Information-Dense Synthesis for Molecular Discovery

Kasper K. Jakobsen; Eli N. Weinstein

stat.MLOct 6, 2026

Information-Dense Synthesis for Molecular Discovery

Authors: Kasper K. Jakobsen, Eli N. Weinstein

Organizations: Department of Chemistry, Technical University of Denmark, Kgs. Lyngby, DK.

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 $d$ candidates from $\mathcal{O}(d)$ to $\mathcal{O}(\log d)$ or $\mathcal{O}(1)$ . In simulation, on estimated protein fitness landscapes, it finds active molecules with an order of magnitude fewer experiments than existing Bayesian optimization methods.

Figures & tables

Figure 1 : Lab-in-the-loop speedup. Optimizing an in silico oracle with LIDS (blue) versus baselines ( Sec. 6 ).

Figure 2 : Overview. Information-dense synthesis uses mixtures of molecules

q

, instead of single molecules

x

. By testing the mixtures as a pool, we encode the activity from many molecules into experiment outcomes

y

. Compared to individual synthesis, this can carry much more information about the optimal molecule

x^{\star}

, especially on sparse molecule-activity maps

\mathrm{f}

Figure 3 : EIG tradeoffs. EIG from Example 3 , plotted alongside MI upper bounds from Proposition 2 , as a function of

d_{q}=|\mathrm{supp}(q)|

Figure 4 : Two rounds of LIDS on a sparse synthetic oracle ( Sec. 6.1 ). UMAP projection of DNA sequences. Row 1: samples from the prior $\pi(x^{\star})$ (left), the mixtures designed at $t=1$ by LIDS $q^{1}_{1:E}$ (middle), and the same mixtures, colored by their measured activity $y^{1}_{1:E}$ (right). Row 2: the posterior after one experiment round, $\mathrm{p}(x^{\star}\mid\mathcal{D}_{1}=(q^{1}_{1:E},y^{1}_{1:E}))$ (left), the mixtures designed at $t=2$ by LIDS $q^{2}_{1:E}$ (middle), and the same mixtures, colored by their activity $y^{2}_{1:E}$ (right). Row 3: the posterior after two experiment rounds, $\mathrm{p}(x^{\star}\mid\mathcal{D}_{2}=(q^{1,2}_{1:E},y^{1,2}_{1:E}))$ (left), and the true activity (right), normalized to be a distribution for visualization $\mathrm{f}(x)/\sum_{x\in\mathcal{X}}\mathrm{f}(x)$ . After these two rounds, LIDS has converged on the global optimum (star).

Table 5

Panel of Figure 5 — Figure 5 : Synthetic $\mathrm{f}$ with rare activity. LIDS, Thompson, and Hamming EI, at varying noise-level $\sigma$ , optimizing a synthetic oracle. The activity of the predicted optimum, $\mathrm{f}(\hat{x}^{\star}_{t})$ , is plotted as a function of experimental round $t$ , averaged over 10 duplicates. The 95% CI is calculated using a logit transform and the delta method, so that it stays bounded in the interval $[0,\mathrm{f}(x^{\star})]$ (See App. P ).

Panel of Figure 6 — Figure 6 : PLM oracle. LIDS, Thompson, Hamming EI, and LOCK EI, applied to a PLM oracle, with low noise ( $\textsc{PSNR}=80$ dB). The maximum activity of the 5 latest test molecules $\max\mathrm{f}(\hat{x}^{\star}_{t-5}),...,\mathrm{f}(\hat{x}^{\star}_{t})$ , is plotted as a function of the experimental round $t$ , averaged over 5 duplicates. The 95% CI is calculated using a logit transform and the delta method, so that it stays bounded in the interval $[0,\mathrm{f}(x^{\star})]$ (See App. P ).

