Yan Xiang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Yan Xiang has a total of 62 co-authors. Here are the top co-authors by number of shared publications:
- Daniel Reker
Duke University
10 shared publications - Guang Lin
School of Mechanical Engineering, Purdue University, West Lafayette, USA
5 shared publications - Huai Sun
School of Chemistry and Chemical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
4 shared publications - Yu-Hang Tang
Computational Research Division, Lawrence Berkeley National Laboratory, Berkeley, California 94720, United States
4 shared publications - Zilu Zhang
Department of Biomedical Engineering
3 shared publications Hongyi Liu
School of Chemistry and Chemical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
2 shared publications
Latest Publications
- Open Access
- 21 October 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Data-efficient molecular and nanoparticle design via fine-tuned large language models
Designing molecules with prescribed properties is central to pharmaceutical and materials innovation, but vast chemical design spaces make development lengthy and costly. Generative artificial intelligence can accelerate molecular discovery, but its deployment in many real-world domains remains limited by the need for large task-specific training datasets. To enable generative design when property-labeled data are scarce, we present a data-efficient foundation-model adaptation strategy combining complementary positive and negative property prompts with parameter-efficient fine-tuning. Generated molecules are then ranked by molecular graph-kernel Gaussian process regression for structure-aware prioritization. In a controlled computational setting, the platform generated candidates with novel scaffolds satisfying multiple design objectives, including predicted target binding, synthesizability, and drug-likeness. Although only a few hundred training molecules per target met all three criteria, the model generated thousands of new molecules matching the desired profiles. For prospective validation, we applied the model to nanoparticle design, a domain of high clinical potential but limited excipient diversity. Fine-tuning on only 1,521 labeled drug-excipient pairs - including just 139 nanoparticle-forming examples identified new nanoparticle-forming excipients for the anticancer drug sorafenib. The model proposed lithocholic acid and the sorafenib metabolite M6, and experiments confirmed that both form stable sorafenib nanoparticles with high encapsulation efficiency and increased in vitro potency. These findings show that foundation-model fine-tuning with complementary prompts and graph-based prioritization enables data-efficient molecular design in low-data chemical domains.
- 21 October 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessData-efficient molecular and nanoparticle design via fine-tuned large language models
Designing molecules with prescribed properties is central to pharmaceutical and materials innovation, but vast chemical design spaces make development lengthy and costly. Generative artificial intelligence can accelerate molecular discovery, but its deployment in many real-world domains remains limited by the need for large task-specific training datasets. To enable generative design when property-labeled data are scarce, we present a data-efficient foundation-model adaptation strategy combining complementary positive and negative property prompts with parameter-efficient fine-tuning. Generated molecules are then ranked by molecular graph-kernel Gaussian process regression for structure-aware prioritization. In a controlled computational setting, the platform generated candidates with novel scaffolds satisfying multiple design objectives, including predicted target binding, synthesizability, and drug-likeness. Although only a few hundred training molecules per target met all three criteria, the model generated thousands of new molecules matching the desired profiles. For prospective validation, we applied the model to nanoparticle design, a domain of high clinical potential but limited excipient diversity. Fine-tuning on only 1,521 labeled drug-excipient pairs - including just 139 nanoparticle-forming examples identified new nanoparticle-forming excipients for the anticancer drug sorafenib. The model proposed lithocholic acid and the sorafenib metabolite M6, and experiments confirmed that both form stable sorafenib nanoparticles with high encapsulation efficiency and increased in vitro potency. These findings show that foundation-model fine-tuning with complementary prompts and graph-based prioritization enables data-efficient molecular design in low-data chemical domains.
- 30 July 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (16) , 9816-9828
Short-Term Memory Active Learning for Drug Development
Active learning is a powerful approach for efficient data selection in machine learning, particularly valuable in domains such as drug discovery where data acquisition is costly. Traditional active learning methods continuously expand the training set with newly labeled data but do not revisit the utility of the previously added data points. Here, we introduce short-term memory active learning (SMAL), a novel active learning framework that can remove training data during learning. We conceptualized, implemented, and evaluated multiple forgetting strategies based on prediction error and uncertainty metrics derived from random forest out-of-bag estimates. We benchmarked SMAL across six absorption, distribution, metabolism, excretion, and toxicity (ADMET) drug development data sets, demonstrating improved or competitive performance compared to classical active learning. We found that SMAL creates more balanced training sets and can reduce labeling costs through its unique data recycling mechanism. Furthermore, when tested on data sets with artificially introduced label errors, SMAL shows robust performance and autonomously identifies and discards corrupted data points. Beyond ADMET tasks, SMAL also shows potential utility for virtual screening: across 99 virtual-screening data sets with simulated temporal splits from the SIMPD benchmark, SMAL matched or outperformed classical active learning in 97.7% of comparisons. These results establish forgetting as a valuable addition to active learning, enabling more robust and efficient model development─especially in noisy, resource-constrained real-world settings.
