acceptodds
Under review as a conference paper at ICLR 2027

Information-Dense Synthesis for Molecular Discovery

Abstract

Machine learning can accelerate molecular discovery by designing molecules and planning experiments. However, many scientific challenges demand molecules with very rare properties, and in this sparse setting, existing algorithms offer little gain over random guessing. We propose a method to efficiently search large regions of molecular space using algorithmically controlled stochastic synthesis. Rather than design, make and test individual molecules, we design and make complex mixtures, test them as a pool, then deconvolute the molecule-activity map. We optimize synthesis to encode maximal information. Theoretically, this approach can reduce the number of experiments required to find the optimal molecule among candidates from to or . In simulation, on estimated peptide fitness landscapes, it finds active molecules with an order of magnitude fewer experiments than existing Bayesian optimization methods.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Bayesian experimental design, Bayesian optimal experimental design, applications to molecular discovery, automated labs, lab-in-the-loop.

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_r7TyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes an approach to efficiently search large regions of molecular spaces for molecules with desired properties. The pruning of the large spaces is done with the lab in the loop approach. Specifically, Bayesian analysis is employed to successively refine the posterior over these spaces when prior for the t+1 step is updated by wet lab experiments on the set of posterior samples at t.

Critical strengths
  1. The paper is very well written. The problem, the solution as well the limitations are all well described. Heartfelt thanks to the authors for writing something which is understandable and does not use convoluted vocabulary.
  2. The approach makes sense and the limitations are well thought of.
  3. The problem is very important, so an approach which may help in the efficiency aspect of it would have very large impact and can open new directions.
Critical weaknesses

The overall idea of trying to prune the search space for such a problem seems very intuitive. I wonder why prior work has not considered this. As a disclaimer I am not an expert in the problem domain so I might be wrong here. But with a bit of search I could find these works : https://pubs.rsc.org/dd/article/4/10/2910/889087/Adaptive-subspace-Bayesian-optimization-over and https://pubmed.ncbi.nlm.nih.gov/34985908/ which atleast at a very high level seems to build methodologies to prune the search space. the first one even uses Bayesian optimisation techniques.

I then wonder about novelty of the contribution. Does it come from lab in the loop proposal or a methodology which has yet not been applied to the current setting?

Questions for authors

What is the novelty of your approach? Does it come from lab in the loop proposal or a methodology which has yet not been applied to the current setting? Are there no prior works which apply efficient search for the current setting?

3 comments

Thank you for your time and careful reading. We greatly appreciate your thorough and clear review, questions, and comments.

Questions:

What is the novelty of your approach? Does it come from lab in the loop proposal or a methodology which has yet not been applied to the current setting?

Thank you for the question. The core novelty stems from designing and synthesizing mixtures of molecules, rather than individual molecules. Achieving this new lab-in-the-loop strategy required new methodology. Previous methods for Bayesian optimization, for instance, do not consider designing and testing mixtures, since computationally it is not as fast to test a mixture as a single point.

Are there no prior works which apply efficient search for the current setting?

Thank you for the important question. We do not know of prior work which applies the EIG to optimize molecular mixtures. Prior work focuses almost entirely on individual synthesis (e.g. Romero & Arnold, 2009; Romero et al., 2013; Frey et al., 2025; Rapp et al., 2024; Liu et al., 2025b; Mints et al., 2022; Hastrup et al., 2026). Recent work has considered mixtures, but they do not optimize them, rather choosing the mixtures at random, and running just one experiment rather than an iterative optimization campaign (e.g. Liu et al., 2025a, Aldridge et al., 2026). We have updated our related work section (Sec 4) to clarify this point.

Concerns:

The overall idea of trying to prune the search space for such a problem seems very intuitive. I wonder why prior work has not considered this. As a disclaimer I am not an expert in the problem domain so I might be wrong here. But with a bit of search I could find these works : https://pubs.rsc.org/dd/article/4/10/2910/889087/Adaptive-subspace-Bayesian-optimization-over and https://pubmed.ncbi.nlm.nih.gov/34985908/ which atleast at a very high level seems to build methodologies to prune the search space. the first one even uses Bayesian optimisation techniques.

