acceptodds
Under review as a conference paper at ICLR 2027

Continuous Variational Synthesis

Abstract

Biological machine learning was long bottlenecked by our ability to synthesize designed DNA. Variational synthesis models control chemical reactions to physically manufacture quadrillions of designed sequences in DNA. However, training these variational synthesis models is challenging: constraints on chemical synthesis can force many parameters into a discrete space, limiting the ability to pre-train and fine-tune. In this article we train "free" variational synthesis models using stochastic gradient descent in continuous space, and then discretize with post-training quantization. This enables variational synthesis models to satisfy stringent reward criteria, while still producing diverse designs, achieving a strictly dominating quality-diversity Pareto frontier. We demonstrate by training variational synthesis models of peptides, antibody CDRH3s, regulatory DNA elements, and enzymes. *In silico* performance is maintained *in vitro*.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
chemical synthesis, quantization, protein design, DNA design

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_sMdKOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Continuous Variational Synthesis (cVS), which addresses the challenge of training variational biological synthesis models under discrete chemical constraints through continuous spatial gradient optimization and post-training quantization. The method combines forward KL pre-training, reverse KL fine-tuning, reduction of gradient variance, and parameter quantization, and is validated on four tasks: enzymes, HLA-presenting peptides, antibody CDRH3, and regulatory DNA. Experiments show that cVS outperforms existing methods in multiple quality-diversity trade-off metrics, and the feasibility of the method is verified through real DNA synthesis and sequencing.

Critical strengths
  1. Continuous optimization followed by quantization is intuitive, practical, and compatible with gradient-based training.

  2. Includes in vitro DNA synthesis and sequencing, providing evidence beyond purely computational evaluation.

  3. Demonstrates improved quality–diversity trade-offs against EM VS and PGLD.

Critical weaknesses
  1. Ablations suggest that continuous relaxation and quantization contribute more than the proposed variance-reduction techniques. Their individual contributions need clearer quantification.

  2. Some diversity metrics, particularly theoretical diversity, may not fully reflect experimentally realizable unique sequences or functional diversity.

Questions for authors
  1. How much improvement comes from continuous relaxation, quantization, pretraining, and variance reduction individually?

  2. How does cVS compare with alternative continuous optimization or discrete optimization methods under identical constraints?

  3. How robust is the method to different synthesis catalogs, numbers of reaction wells, and quantization errors?

1 comment

We are grateful to the reviewer for their time, their careful reading and their constructive suggestions. A point-by-point response is below (since some comments and questions cover similar points, we address them jointly).

"Ablations suggest that continuous relaxation and quantization contribute more than the proposed variance-reduction techniques. Their individual contributions need clearer quantification."... "How much improvement comes from continuous relaxation, quantization, pretraining, and variance reduction individually?"

Thank you for the valuable suggestion, we have run additional experiments quantifying the contribution of each component of our training algorithm. First, we ran additional experiments evaluating each component of our variance reduction strategy. Each separate technique offers a small but reliable improvement, which add together to a more substantial gap over the naive approach (Figure 21). Next, we evaluated our quantization strategy in greater depth. Our new experiments show our reverse KL quantization objective offers large improvements over a naive Euclidean quantization objective. We also considered quantizing during training rather than only post-training, by adding a penalty for violations of the quantization objective. We find this only offers modest improvements, at the cost of reduced wetware inter-operability (Figure 22). To evaluate the continuous relaxation, we can compare to PGLD, which does not use a continuous relaxation; here we find substantial gains (Figures 2, 3, 4). Finally, we found that pre-training offers substantial gains (Figure 11).

"Some diversity metrics, particularly theoretical diversity, may not fully reflect experimentally realizable unique sequences or functional diversity."

Thank you for raising this important point. Indeed, we only include “theoretical diversity” in the peptide design evaluation because it was used in a previous benchmark proposed in the PGLD paper (Sussex et al. 2026); we do not believe it is a useful estimate of diversity in practice, precisely because it does not fully reflect experimentally realizable unique sequences. We clarify these points in Section 2.2. We also agree that evaluating functional diversity is both important and challenging. In the absence of functional measurements, we evaluated diversity according to a range of different strategies: besides KL divergence to the prior, we looked at Shannon entropy (Figure 13), the kernel 2-Renyi entropy across many different kernel bandwidths (Figure 14), and the kernelized Stein discrepancy (KSD-B) across different kernels and bandwidths (Figure 18). The KSD-B is of particular note, since it is an integral probability metric which works by integrating over an infinite number of possible sequence-function relationships, effectively averaging over many possible measures of functional diversity (Amin et al. 2023).

