acceptodds
Under review as a conference paper at ICLR 2027

FIXED-CARDINALITY JOINT DIFFUSION FOR COMBINATORIAL PROTEIN DESIGN

Abstract

Fixed-cardinality design seeks a batch of sequences with exactly K mutations relative to a reference, a constraint libraries impose at cloning that post-hoc filtering cannot restore. We introduce FC-JD (Fixed-Cardinality Joint Diffusion), a two-stream diffusion model. Its position stream diffuses mutation masks on the Johnson graph J(N, K), making exact-K a construction invariant. Its simulation-free objective replaces forward rollouts with an analytic posterior over reverse transitions. We answer three questions. (Q1) Can FC-JD guarantee exact-K generation without post-hoc correction? (Q2) How does FC-JD compare with search-based methods (simulated annealing, Bayesian optimization) under the same batch budget? (Q3) When is FC-JD useful, and what are its limitations? In a pre-registered, oracle-scored benchmark (three complete-oracle landscapes, one partial-oracle case study, one regression-only dataset, four arms), FC-JD produced exact-K designs in every round, achieved substantially higher valid rates on partial-oracle landscapes, delivered batches in under two seconds, and descriptively matched or exceeded GP-BO on the extreme tail in every oracle-evaluable cell at K ∈ 5, 7. Surrogate-guided simulated annealing retained the best mean fitness on all four oracle-evaluable landscapes, and pre-registered criteria confirm this division of labor: FC-JD meets the generation criteria (valid, time) while SA96 keeps the best mean. We position FC-JD as a fast, diversity-preserving generator for exact-K design. Additionally, ESM-2-conditioned variants and cross-target transfer are reported as supporting analyses.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein design, discrete diffusion, combinatorial optimization, fixed-cardinality, Johnson graph, deep mutational scanning, generative models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_7mtROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

This paper proposes Fixed-Cardinality Joint Diffusion (FC-JD), a discrete diffusion framework for combinatorial protein design under exact mutation-cardinality constraint. The method models mutation positions through a diffusion process on the Johnson graph, where each transition preserves exactly K mutated positions, and uses a separate residue diffusion stream to model amino-acid identities when needed. The authors further introduce a simulation-free training objective based on analytically computed reverse-transition posteriors.

The proposed method is evaluated on several protein fitness landscapes, with models trained on low-order mutations and tested on higher-order designs. The experiments compare FC-JD against simulated annealing and Bayesian optimization in terms of fitness, diversity, validity, and computational cost. Results show that FC-JD guarantees exact-K generation, produces diverse candidate batches with low sampling latency, and achieves competitive extreme-tail fitness compared with Bayesian optimization in the reported settings.

Critical strengths
  1. Well-motivated problem formulation. This paper addresses the problem that a practically relevant setting where generated protein sequences must contain exactly K mutations relative to a reference. Enforcing this constraint directly during generation is a meaningful alternative to post-hoc filtering or rejection sampling.

  2. Interesting methodological design. Modeling mutation-position sets through diffusion on the Johnson graph provides a natural way to preserve the exact-K constraint throughout the generation process. The use of an analytically computed reverse-transition posterior is also an interesting aspect of the framework.

  3. Comprehensive evaluation and transparent reporting. The paper evaluates the proposed method across multiple protein fitness landscapes and considers several relevant metrics, including fitness, diversity, validity, and computational efficiency.

Critical weaknesses

1. Missing comparisons with constrained generative baselines.
The current baselines do not establish the advantages of Johnson-graph diffusion over simpler exact-K methods, such as stochastic Top-K sampling or other fixed-cardinality generators. Unconstrained diffusion only demonstrates the need for cardinality enforcement. Comparisons with diversity-oriented methods such as GFlowNets (Jain et al., ICML 2022) would further strengthen the evaluation.

2. Limited evidence of a superior quality–diversity trade-off.
SA achieves higher mean fitness across all four oracle-evaluable landscapes, while BO offers comparable diversity in several settings. The extreme-tail comparison selects the better result between FC-JD and P+R, and only 2 of 7 pre-registered criteria are met. A systematic quality–diversity comparison under matched budgets is needed to clarify FC-JD's practical advantages.

3. Limited evaluation on realistic protein design tasks.
Most experiments use small combinatorial landscapes, with antibody tasks restricted to predefined substitutions and DHFR using a small alphabet. GFP lacks reliable exact-K fitness evaluation. Thus, scalability to longer protein sequences, larger amino-acid alphabets, and complex mutational interactions remains unclear.

Questions for authors
  1. Could the authors clarify what advantages Johnson-graph diffusion offers over simpler exact-K generation methods, such as stochastic Top-K sampling? Would comparisons with constrained generative baselines or GFlowNets change the conclusions?

  2. Given that SA achieves higher mean fitness and only 2 of 7 pre-registered criteria are met, under what conditions does FC-JD provide a clear advantage in the quality–diversity trade-off? Could the authors provide comparisons under matched computational and candidate budgets?