Thank you for the comment, and for the useful pointers. Both MolDAIS and the second article you cite are individual synthesis methods, like our baselines. MolDAIS is a GP based method, and it uses the EI acquisition function. It is very similar to the LOCK EI, and Hamming EI baseline, described on l. 364-366 of the main text. The second article focuses not on laboratory experiments but on simulations. In these simulations, it evaluates one molecular design at a time.

In short, while prior work has put considerable effort into pruning the search space, it has paid less attention to what can be accomplished physically in the lab versus in silico. Our approach starts from the observation that it is possible, and just as efficient, to make mixtures rather than individual molecules, and this leads us to new and effective methodology. We have updated the related work section to further clarify where our approach sits in the literature (Sec. 4).

References (in next comment)

References

Matthew Aldridge, Oliver Johnson, and Jonathan Scarlett. Group Testing: An Information Theory Perspective. Foundations and Trends® in Communications and Information Theory, 23(1-2): 1–221, May 2026. ISSN 1567-2190, 1567-2328. doi: 10.1108/FTCIT-11-2025-0150. URL http://arxiv.org/abs/1902.06002. arXiv:1902.06002 [cs.IT].

Nathan C. Frey, Isidro H¨otzel, Samuel D. Stanton, Ryan Kelly, Robert G. Alberstein, Emily K. Makowski, Karolis Martinkus, Daniel Berenberg, Jack Bevers, Tyler Bryson, Pamela Chan, Yongmei Chen, Alicja Czubaty, Tamica D’Souza, Henri Dwyer, Anna Dziewulska, James W. Fairman, Allen Goodman, Jennifer Hofmann, Henry Isaacson, Aya Ismail, Samantha James, Taylor Joren, Simon Kelow, James R. Kiefer, Matthieu Kirchmeyer, Joseph Kleinhenz, James T. Koerber, Julien Lafrance-Vanasse, Andrew Leaver-Fay, Jae Hyeon Lee, Edith Lee, Donald Lee, Wei Ching Liang, Joshua Yao-Yu Lin, Sidney Lisanza, Andreas Loukas, Jan Ludwiczak, Sai Pooja Mahajan, Omar Mahmood, Homa Mohammadi-Peyhani, Santrupti Nerli, Ji Won Park, Jaewoo Park, Stephen Ra, Sarah Robinson, Saeed Saremi, Franziska Seeger, Imee Sinha, Anna M. Sokol, Natasa Tagasovska, Hao To, Edward Wagstaff, Amy Wang, Andrew M. Watkins, Blair Wilson, Shuang Wu, Karina Zadorozhny, John Marioni, Aviv Regev, Yan Wu, Kyunghyun Cho, Richard Bonneau, and Vladimir Gligorijevi´c. Lab-in-the-loop therapeutic antibody design with deep learning, February 2025. URL http://biorxiv.org/lookup/doi/10.1101/2025.02.19.639050.

Bjarke Hastrup, Francois Cornet, Tejs Vegge, and Arghya Bhowmik. AtomComposer: Discovering Chemical Space from First Principles with Reinforcement Learning, May 2026. URL http://arxiv.org/abs/2605.28287. arXiv:2605.28287 [cs.LG].

Nuo Liu, Walaa E. Kattan, Benjamin E. Mead, Conner Kummerlowe, Thomas Cheng, Sarah Ingabire, Jaime H. Cheah, Christian K. Soule, Anita Vrcic, Jane K. McIninch, Sergio Triana, Manuel Guzman, Tyler T. Dao, Joshua M. Peters, Kristen E. Lowder, Lorin Crawford, Ava P. Amini, Paul C. Blainey, William C. Hahn, Brian Cleary, Bryan Bryson, Peter S. Winter, Srivatsan Raghavan, and Alex K. Shalek. Scalable, compressed phenotypic screening using pooled perturbations. Nature Biotechnology, 43(8):1324–1336, August 2025a. ISSN 1546-1696. doi: 10.1038/s41587-024-02403-z. URL https://www.nature.com/articles/s41587-024-02403-z.

Yuanbin Liu, Ata Madanchi, Andy S. Anker, Lena Simine, and Volker L. Deringer. The amorphous state as a frontier in computational materials design. Nature Reviews Materials, 10(3):228–241, March 2025b. ISSN 2058-8437. doi: 10.1038/s41578-024-00754-2. URL https://www.nature.com/articles/s41578-024-00754-2.