"How does cVS compare with alternative continuous optimization or discrete optimization methods under identical constraints?"

Thank you for raising this important point. We do not know of any directly comparable method besides PGLD, which focuses just on the constraint of equal nucleotide mixtures. We have performed additional experiments evaluating different optimizers for the training phase of cVS, including REINFORCE leave-one-out, a widely used method (Figure 21). We also performed additional experiments evaluating different quantization methods used for discrete optimization, including quantization with the standard Euclidean distance, which performs much worse (Figure 22).

"How robust is the method to different synthesis catalogs, numbers of reaction wells, and quantization errors?"

Thank you for raising this important point. We have evaluated the method with different synthesis catalogs, reaction wells and quantization errors, observing substantial robustness. As the number of reaction wells increases, the performance steadily increases, approaching an asymptotic limit; small changes in wells have little effect (Figure 10). As the synthesis catalog increases in size, the quantization error decreases, and performance improves steadily (Figure 3b). Some amount of quantization error is unavoidable when synthesis uses mixtures of nucleotides, rather than trinucleotides (codons), even with a large catalog. Despite this irreducible quantization error, we do not observe a large gap between the performance of the unconstrained free cVS model and the quantized model in practice.

Reviewer_2YX2OpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work trains model to give the synthetic procedure to synthesize a molecule (DNA/RNA/Enzyme) X, then it fine tunes the model to select the optimal synthetic route. The problem the paper address is of interest to both biologists and generative modelling researchers as in synthesis, mixtures are pulled from a finite set of flasks and preset nucleotide mixes.

Critical strengths
  1. Wet lab synthesized 101510^15 molecules from a quantized CVS model and showed that quantization and synthesis doesn't break the in silico optimum. Fig 5.g shows the in vitro lands on top of the cloud of the in silico designed promoters, although in UMAP space.
  2. Borrows idea from LLM pretrain/fine-tunning in eq 4 and 5 to improve the pre-llm approach of EM VS which is optimization based.
Critical weaknesses
  1. θ\theta has a few meanings, the model weights and the nucleotides combinations living on Δ4\Delta^4 and others. They are confusing.
  2. There is concern of reward hacking. CVS is reward finetuned. For the CDRH3 experiment, the binding counts reaches 10^4, but in training data for the reward are their samples with such high binding counts? are you extrapolating the reward upwards far beyond training data? Can you convince that the reward can generalize to such high counts. Are there held out rewards you can check against.
Questions for authors

per weakness 2.

2 comments

We are grateful to the reviewer for their time, their careful reading and their constructive suggestions. A point-by-point response is below.

"Theta has a few meanings, the model weights and the nucleotides combinations living on \Delta^4 and others. They are confusing."

Thank you for the important question. Indeed, in variational synthesis, the model parameters (weights) correspond exactly to the nucleotide combinations (that is, concentrations) used during chemical synthesis. We are using this generative model to design an experiment: the model parameters specify the experiment to run. We have revised our background section to clarify this important point (Section 0.1; revisions in blue).

"There is concern of reward hacking. CVS is reward finetuned. For the CDRH3 experiment, the binding counts reaches 10^4, but in training data for the reward are their samples with such high binding counts? are you extrapolating the reward upwards far beyond training data? Can you convince that the reward can generalize to such high counts. Are there held out rewards you can check against."

Thank you for the important question, we agree that reward hacking is a major concern in molecular design . In this paper, the goal is to develop synthesis methods for a given reward model, not to evaluate the quality or generalization of specific reward models. Nonetheless, we took several measures to avoid reward hacking. Most critically, we design our libraries using a regularized reinforcement learning objective: in addition to the reward, we add the KL divergence to the prior (Eqn. 3). For the CDRH3 experiments, the reward models were also trained with regularization that explicitly encourages predictions to revert to zero outside the training data distribution (Weinstein et al. 2025).