  3. How well is FC-JD expected to scale to realistic protein design tasks involving longer sequences and larger amino-acid alphabets? Is there additional evidence beyond the current small combinatorial landscapes?

I would be willing to reconsider my rating if the authors can adequately address these questions.

Reviewer_JoUAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper focusses on the problem of generating a batch of sequences with exactly K mutations relative to a reference or fixed cardinality design. To solve this problem, the authors propose Fixed-Cardinality Joint Diffusion (FC-JD). FC-JD uses two stream diffusion architecture where diffusing mutation masks on Johnson graph they guarantee fixed cardinality. Further, they derive an analytic posterior over reverse transitions enabling simulation-free training. They compare their approach with simulated annealing and Bayesian optimization on pre-registered, oracle-evaluated benchmark. FC-JD generates exact-K designs faster and with higher diversity than simulated annealing. However, simulated annealing achieves higher mean fitness.

Critical strengths
  • I found the approach of diffusing mutation masks on Johnson graph to guarantee fixed cardinality interesting. Constrained diffusion is an area with a wide range of practical applications where FC-JD idea could be applicable.

  • Their analytic posterior or closed-form kernel makes training simulation-free thus no simulation of the forward process is required during training, which is another technically valuable contribution.

  • I also likes their detailed experiments. They compared FC-JD wrt surrogate-guided simulated annealing, Gaussian-process based Bayesian optimization, and unconstrained diffusion baselines. Training on K=3 and testing K = 5,7 is also informative of the generalization. They also evaluated various metrics like mean fitness, fitness hit rates, diversity, exact- K validity, and runtime, giving clearer picture of strengths and weaknesses of different methods.

Critical weaknesses

FC-JD achieves faster generation and greater diversity than simulated annealing, but SA consistently does better in mean fitness. Similarly, BO based approach achieves comparable diversity with mixed fitness performance. Moreover, FC-JD satisfies only 2/7 pre-registered success criteria (Table 2). These results make me wonder if the proposed FS-JD approach offers a sufficiently compelling quality–diversity–cost tradeoff to justify its additional method level complexity.

Questions for authors
  • Could the authors clarify the advantages of the proposed diffusion framework over simpler approaches for enforcing fixed-cardinality constraints? I am trying to understand what aspects of the observed performance you can attributed to the diffusion mechanism?

  • I appreciate you highlighting the trade-offs different approaches achieve. It would be useful to further highlight scenarios where one should prefer one over another -- would help practitioners. For example, in which applications the observed improvements in diversity and runtime preferred at the cost fitness?

  • How well these findings generalize to protein-design settings involving more diverse amino-acid substitutions and larger design spaces?

  • Can authors provide motivation behind their testing setup -- train on K = 3 and test of K = 5, 7

Reviewer_g4Z6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses a very practical bottleneck in protein library design: generating sequences with strictly exact-K mutations. Standard (unconstrained) discrete diffusion models struggle.

To fix this, the authors propose FC-JD, decoupling generation into a position stream and a residue stream. By diffusing position masks directly over the Johnson graph J(N, K) with a closed-form analytic posterior, the exact-K cardinality is naturally guaranteed by construction without any post-hoc filtering.

Across pre-registered benchmarks, FC-JD strictly hits 100% validity and generates batches in under two seconds, outperforming GP-BO on extreme tail fitness. The authors also honestly show that surrogate-guided simulated annealing still wins on mean fitness.

Critical strengths
  1. This is a principled method for exact mutation constraint. Restricting the position mask diffusion to the Johnson graph J(N, K) via a single-swap Markov chain is elegant and theoretically justified. Unlike typical unconstrained masked diffusion models that may need a lot of post filtering, FC-JD guarantees exact-K satisfaction by design.

  2. The experiments are well methodologically designed and validated. The authors pre-registered 7 specific evaluation criteria, executed 40,000 evaluated rounds across 10x10 seeds, and honestly reported that FC-JD met only 2/7 criteria (failing on mean fitness against SA96), rather than cherry-picking metrics.

Critical weaknesses
  1. While FC-JD excels in exact-K validity, batch inference speed (<2 s), and sequence diversity, surrogate-guided simulated annealing (SA96) consistently outperforms FC-JD in mean fitness across all four oracle-evaluable landscapes (with gaps ranging from 0.21 to 0.42). In downstream experimental library synthesis where total budget is small and mean hit probability is prioritized, local search or hybrid pipelines (such as Proposal+Ranker) may still be preferred.
Questions for authors
  1. Scalability with sequence length N: The benchmark primarily validates small combinatorial sites (N <= 16). How does the sampling dynamics and computational overhead of the Johnson graph scale when N is substantially larger (e.g., N > 200)?

  2. On CB6 and DHFR, applying the surrogate ranker (P+R) caused substantial diversity collapse compared to raw FC-JD. Could the authors clarify why this collapse might happen and what this implies for practical application of FC-JD?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.