Vladislav A. Mints, Jack K. Pedersen, Alexander Bagger, Jonathan Quinson, Andy S. Anker, Kirsten M. Ø. Jensen, Jan Rossmeisl, and Matthias Arenz. Exploring the Composition Space of High-Entropy Alloy Nanoparticles for the Electrocatalytic H2 /CO Oxidation with Bayesian Optimization. ACS Catalysis, 12(18):11263–11271, September 2022. ISSN 2155-5435, 2155-5435. doi: 10.1021/acscatal.2c02563. URL https://pubs.acs.org/doi/10.1021/acscatal.2c02563.

Jacob T. Rapp, Bennett J. Bremer, and Philip A. Romero. Self-driving laboratories to autonomously navigate the protein fitness landscape. Nature Chemical Engineering, 1(1):97–107, January 2024. ISSN 2948-1198. doi: 10.1038/s44286-023-00002-4. URL https://www.nature.com/articles/s44286-023-00002-4

Philip A. Romero and Frances H. Arnold. Exploring protein fitness landscapes by directed evolution. Nature Reviews Molecular Cell Biology, 10(12):866–876, December 2009. ISSN 1471-0080. doi: 10.1038/nrm2805. URL https://www.nature.com/articles/nrm2805.

Philip A. Romero, Andreas Krause, and Frances H. Arnold. Navigating the protein fitness landscape with Gaussian processes. Proceedings of the National Academy of Sciences, 110(3), January 2013. ISSN 0027-8424, 1091-6490. doi: 10.1073/pnas.1215251110. URL https://pnas.org/doi/full/10.1073/pnas.1215251110.

Reviewer_r7TyOpenReview ↗

Thanks for the response

Dear Authors,

Thank you for your reponse. I have no further questions. I will update my score based on the reviewer discussion.

Reviewer_KwroOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  1. LIDS (Learning with Information-Dense Synthesis) is proposed to accelerate the discovery of rare active molecules by designing, synthesizing, and testing molecular mixtures, then inferring the optimal molecule from pooled assay measurements.
  2. Algorithms are proposed to design stochastic syntheses by maximizing the expected information gain about the optimal molecule and to learn the molecule-activity map from pooled assay measurements using approximate Bayesian inference.
Critical strengths
  1. The research approach is insightful: it expands the scope of optimization from individual molecules to controllably synthesized molecular mixtures. By increasing the amount of information extracted from each measurement, it addresses the challenge of discovering rare, active molecules.
  2. The theoretical analysis is quite thorough: it not only demonstrates the advantages in experimental complexity under ideal conditions but also examines how measurement noise constrains information gain and the overall effectiveness of the method.
Critical weaknesses
  1. Reliance on strong measurement assumptions: The method assumes that mixture activity is a linear combination of individual molecular activities. However, real experiments may involve intermolecular interactions and nonlinear dose-response relationships.
  2. High computational cost: Mixture design requires optimizing the expected information gain, while posterior inference relies on extensive sampling, resulting in substantial computational overhead.
Questions for authors

Could the authors report the computational time required for the different posterior sample sizes shown in Figure 6, in order to illustrate the trade-off between computational cost and optimization performance?

1 comment

Thank you for your time and attention to the paper, and for your clear and comprehensive questions and comments.

Questions:

Could the authors report the computational time required for the different posterior sample sizes shown in Figure 6, in order to illustrate the trade-off between computational cost and optimization performance?

Thank you for the question, and for highlighting this important point. We agree that compute/runtime is an important point of comparison. Each duplicate takes 0.9 hours using M=1000, 3.1 hours for M=10000, and 5.0 hours for M=100000. Per round, this corresponds to 10.6 min, 36.6 min, and 60.1 min, respectively. We have added Table 4, summarizing this information. In our experiments, LIDS requires less compute to reach high activity molecules (e.g. 5 hours vs 16 hours for LOCK on the Rubisco landscape) but more compute on a per-round basis (e.g. 60.1 min versus 9.6 min). Full details have been added in (App. P.1) Table 3.

Concerns:

Reliance on strong measurement assumptions: The method assumes that mixture activity is a linear combination of individual molecular activities. However, real experiments may involve intermolecular interactions and nonlinear dose-response relationships.

