💬 NLP

Closed-Loop Bayesian Molecular Inverse Design with Semantic LLM Surrogates

This paper introduces \textbf{\method}, a closed-loop Bayesian molecular inverse design framework that leverages a frozen large language model as a semantic surrogate to reason directly over textual instructions and optimization history, thereby selecting informative reference molecules to guide a frozen generator and outperforming or matching traditional Gaussian process baselines across drug and material design tasks.

Original authors: Yaoyao Xu, Xinjian Zhao, Xiaozhuang Song, Lei Bai, Tianshu Yu

Published 2026-08-25

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yaoyao Xu, Xinjian Zhao, Xiaozhuang Song, Lei Bai, Tianshu Yu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The search for new medicines and advanced materials often begins with a question: what does a molecule need to look like to perform a specific job? Scientists call this inverse design. Instead of testing thousands of existing chemicals to see what they do, researchers try to build a molecule from scratch that fits a precise set of requirements, such as blocking a virus or allowing gas to pass through a filter. The challenge is that the universe of possible chemical structures is vast, estimated to contain around 10 to the power of 60 drug-like molecules. Finding the few that work is like searching for a needle in a haystack that is constantly changing shape. Traditionally, computers have tried to solve this by generating random candidates and checking them against a computer model, but this process is often inefficient, producing many poor options before finding a good one.

A team of researchers at the Chinese University of Hong Kong, Shenzhen, and the Shanghai Artificial Intelligence Laboratory has developed a new way to guide this search. They created a system called BoMolLLM, which treats the process of finding a good molecule not as a single guess, but as a conversation between a computer program and a large language model. In their approach, the computer generates a batch of molecules, and a specialized artificial intelligence acts as a guide. This guide reads the results of the previous batch, analyzes which structures worked best, and then writes a new set of instructions for the generator. Instead of just picking the best molecule from the last round, the guide selects a few interesting examples and explains why they are promising, effectively teaching the generator what to try next. This creates a closed loop where the search gets smarter with every step, refining its focus until it finds molecules that closely match the desired properties.

The researchers tested this system on two very different types of problems. First, they asked it to design molecules for drug discovery, specifically looking for compounds that could interact with HIV, cross the blood-brain barrier, or inhibit a specific enzyme called BACE. These tasks require the molecule to have a binary outcome: it either works or it does not. Second, they tested it on material science, asking the system to design polymers that could control the flow of gases like carbon dioxide, oxygen, and nitrogen. These tasks are more continuous, requiring the molecule to hit a specific target value for permeability rather than just passing or failing a test. In both cases, the system was given a limited number of attempts to improve its results, simulating a scenario where testing real chemicals is expensive and time-consuming.

The results showed that this conversational approach outperformed standard methods. When the researchers compared their system to a one-shot attempt, where the computer tries to generate the perfect molecule in a single go without any feedback, the new method produced a much higher fraction of successful candidates. It also performed as well as, or better than, traditional mathematical optimization techniques that rely on complex statistical models to predict the next best step. The key difference was that the new system could "read" the history of its own attempts. It could look at a list of molecules that failed and say, "These had too many heavy metals," or "These lacked the right ring structure," and then pass that insight to the next round of generation. This ability to reason over the text of the chemical structures and the scores they received allowed the system to navigate the chemical space more effectively than methods that only looked at compressed numerical data.

One of the most interesting findings was that the system adapted its strategy depending on the type of problem. For the drug targets, which had clear pass-or-fail criteria, the system worked best when it simply pointed to the best examples from the previous round. The specific structures of those molecules were enough to guide the next step. However, for the material science tasks, where the goal was to hit a precise numerical target, the system needed more help. In these cases, the guide provided a short, one-sentence summary of the strategy, such as "focus on fluorinated structures," which helped the generator understand the broader pattern needed to reach the target. This suggests that while the system is flexible, the way it communicates its findings must be tailored to the nature of the problem it is solving.

The researchers also observed that the system naturally shifted its behavior over time. In the early rounds, it explored a wide variety of chemical structures, trying many different shapes and compositions to see what was possible. As the rounds progressed, it began to focus more narrowly, refining the most promising structures it had found. This mirrors the way a human scientist might work, starting with broad ideas and gradually narrowing down to the most viable options. The system did not just get lucky; it consistently discovered high-scoring molecules earlier and more reliably than the other methods tested. By the end of the process, the molecules it generated were clustered in the regions of chemical space where the best candidates were found, rather than being scattered randomly.

This work demonstrates that large language models can serve as effective guides for scientific discovery, not just by generating text, but by making strategic decisions based on data. The system does not replace the need for real-world testing, but it significantly improves the quality of the candidates that reach the lab bench. It turns the search for new molecules into a more transparent process, where the reasoning behind each step can be read and understood. The researchers found that by treating the optimization history as a story to be analyzed rather than just a list of numbers, they could build a more efficient and intelligent search engine for the molecular world. This approach offers a promising path forward for accelerating the discovery of new drugs and materials, turning a vast and chaotic search space into a guided journey toward a solution.

Technical Summary: Closed-Loop Bayesian Molecular Inverse Design with Semantic LLM Surrogates

Problem Statement

Practical molecular inverse design is rarely a one-shot generation problem. Instead, it typically manifests as a closed-loop candidate-pool enrichment task. Under a limited oracle budget (e.g., computational cost for property prediction), the objective is not merely to find a single high-scoring molecule but to increase the fraction of generated molecules that match a desired property profile.

