PaperlayerView exact revision on arXiv

Source-linked AI summary

dZiner: Rational Inverse Design of Materials with AI Agents

Mehrad Ansari, Jeffrey Watchorn, Carla E. Brown, Joseph S. Brown

arXiv:2410.03963v1physics.chem-ph

TL;DR

Materials design faces costly screening and a dependence on extensive training data, while scientific literature contains domain knowledge that can inform design. dZiner is an LLM agent that retrieves such guidance and iteratively proposes and evaluates candidates; across three tasks, Claude 3.5 Sonnet outperforms GPT-4o, with especially high success rates when using literature-derived guidance.

  • Problem

    Traditional screening is time- and resource-intensive, and data-driven surrogate approaches rely heavily on extensive training datasets that may be unavailable or costly to obtain.

    Source: §1 · p. 1

  • Method

    dZiner uses an LLM agent to retrieve domain guidance, propose candidates under design constraints, assess feasibility, and evaluate properties with surrogate models in an iterative workflow.

    Source: §1 · p. 2

  • Results

    Across CMC, WDR5 docking, and CO2 adsorption tasks, Claude 3.5 Sonnet significantly outperforms GPT-4o in both conditions, with especially high success rates when using literature-derived domain knowledge.

    Source: §3 · pp. 9–10

  • Takeaways & Limitations

    The framework adapts across several property-to-structure tasks, including surfactants, ligand and drug design, and metal-organic frameworks.

    Source: §3 · p. 9

    The discussion identifies interpreting images and schemes in supporting literature as future work and notes that SMILES can oversimplify complex structures such as MOFs.

    Source: §3.1 · p. 11

Abstract

from arXiv · show

Recent breakthroughs in machine learning and artificial intelligence, fueled by scientific data, are revolutionizing the discovery of new materials. Despite the wealth of existing scientific literature, the availability of both structured experimental data and chemical domain knowledge that can be easily integrated into data-driven workflows is limited. The motivation to integrate this information, as well as additional context from first-principle calculations and physics-informed deep learning surrogate models, is to enable efficient exploration of the relevant chemical space and to predict structure-property relationships of new materials a priori. Ultimately, such a framework could replicate the expertise of human subject-matter experts. In this work, we present dZiner, a chemist AI agent, powered by large language models (LLMs), that discovers new compounds with desired properties via inverse design (property-to-structure). In specific, the agent leverages domain-specific insights from foundational scientific literature to propose new materials with enhanced chemical properties, iteratively evaluating them using relevant surrogate models in a rational design process, while accounting for design constraints. The model supports both closed-loop and human-in-the-loop feedback cycles, enabling human-AI collaboration in molecular design with real-time property inference, and uncertainty and chemical feasibility assessment. We demonstrate the flexibility of this agent by applying it to various materials target properties, including surfactants, ligand and drug candidates, and metal-organic frameworks. Our approach holds promise to both accelerate the discovery of new materials and enable the targeted design of materials with desired functionalities. The methodology is available as an open-source software on https://github.com/mehradans92/dZiner.

1 Introduction

Materials discovery has moved toward data-driven methods, but traditional screening is costly and machine-learning approaches depend heavily on extensive training data. dZiner is introduced as an LLM agent that uses domain knowledge to propose and evaluate candidates iteratively, with optional chemist feedback.

1 Introduction

  • Materials innovation progressed from trial and error through scientific laws and computational chemistry to a fourth paradigm integrating theoretical, experimental, and computational methods with data-driven techniques.The passage illustrates computational chemistry with ab initio calculations and density functional theory [3].Source: §1 · pp. 1–2
  • Traditional screening is time- and resource-intensive, while data-driven surrogate models rely on extensive training datasets that may be unavailable or costly to obtain.The introduction contrasts these limits with human experts’ ability to use domain knowledge and reason from limited examples.Source: §1 · p. 2
  • LLMs’ natural-language capabilities and external-tool use motivate chemistry agents, with prior work spanning data extraction, chemical innovation, experiment planning, prediction, hypothesis generation, and molecular discovery.The introduction points to an overview of LLM-based autonomous agents in chemistry and materials [35].Source: §1 · p. 2
  • dZiner uses a textual starting structure and property task to retrieve domain guidance, apply constraints, validate candidates, estimate properties with a surrogate, and iterate toward convergence.A chemist can optionally review candidates and reasoning, then provide feedback or further constraints.Source: §1 · p. 2, Figure 1 · p. 3
  • The paper benchmarks dZiner across three materials inverse-design tasks, then discusses implications and limitations before detailing its methodology and tools.The benchmark section is followed by discussion and a methodology section covering agent toolkits, domain expertise, surrogate models, and synthesizability assessment.Source: §1 · pp. 2–3