Thank you for highlighting this important point. Indeed LIDS relies on a linearity assumption, and while there is empirical evidence for it across many different assays (Liu et al., 2025a, Aldridge et al., 2026), for others it may fail. We believe developing methods that are robust to violations of this assumption is an important area for future work. One approach is to add an interaction term to the model, e.g. Ex∼q(x),x′∼q(x)[h(x,x′)]\mathbb{E}_{x \sim q(x), x' \sim q(x)}[h(x, x')]. Another approach is to develop methods that are more generally robust to model misspecification, building on advances in Bayesian inference and experimental design under misspecification, such as Gibbs optimal design (Huggins & Miller, 2019; Overstall, 2026). Developing methods that can diagnose misspecification and nonlinearities is also an important area for future work. We have added a discussion of the linearity assumption and this possible future work to the limitations section (Sec. 7, l. 510-513).

The framework of LIDS can be directly extended to model non-linear dose-response relationships. Consider the model y=g(⟨f,q⟩)+ϵy = g(\langle f,q \rangle) + \epsilon, where g:R→Rg: \mathbb{R} \to \mathbb{R} is the non-linear dose-response relationship. Under this model, one can optimize the EIG using the same bound as LIDS. We have added a description of this to Sec 7 (l. 513-517), and a discussion to the limitations section. However, we have not empirically assessed LIDS in this setting, which we leave for future work.

High computational cost: Mixture design requires optimizing the expected information gain, while posterior inference relies on extensive sampling, resulting in substantial computational overhead.

Thank you for raising this important point, indeed LIDS has increased computational cost per round even while it reduces the experimental cost. In our experiments, LIDS requires less compute to reach high activity molecules (e.g. 5 hours for LIDS vs 16 hours for LOCK on the Rubisco landscape) but more compute on a per-round basis (e.g. 60.1 min versus 9.6 min). Full details have been added in (App. P.1,) Table 3. For molecular discovery, the typical bottleneck of a lab-in-the-loop system is physical experimentation (it takes days or weeks to synthesize molecules and conduct assays), not computational (both LIDS and the baselines take minutes or hours to generate designs). We therefore prioritized experimental efficiency over computational speed.

References

Huggins, J.H. and Miller, J.W., 2019. Robust inference and model criticism using bagged posteriors. arXiv preprint arXiv:1912.07104.

Overstall, A.M., Holloway-Brown, J. and McGree, J.M., 2026. Gibbs optimal design of experiments. Statistical Science, 41(3), pp.582-597.

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

The method (LIDS) wants to search large regions of molecular space controlled by stochastic synthesis. The proposed method includes lab-in-the-loop learning. The motivation is to design mixtures rather than molecules directly. The results are tested on various oracles like sparse and protein language models, where there is further analysis on noise levels and proteins.

Critical strengths

Provide a framework to choose mixtures instead of molecules though maximizing the expected information gain, and how to they find the optimal molecule LIDS is able to find the optimal molecule in significantly lower rounds, shown through the empirical results, .

Critical weaknesses

While the number of rounds needed is lower, the computational cost appears to be high. Adding in additional information regarding the compute and runtime between methods would be beneficial. Based on Figure 4, although LIDS approaches the optimal very quickly it doesn't seem to converge with f(x*) like LOCK EI. The performance overall appears worse, and the reasoning for the LIDS 5 rounds restriction is due to cost. The individual synthesis methods ran for 100 rounds, so it is a bit concerning regarding how much LIDS speeds up in runtime and cost.

Questions for authors

Mainly addressing a weakness, what is the compute/runtime comparison between other methods that do not utilize a mixture? Similarly, how many extra molecules are generated/needed compared to methods that generate molecules directly?

2 comments

Thank your time and careful reading. We appreciate your comprehensive review, and your relevant questions and comments.

Questions:

what is the compute/runtime comparison between other methods that do not utilize a mixture?

Thank you for the important question, we agree that compute/runtime is an important point of comparison. In our experiments, LIDS requires less compute to reach high activity molecules (e.g. 5 hours for LIDS vs 16 hours for LOCK on the Rubisco landscape) but more compute on a per-round basis (e.g. 60.1 min versus 9.6 min). We have added a full compute/runtime comparison in App. P.1, Table 3.

how many extra molecules are generated/needed compared to methods that generate molecules directly?