While Bayesian Optimization (BO) offers a natural framework for this setting, standard approaches rely on Gaussian Process (GP) surrogates operating in compressed continuous latent embeddings. This paradigm discards substructural and reference-similarity signals that chemists naturally use to guide search. Furthermore, GPs scale poorly in high dimensions and often require dimensionality reduction, which further compresses the chemical signal available to the surrogate. Recent Large Language Models (LLMs) have demonstrated strong capabilities in molecular representation and reasoning but are typically used in a one-shot manner, lacking mechanisms to explicitly utilize oracle feedback to correct search trajectories iteratively.

Methodology: BoMolLLM

The authors propose BoMolLLM, a closed-loop framework where the surrogate (rather than the generator) is the locus of design choice. The framework treats molecular optimization histories as semantic objects containing SMILES strings, target descriptions, and oracle scores.

Core Components

Frozen Molecular Generator ( $M$ ): An instruction-conditioned generator (instantiated as Llamole) that maps text prompts to candidate molecules. It remains frozen throughout the optimization process.
Task Oracle ( $O_\tau$ ): Evaluates generated molecules against specific targets (binary classification for drug tasks like HIV, BBBP, BACE; continuous regression for material tasks like $CO_2$ , $O_2$ , $N_2$ permeability).
Semantic LLM Surrogate ( $L$ ): A frozen LLM that acts as the decision module. Unlike classical BO, it operates directly on the discrete optimization history in its native textual form.

The Optimization Loop

The process runs for $T$ rounds:

Round 0 (Warm Start): The generator produces an initial batch based solely on the original design instruction.
Rounds $t \ge 1$ :
1. The surrogate reads the accumulated history $D_{t-1}$ , the task context, and the current round index.
2. Under an exploration-exploitation principle, the surrogate selects a set of top- $k$ reference molecules ( $R_t$ ) from the history.
3. Optionally, the surrogate generates a concise guidance sentence ( $g_t$ ) summarizing the trajectory.
4. These elements are formatted into an augmented prompt $p_t$ for the generator.
5. The generator produces a new batch $S_t$ , which is scored and added to the history.

Surrogate Instantiations

The paper compares two surrogate types within the same loop:

Semantic LLM Surrogate (BoMolLLM): The LLM reasons over the history to select references and provide guidance. It interprets pretrained chemical knowledge as a prior, in-context conditioning as a posterior update, and reference selection as the acquisition step.
Classical GP Surrogate: A baseline adapted from BOPRO, fitting a GP on PCA-compressed embeddings of the generator's conditioning vectors, optimizing a LogEI acquisition function, and retrieving nearest neighbors as references.

Key Contributions

Closed-Loop Framework with Replaceable Surrogate: The authors formulate molecular inverse design as a candidate-pool enrichment problem where the surrogate layer is isolated as the locus of design choice. This allows for the direct comparison of classical GP and LLM-based surrogates within the same iterative loop.
Semantic LLM Surrogate: BoMolLLM introduces an LLM instantiation that operates directly on SMILES strings, scores, and target context. It provides an inspectable optimization trace in natural language, offering design rationales unavailable from numeric surrogates.
Empirical Evidence and Analysis: The study demonstrates that BoMolLLM matches or surpasses GP-BO baselines across three generator backbones (Qwen, Mistral, Llama) and six tasks. Crucially, the surrogate exhibits a BO-style transition from exploration to exploitation and discovers high-scoring molecules earlier than baselines.

Experimental Results

Experiments were conducted on the MolQA benchmark, covering drug design (HIV, BBBP, BACE) and material design ( $CO_2$ , $O_2$ , $N_2$ ).

Performance: BoMolLLM consistently improves over one-shot prompting (Llamole-OneShot) and is competitive with or stronger than GP-BO baselines. The advantage is most pronounced in drug tasks, where BoMolLLM achieves the best AUC scores across all targets for Mistral and Llama backbones.
Domain-Dependent Interface: An ablation study revealed a critical interface difference:
- Drug Tasks (Binary Targets): "Reference-only" transfer (Top-k molecules + scores) works best. The structural signal from high-scoring references is sufficient.
- Material Tasks (Continuous Targets): Adding a "concise surrogate summary" (guidance) significantly improves performance. The softer, graded nature of continuous targets benefits from the surrogate's ability to compress trajectory patterns into actionable directions.
Search Behavior: Analysis of the search trajectory shows that BoMolLLM discovers high-scoring molecules earlier and more reliably than baselines. Mechanism analysis confirms the surrogate balances exploration (selecting diverse structures) and exploitation (selecting high-scoring structures) over rounds.
Chemical Space Allocation: Visualization via t-SNE of ECFP4 fingerprints shows that BoMolLLM concentrates its search budget in regions enriched with high-scoring candidates without collapsing into a single cluster, unlike random or one-shot approaches.

Significance and Claims

The paper claims that LLM surrogates are a useful closed-loop alternative to classical BO surrogates for discrete molecular design.

By shifting the surrogate decision from a compressed numerical space to a chemically meaningful history expressed in molecules and text, BoMolLLM preserves the decision logic of Bayesian Optimization while making the search process inspectable and interpretable. The framework demonstrates that the "semantic" nature of molecular optimization histories can be leveraged directly by LLMs to guide generation, offering a path toward more transparent and effective inverse design workflows.

The authors note limitations, acknowledging that their oracles are learned predictors rather than experimental validators and that the LLM surrogate is not a calibrated probabilistic model with theoretical guarantees like classical BO. However, the results suggest that the semantic approach effectively enriches candidate pools under limited oracle budgets.