2 Results

Across surfactant, WDR5 ligand, and MOF-linker design, dZiner iteratively modifies candidates under chemical constraints; results include improved predicted properties and a chemist-guided design cycle.

2.1 Surfactant Design and Critical Micelle Concentration Inference

  • Figure 2's surfactant campaign reduced expected CMC by roughly two orders; iteration 17 reached log(CMC) −1.257 after a 1.618 improvement.Fluorine additions produced the largest reduction, while the lowest-CMC candidate had an SA score of 3.33.Source: Figure 2 · p. 4, §2.1 · pp. 4–5

2.2 Drug Design and Targeted Docking Inference

  • For WDR5, docking reached −9.2 at iteration 8 versus −6.7 initially, and yielded ten analogs below −8.5; every iteration stayed under 600 g/mol.The authors describe docking as low fidelity and propose the ten analogs as candidates for synthesis.Source: Figure 3 · p. 6, §2.2 · p. 6

2.2.1 Human-in-the-loop Design

  • In human-guided WDR5 design, the final molecule scored −9.3, and dZiner enacted requested structural changes and generated positional isomers.The human chemist flagged a potential toxicophore and suggested relocating an amine based on a crystal structure.Source: Figure 4 · p. 7, §2.2.1 · pp. 7–8

2.3 MOF Organic Linker Design and CO2 Adsorption Capacity Inference

  • For MOF linkers, the accepted iteration 17 reached CO2 adsorption of 1.978 from 0.885 initially; iteration 19's 2.289 exceeded the 600 g/mol limit.Fluorination improved predicted adsorption, but the last three suggested molecules were not accepted because they exceeded the molecular-weight constraint.Source: Figure 5 · p. 8, §2.3 · p. 9

3 Discussion

Across three inverse-design tasks, dZiner performed better with literature-derived design guidance, while Claude 3.5 Sonnet outperformed GPT-4o. The discussion also highlights the workflow’s flexibility and identifies limitations in molecular representations and human-in-the-loop positional instructions.

3 Discussion

  • Across CMC, WDR5 docking, and CO2 adsorption, Claude 3.5 Sonnet outperformed GPT-4o in both settings, while guideline-free runs often failed at validity and property optimization.Without retrieved guidelines, agents primarily made random functional-group or element changes.Source: §3 · pp. 9–10
  • Success was assessed by best-candidate average improvement and per-iteration gains over the initial candidate; Claude also respected molecular-weight and forbidden-group constraints while optimizing targets.The improvement metric compares each run’s best candidate with iteration 0.Source: §3 · p. 10, Table 2 · p. 10
  • Natural-language prompts can combine property optimization with design constraints, while customizable surrogate models support property prediction and uncertainty estimation.The framework also provides chain-of-thought reasoning for more interpretable results.Source: §3 · p. 10
  • Human-in-the-loop users can give feedback, suggest candidate modifications, or add constraints; iterative design retained relative synthesizability in CMC and WDR5 tasks compared with completely generative approaches.Source: §3 · p. 11

3.1 Limitations and Future Work

  • SMILES can oversimplify complex structures such as MOFs, and human-in-the-loop runs struggled with ortho, meta, and para placement on rings bearing three or more substituents.The authors suggest multimodal data, domain-specific fine-tuning, more human feedback, or validation tools as possible improvements.Source: §3.1 · p. 11

4.1 AI Agent

dZiner uses an LLM agent that combines scientific-literature knowledge, chemical tools, and surrogate-model feedback in an iterative materials-design workflow. Its retrieval step selects relevant and diverse design guidance for modifying molecules toward a target property.