Thank you for the excellent question.All our comparisons were done at a fixed number of physically made molecules, meaning specifically a fixed yield c, which is kept the same between LIDS and all baselines. We have updated our description of the empirical results to clarify this point (Section 6). Our theoretical and empirical results suggest that with larger yields, LIDS would be expected to perform even better: SNR depends on the ratio of yield to noise, SNR∝c2σ2SNR \propto \frac{c^2}{\sigma^2}, and with increasing SNR, LIDS reaches high activity molecules faster and faster (Fig. 4).

Concerns:

While the number of rounds needed is lower, the computational cost appears to be high. Adding in additional information regarding the compute and runtime between methods would be beneficial.

Thank you for the excellent suggestion. . Indeed, LIDS computational cost can be high. We have added compute and runtime comparisons (App. P.1, Table 3), discussed above. We also further examined LIDS' dependence on compute and runtime. Its computational cost scales based on the number of posterior samples M. Our empirical results use M=100000, which takes 60.1 min per round. Decreasing to M=10000 or M=1000 can significantly decrease the runtime, however this comes at the cost of performance (App. S, Fig. 9). We have added this information to Table 4. An important area of future work is improving the accuracy of the posterior approximation, while keeping the computational cost manageable. We have added further discussion in the Limitations section of the Discussion (Sec. 7).

Based on Figure 4, although LIDS approaches the optimal very quickly it doesn't seem to converge with f(x⋆)f(x^\star) like LOCK EI. The performance overall appears worse, and the reasoning for the LIDS 5 rounds restriction is due to cost.

Thank you for the comment, and for highlighting this limitation. We investigated the lack of convergence further. We found substantial variability across runs: some LIDS duplicates get stuck in a local optimum, while the rest converge on x⋆x^\star without issues. We have added a figure that shows the individual duplicates, not just the average (App. R.2, Fig. 8). We believe the duplicates become stuck due to inefficient posterior approximation (line 7 of Algorithm 1). We clarify this limitation in both the results section and the limitations section of the discussion. We also describe possible directions for future work to address this limitation using improved Bayesian experimental design algorithms.

The individual synthesis methods ran for 100 rounds, so it is a bit concerning regarding how much LIDS speeds up in runtime and cost.

Thank you for raising this important point, indeed LIDS has increased computational cost per round even while it reduces the experimental cost. For molecular discovery, the typical bottleneck of a lab-in-the-loop system is physical experimentation (it takes days or weeks to synthesize molecules and conduct assays), not computational (both LIDS and the baselines take minutes or hours to generate designs). We therefore prioritized experimental efficiency over computational speed.

Thank you for your response. I have no further questions.

Comments on the paper

1 comment

We are grateful to the reviewers for their time, attention, and careful reading. To address the reviewers questions and concerns we have made corrections to the paper, as detailed below and highlighted in blue in the main text of the updated pdf. To further improve the paper, we also added a graphical overview to aid understanding (Fig. 2) and expanded the related works section (Sec 4).

The major points raised were:

  1. LIDS' computational cost was unclear.

We added a detailed description of LIDS' computational cost in comparison to baselines (App. P and S, Table 3 and 4). Briefly, LIDS requires more compute per-round compared to previous methods, but requires less compute overall to find high activity molecules. We discuss this tradeoff in the Limitations section of the discussion (Sec. 7).

  1. LIDS assumes a linear combination of individual molecular activities.

This is an important assumption, and we have reiterated it clearly in the Limitations section of the discussion. We have also described several areas for future work that can relax this assumption or make the method more robust to violations (Sec. 7 l. 510-517).

  1. The number of molecules in a mixture was unclear.

We have clarified that all experiments are run with the same number of molecules (the yield c) in LIDS versus baseline (Sec. 6). We have also clarified the method's dependence on the total number of molecules.

  1. Empirical studies don't always converge

We investigated this failure in greater detail (App. R.2, Fig. 8) and believe it is likely due to poor posterior approximations (line 7 of Algorithm 1). We clarify this limitation in the results (l. 485-487) and the Limitations section of the Discussion (Section 7). We further describe how future work may overcome this challenge.

  1. The novelty of the approach was unclear.

The core novelty stems from designing and synthesizing mixtures of molecules, rather than individual molecules. Achieving this new lab-in-the-loop strategy required new methodology. We clarified this novelty further in the related works section (Sec. 4).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.