q9BXS1; PDB 2PNY).

We generated 1,000 sequences for each target. The table shows one selected binder design for each target.

TargetBinder lengthipTM ↑ipAE (Å) ↓Complex pLDDT ↑Binder pLDDT ↑
HNMT820.8267.7889.4386.58
IDI2870.8628.6086.3081.48

FBindGen generated designs with ipTM > 0.8 for both targets, despite their low sequence similarity to the training data.

6. Additionally, the model also needs to show its diversity and novelty.

Diversity

We assessed diversity using all 100 designs for each of the 275 EnzBind cases. For each case, we calculated the average percentage of amino acid positions that differed across all sequence pairs and counted unique sequences. Both measures were then averaged equally across cases.

MetricBoltzGenProteina-ComplexaFBindGen
Diversity (%) ↑74.93%87.79%78.90%
Unique sequences / 100 ↑99.9595.2499.89

Novelty

For each design, we found the closest sequence in UniRef50 and the closest structure in PDB. The table reports average similarity across all designs. TM-score measures structural similarity; higher scores mean more similar structures. Lower similarity indicates greater novelty.

MetricBoltzGenProteina-ComplexaFBindGen
UniRef50 identity (%) ↓37.99%39.76%29.56%
PDB TM-score ↓0.63080.60200.4884

FBindGen had lower average sequence and structure similarity than both baselines.

We hope these results address your concerns. We would be happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.