4.1 AI Agent

  • The agent uses LLM reasoning and actions to draw design insights from literature, datasets, and chemical tools, then evaluates modifications with surrogate models and synthesizability assessment.The implementation uses GPT-4o and Claude 3.5 Sonnet at temperature 0.3, with LangChain.Source: §4.1 · p. 11

4.1.2 Domain-expert Knowledge

  • Each design iteration uses the current molecule, its predicted property, the target, literature, and history to request a modification and guideline from the LLM.The proposed structure is checked for validity before it proceeds to property evaluation.Source: §4.1.2 · p. 12
  • For domain-expert guidance, dZiner retrieves relevant literature passages for the property query and ranks them with MMR to balance query relevance and diversity.The retrieval process embeds papers and queries, selects top-k passages, and adjusts k for context length; the default MMR λ is 0.5.Source: §4.1.2 · pp. 11–12

4.2 Domain-expert Surrogate Models

The domain-expert surrogate models cover surfactant CMC, ligand docking, and MOF CO2 adsorption. For two case studies, ensembles provide prediction averages and epistemic uncertainty estimates.

4.2 Domain-expert Surrogate Models

  • For ensemble-based case studies, the prediction average summarizes model outputs, while epistemic uncertainty quantifies disagreement among model estimates.The paper uses ensemble modeling for the surfactant and MOF case studies.Source: §4.2 · pp. 12–13

4.2.1 Critical Micelle Concentration Inference Model

  • The surfactant GCN has test-data RMSE 0.23 and R2 0.96 for nonionic surfactants; the study ensemble has mean RMSE 0.32.The model predicts log CMC from molecular structure and was trained on experimental measurements for 202 surfactants.Source: §4.2.1 · p. 13

4.2.2 Targeted Docking Inference Model

  • The WDR5 ligand model uses targeted molecular docking with Dockstring and AutoDock Vina, returning single docking measurements without rescoring.Docking used a 30 Angstrom grid box around the WDR5 binding pocket and default exhaustiveness and energy range.Source: §4.2.2 · p. 13

4.2.3 CO2 Adsorption Inference Model

  • For CO2 adsorption at 0.5 bar, MOFormer has MAE 0.545 mol/kg, while the fine-tuned linker-SMILES ensemble has MAE 0.894 mol/kg.MOFormer uses text-based MOFids and was pretrained on over 400,000 MOF structures.Source: §4.2.3 · p. 13

4.3 Synthesizability Assessment

The assessment combines chemical-validity checks with synthetic accessibility and, for the drug-design section, a quantitative drug-likeness estimate.

4.3 Synthesizability Assessment

  • RDKit checks SMILES parsing, atoms and bonds, valences, ring closure, implicit hydrogens, and aromatic systems to assess basic chemical validity.The workflow also uses synthetic accessibility score, based on PubChem fragments and a complexity penalty.Source: §4.3 · p. 13
  • For section 2.2, drug-likeness is additionally estimated with QED, which integrates molecular weight, logP, polar surface area, and hydrogen-bond donors and acceptors.Source: §4.3 · p. 14

Supporting Information for dZiner: Rational Inverse

The supporting-information record identifies the dZiner paper and its authors.

Supporting Information for dZiner: Rational Inverse

1 Surfactant Design and Critical Micelle Concentration Inference with GPT-4o

dZiner iteratively modified a surfactant to lower predicted CMC, using literature-derived guidance and a molecular-weight constraint. The design combined tail and head-group changes, with the final candidate reaching the lowest reported log(CMC).