We checked our reward models’ extrapolation and added additional evaluations to check the performance of our designs on held-out rewards. High counts, around 10^4, indeed appear in the training data. To check our designs against held-out rewards, we used an ensemble of models trained on the TCRm screening data, trained via 10-fold cross validation. We designed using one of these models as the reward; we used the remaining models, trained on different folds, to estimate the performance and provide a measure of uncertainty. With this evaluation, the estimated rewards from cVS drops, but they are still large in absolute terms, and higher than PGLD (Figure 20). We do not see evidence of reward hacking, as designs with higher training reward also achieve higher held-out rewards.

Dear Authors,

Thank you so much for the rebuttal, and yes figure 20 is illustrative. I don't have further questions and will continue recommend acceptance.

Reviewer_1B4TOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes Continuous Variational Synthesis (cVS), a framework for training generative models that control stochastic DNA synthesis. The method optimizes a mixture of independent categorical distributions in continuous space using forward and reverse KL objectives, followed by post-training quantization to satisfy hardware synthesis constraints. cVS shows general improvements over other baselines, especially on quality-diversity Pareto frontier.

Critical strengths
  1. Strong empirical results, especially the improved quality-diversity Pareto frontier compared with existing methods.

  2. evaluation across four different biological design tasks, including enzymes, peptides, antibodies, and regulatory DNA.

  3. in vitro validation showing that the synthesized DNA library maintains a distribution close to the intended computational distribution.

Critical weaknesses
  1. Accessibility and clarity. The paper is difficult to follow for a general ML audience due to extensive domain-specific terminology. The training data, reference distributions, and reward models are also not clearly presented across the different experiments.

2.motivation/justification for the restricted generative model. The paper does not sufficiently motivate the use of mixtures of independent categorical distributions compared with more expressive biological generative models that can capture dependencies between sequence positions. An alternative would be to generate diverse sequences using existing generative models and then filter or adapt them for synthesis. The practical trade-offs between these approaches are not discussed or evaluated.

3.Reliability of functional evaluation. Functional improvements are primarily measured using computational predictors without experimental validation. The predictors' accuracy, especially on potentially out-of-distribution generated sequences, is insufficiently evaluated. Details on predictor training and validation are not clearly presented in the paper, making the reported improvements difficult to assess.

4.(relatively minor) lack of experimental functional validation. The in vitro validation does not demonstrate that the generated sequences actually achieve better biological function.

Questions for authors
  1. What are the training datasets and reference distributions used for each task? A summarized explanation would improve clarity.

  2. What is the practical advantage of the proposed synthesis-compatible generative model over more expressive biological generative models followed by sequence filtering or other synthesis strategies? Is there an advantage in cost, scalability, or functional quality?

  3. Can the authors provide more details on the training and predictive performance of the independently trained functional evaluators, particularly on held-out data?

  4. Some improvements over PGLD appear modest, particularly in peptide design. Can the authors report variability across independent runs to establish statistical significance?

2 comments

We are grateful to the reviewer for their time, their careful reading and their constructive suggestions. A point-by-point response is below (since some comments and questions cover similar points, we address them jointly).

"Accessibility and clarity. The paper is difficult to follow for a general ML audience due to extensive domain-specific terminology. The training data, reference distributions, and reward models are also not clearly presented across the different experiments." ... "What are the training datasets and reference distributions used for each task? A summarized explanation would improve clarity."

Thank you for the valuable suggestion. We have added definitions of key domain-specific terms (see e.g. the revised Section 0.1). We have added an up-front summary of the training data, reference distributions and reward models across each experiment at the start of the results (Section 2, highlighted in blue).

"motivation/justification for the restricted generative model. The paper does not sufficiently motivate the use of mixtures of independent categorical distributions compared with more expressive biological generative models that can capture dependencies between sequence positions. An alternative would be to generate diverse sequences using existing generative models and then filter or adapt them for synthesis. The practical trade-offs between these approaches are not discussed or evaluated." ... "What is the practical advantage of the proposed synthesis-compatible generative model over more expressive biological generative models followed by sequence filtering or other synthesis strategies? Is there an advantage in cost, scalability, or functional quality?"