Figure 7 : LIDS with synthesis errors. We add random errors to the designed mixtures

q_{t}

by replacing

\beta_{sl}

with

(1-\rho)\beta_{sl}+\rho\omega

where

\omega\sim\mathrm{Dir}(\vec{1}_{B})

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Notation	Description
$\mathcal{X}$	The molecular space, usually a Hamming space
$\mathcal{P}(\mathcal{X})$	$=\{\mathrm{p}(\cdot):\mathrm{p}(x)\geq 0,\int_{\mathcal{X}}\mathrm{p}(x)\mathrm{d}x=1\}$ The set of probability distributions on $\mathcal{X}$
$\mathrm{f}:\mathcal{X}\to\mathbb{R}_{+}$	The molecule-activity map
$q:\mathcal{X}\to\mathbb{R}_{+}$	The design - can be a mixture or a delta. The restriction on $q$ comes from the synthesis model.
$\bar{q}=\frac{q}{c}\in\mathcal{P}(\mathcal{X})$	The normalized design, where $c$ is the yield, the total number of molecules produced by a synthesis.
$c\in\mathbb{R}_{+}$	The chemical yield. The total number of molecules produced by the synthesis model.

Appendix

Table 9

Panel of Figure 8 — Figure 8 : Hierarchical causal model. (a) Our model can be viewed as a hierarchical causal model, where multiple molecules $x_{ij}$ affect the outcome $y_{i}$ in each experiment $i$ , and there are $c$ total molecules. (b) The collapsed hierarchical causal model, where $q^{x}$ is the distribution of $x$ (elsewhere we drop the superscript x for conciseness). This model leads to the estimator in Eqs. 3 and 4 .

Figure 9 : Posterior approximation. LIDS using different number of posterior samples

M

. The setup is the same as in Sec. 6.2 .

Explore similar work

LLM-Guided Dynamic Action Spaces for Synthesizable Molecular Optimization
Apr 9, 2026Tao Li, Kaiyuan Hou, Tuan Vinh +4Molecular Optimization Reaction
Department of Computer Science, Emory University · School of Computer Science, Carnegie Mellon University · MRC Brain Network Dynamics Unit, University of Oxford +3
Sample Efficient Generative Optimization for Molecular Design
Jul 14, 2026Sarina Kopf, Cristina Nevado, Philippe SchwallerMolecular Optimization Molecular Design
LIAC, ISIC, EPFL, Lausanne, Switzerland · Department of Chemistry, University of Zurich, Zurich, Switzerland · Centre of Competence in Research (NCCR) Catalysis, Zurich, Switzerland
Active Learning Enables Generation of Molecules that Advance the Known Pareto Front
Jan 3, 2025Evan R. Antoniuk, Peggy Li, Nathan Keilbart +3Molecular Generation Molecular Property Prediction

Apr 9, 2026, cs.LGJ/K move · Enter open · S save

LLM-Guided Dynamic Action Spaces for Synthesizable Molecular Optimization

Tao Li, Kaiyuan Hou, Tuan Vinh, Fanglei Xue+3

Department of Computer Science, Emory University · School of Computer Science, Carnegie Mellon University · MRC Brain Network Dynamics Unit, University of Oxford+3

Molecular Optimization Reaction High-Level Synthesis Action Space

Open paper AskPDF arXiv

Synthesizable molecular optimization seeks to improve target properties while ensuring that molecular modifications follow feasible synthetic pathways. Existing synthesis-aware methods typically rely on exploring a large space of candidate transformations defined by reaction templates and purchasable building blocks. This search becomes even more challenging when property improvement requires multiple reaction steps, as the space expands further along the pathway. To address this challenge, we introduce MolReAct, which reformulates molecular optimization as search over compact reaction spaces proposed by a tool-augmented large language model (LLM). At each step, the LLM combines its prior chemical knowledge with cheminformatics tools to identify a molecule-specific set of compatible reactions, preserving synthesizability while making multi-step optimization feasible. Given this compact action space, we further leverage Group Relative Policy Optimization (GRPO) with the terminal oracle reward to improve long-term decision-making over multiple reaction steps. Across diverse molecular optimization tasks, MolReAct achieves the highest Top-10 score on 11 of 14 tasks and the best sample efficiency on 12 of 14 tasks, outperforming existing baselines under limited oracle budgets. Beyond these gains, MolReAct also provides each optimized molecule with a template-grounded synthetic pathway.

Abstract

Explore similar work

LLM-Guided Dynamic Action Spaces for Synthesizable Molecular Optimization

Sample Efficient Generative Optimization for Molecular Design

Active Learning Enables Generation of Molecules that Advance the Known Pareto Front