1 Surfactant Design and Critical Micelle Concentration Inference with GPT-4o

  • Adding methyl-type units to the hydrophobic tail in iteration 1 and replacing tail hydrogens with fluorine in iteration 3 reduced CMC across the first three iterations.A branching change in iteration 2 was rejected because evaluation showed no improvement over iteration 1.Source: §1 · p. 20
  • At iteration 8, log(CMC) reached -0.424 after a further hydrophobic-tail extension, following the largest reduction from 0.633 to 0.102 at iteration 7.The overall CMC reduction was roughly two orders of magnitude; the accepted molecule had 0.74 Tanimoto similarity to the starting molecule after eight iterations.Source: §1 · p. 20, Figure S1 · p. 21
  • Replacing the head group with amide-linked cyclic ethers drove the largest iteration-7 log(CMC) reduction, while the lowest-CMC candidate had an SA score of 3.29.Across the experiments, SA scores ranged from 2.80 to 4.01; the final iteration further lengthened the hydrophobic tail.Source: §1 · p. 20
  • The 100-molecule visualizations report no potentially unstable functional groups for GPT-4o or Claude 3.5 Sonnet, and mark invalid SMILES for review.The captions do not state that no invalid SMILES were generated in these surfactant runs.Source: Figure S2 · p. 22, Figure S3 · p. 23

2 Drug Design and Targeted Docking Inference with GPT-4o

The WDR5 case studies used iterative docking feedback to refine candidate structures, and also tested medicinal-chemist feedback in a human-in-the-loop workflow. Both workflows produced improved docking candidates, with the human-in-the-loop runs yielding diverse structures.

2 Drug Design and Targeted Docking Inference with GPT-4o

  • Docking score fell just over two orders of magnitude; the best reported score was -8.4 after iteration 11 added trifluoromethyl to the butoxy-substituted structure.Earlier aromatic, cyano, and ether modifications improved docking scores, while later pentoxy and hexoxy substitutions did not improve on the butoxy result.Source: §2 · p. 23, Figure S4 · p. 24
  • The Claude 3.5 Sonnet panel reports no invalid or potentially unstable molecules, while the GPT-4o panel marks potentially unstable groups and invalid SMILES.The chain-of-thought example reports a 0.46 Tanimoto similarity to the initial WDR5 candidate after substantial structural changes.Source: Figure S4 · p. 24, Figure S5 · p. 25, Figure S6 · p. 26

2.1 Human-in-the-loop design with GPT-4o

  • Human-in-the-loop experiments produced diverse structures with improved WDR5 docking scores, and dZiner enacted requested changes, including reverting to prior iterations.Medicinal chemists with no prior AI experience could use the cooperative workflow; accepted changes were evaluated using docking scores and human input.Source: §2.1 · p. 26, Figure S7 · p. 27

3 MOF Organic Linker Design and CO2 Adsorption Capacity Inference with GPT-4o

dZiner iteratively modified organic linkers to improve CO2 adsorption, evaluating chemical changes under molecular-weight and functional-group constraints. Electronegative atoms and nitrogen-containing groups were reported as effective modifications.

3 MOF Organic Linker Design and CO2 Adsorption Capacity Inference with GPT-4o

  • Hydroxyl and amino additions raised adsorption to 0.992, a pyridine ring raised it to 1.278, and fluorine raised it to 1.644; a later chlorine addition yielded 1.409.The account identifies electronegative atoms and nitrogen-containing groups as the most effective combination; molecular weight increased from 430.424 to 516.945 g/mol.Source: §3 · p. 27
  • The SA score peaked at 4.595 in iteration 5 after hydroxyl additions, indicating increased synthetic complexity in that step.Across the optimization, molecular weight rose from 430.424 g/mol at iteration 0 to 516.945 g/mol at iteration 9.Source: §3 · p. 27
  • The constrained design asked for MW <600 g/mol and exclusion of nitrosylated, chloro-, and fluoro-amines; its 100-molecule visualization reports no invalid SMILES.The figure also describes literature-derived design guidelines.Source: Figure S8 · p. 28, Figure S11 · p. 31
  • 75% improvement in CO2 adsorption was reported under the MW <600 g/mol constraint; adsorption reached 1.644 in iteration 8 after fluorine was introduced.The iteration-8 value was the highest observed in the described run.Source: §3 · p. 27, Figure S9 · p. 29

Related papers

Loading 2410.03963v1…