Thank you for the question. The key advantage of variational synthesis is cost and scale: by constraining the generative model to be synthesis compatible, we can physically manufacture quadrillions of independent samples from the model for the same price as synthesizing a few thousand individually, i.e. the synthesis constraint gives us a trillion-fold increase in scale of physically realized samples. The reason is that, unlike for other biological generative models, we do not need to make samples from the model individually, but rather can run stochastic synthesis in the lab according to the learned parameters of the model. While synthesis does impose constraints on the generative model, synthesis models are still expressive in practice, and can capture dependence between sequence positions; see Weinstein et al 2022 and Weinstein et al 2026a for detailed theoretical proofs and empirical evaluations. We have added a clear explanation of the motivation for using a synthesis-constrained model in Section 0.1 (highlighted in blue).

"Reliability of functional evaluation. Functional improvements are primarily measured using computational predictors without experimental validation. The predictors' accuracy, especially on potentially out-of-distribution generated sequences, is insufficiently evaluated. Details on predictor training and validation are not clearly presented in the paper, making the reported improvements difficult to assess." ... "Can the authors provide more details on the training and predictive performance of the independently trained functional evaluators, particularly on held-out data?"

Thank you for the helpful suggestion. Note our goal in this paper is not to evaluate the quality of the reward models, but rather to demonstrate a synthesis method which is compatible with diverse rewards.

We used trained evaluators for the antibody example (Section 2.3) and the regulatory element example (Section 2.4). We added evaluations for both on held-out data. The antibody predictor enriches for specific pHLA-binders by a factor of 100x over a baseline random predictor, when evaluated on held-out cells from the screen (on average across cross-validation folds) (Section B.3). The ATAC-seq predictor shows an AUROC of 0.91 on a held-out chromosome (Section B.4).

We also evaluated how sensitive performance is to inaccuracy in the predictor. For the antibody design challenge, we trained 10 separate predictors on separate cross-validation folds. Only one model was used for design. We used the others to re-evaluate the designs, and provide a measure of uncertainty. This performance estimate is somewhat lower than that from the reward model used for design, but still high in absolute terms, and still higher than PGLD (Figure 20). We do not see evidence that our designs are overfit to the reward model, suggesting the strategy of KL regularization to the prior is effective.

"Some improvements over PGLD appear modest, particularly in peptide design. Can the authors report variability across independent runs to establish statistical significance?"

Thank you for the suggestion. Indeed, in the peptide example, while cVS showed a major advantage over the previously published PGLD Pareto frontier (compare blue and light gray lines in Figure 2), cVS had a smaller advantage over our own hyperparameter-optimized version of PGLD (dark gray). We checked the statistical significance of this result: we found cVS shows a statistically significant advantage over hyperparameter-optimized PGLD in the hypervolume of the Pareto front (percentile bootstrap p < 0.05; see Figure 19).

"4.(relatively minor) lack of experimental functional validation. The in vitro validation does not demonstrate that the generated sequences actually achieve better biological function."

We agree with this limitation, and have added a discussion of it in our Discussion section (Section 3). Functional experimental validation is ongoing, and we expect to have data to present in time for the camera-ready deadline.

Comments on the paper

1 comment

We are grateful to all the reviewers for their time, their careful reading and their constructive suggestions. We have run additional experiments and analyses, and revised the paper, to address each comment and suggestion. Our revisions are highlighted in blue in the updated paper pdf. To summarize, the major points raised by the reviewers were:

  1. Uncertainty in the reward model, and the possibility of reward hacking.

We ran additional experiments and evaluations checking (a) the performance of the reward models on held-out data, and (b) the performance of our designs on held-out reward models (Figure 20).

  1. The contribution of each component of the training algorithm, especially the variance-reduction and quantization techniques.

We ran additional experiments and evaluations ablating each component of our training algorithm, and comparing to strong baseline alternatives. We test a range of variance reduction strategies (Figure 21) and a range of quantization strategies (Figure 22).

  1. Clarity explaining the motivation and interpretation of the synthesis model.

We edited the paper to clarify the motivation behind the restrictions on the synthesis model, and the relationship between the model parameters and the laboratory synthesis (revised Section 0.1).

Full point-by-point responses to each reviewer are found below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.