- 30 July 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (16) , 9816-9828
Short-Term Memory Active Learning for Drug Development
Active learning is a powerful approach for efficient data selection in machine learning, particularly valuable in domains such as drug discovery where data acquisition is costly. Traditional active learning methods continuously expand the training set with newly labeled data but do not revisit the utility of the previously added data points. Here, we introduce short-term memory active learning (SMAL), a novel active learning framework that can remove training data during learning. We conceptualized, implemented, and evaluated multiple forgetting strategies based on prediction error and uncertainty metrics derived from random forest out-of-bag estimates. We benchmarked SMAL across six absorption, distribution, metabolism, excretion, and toxicity (ADMET) drug development data sets, demonstrating improved or competitive performance compared to classical active learning. We found that SMAL creates more balanced training sets and can reduce labeling costs through its unique data recycling mechanism. Furthermore, when tested on data sets with artificially introduced label errors, SMAL shows robust performance and autonomously identifies and discards corrupted data points. Beyond ADMET tasks, SMAL also shows potential utility for virtual screening: across 99 virtual-screening data sets with simulated temporal splits from the SIMPD benchmark, SMAL matched or outperformed classical active learning in 97.7% of comparisons. These results establish forgetting as a valuable addition to active learning, enabling more robust and efficient model development─especially in noisy, resource-constrained real-world settings.
- 24 June 2026
- Published by American Chemical Society (ACS) in ACS Pharmacology & Translational Science
- Vol. 9 (7) , 1756-1766
Identifying Antibiotic Effects of Investigational Drugs on Commensal Bacteria with Machine Learning
Many human-targeted medications have been found to impact patients’ gastrointestinal microbiomes, which has been proposed as an unrecognized source of drug side effects, comorbidities, and reduced treatment efficiencies. However, current methods for detecting such effects, such as patient sample analysis or in vitro high-throughput screening, are both labor- and resource-intensive. To accelerate the discovery of drug effects on the microbiome, we developed machine learning models that predict whether a small, drug-like molecule is likely to inhibit the growth of any of 40 representative human gut commensal microbes. We employed these models to virtually screen thousands of investigational drugs, revealing a strong propensity for human-targeted compounds to potentially modulate commensal microbes. Prospective in vitro validations uncovered two nonantibiotic drugs, the recently approved anti-cancer agent entrectinib and the clinical drug candidate PSI-697, to have previously unknown growth inhibition effects on multiple commensal gut microbes. Furthermore, we show that resistance to the effects of these drugs is mediated by known antibiotic resistance mechanisms BamB and TolC. Additionally, entrectinib significantly reduced microbial richness in a synthetic microbial model community. Taken together, our machine learning-assisted workflow and future extensions can triage microbiome-drug interactions to prioritize experimental testing and validation.
- 24 June 2026
- Published by American Chemical Society (ACS) in ACS Pharmacology & Translational Science
- Vol. 9 (7) , 1756-1766
Identifying Antibiotic Effects of Investigational Drugs on Commensal Bacteria with Machine Learning
Many human-targeted medications have been found to impact patients’ gastrointestinal microbiomes, which has been proposed as an unrecognized source of drug side effects, comorbidities, and reduced treatment efficiencies. However, current methods for detecting such effects, such as patient sample analysis or in vitro high-throughput screening, are both labor- and resource-intensive. To accelerate the discovery of drug effects on the microbiome, we developed machine learning models that predict whether a small, drug-like molecule is likely to inhibit the growth of any of 40 representative human gut commensal microbes. We employed these models to virtually screen thousands of investigational drugs, revealing a strong propensity for human-targeted compounds to potentially modulate commensal microbes. Prospective in vitro validations uncovered two nonantibiotic drugs, the recently approved anti-cancer agent entrectinib and the clinical drug candidate PSI-697, to have previously unknown growth inhibition effects on multiple commensal gut microbes. Furthermore, we show that resistance to the effects of these drugs is mediated by known antibiotic resistance mechanisms BamB and TolC. Additionally, entrectinib significantly reduced microbial richness in a synthetic microbial model community. Taken together, our machine learning-assisted workflow and future extensions can triage microbiome-drug interactions to prioritize experimental testing and validation.
- Open Access
- 6 April 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Physics-informed design of drug-excipient nanoparticles via free energy calculation and yoked deep learning
The discovery of self-assembling molecular systems is often constrained by the high computational cost of physics-based simulations and the limited generalizability of data-driven models trained on experimental datasets. This limits our ability to efficiently navigate massive design spaces. Here, we introduce a physics-informed discovery framework that integrates free-energy calculations with sequential-pool active learning and adaptive yoked deep learning to effectively navigate vast chemical coassembly spaces. Using pairwise free energy as a physically grounded descriptor of co-assembly propensity and excipient net charge as a proxy for interfacial stabilization, our approach predicts nanoparticle-forming drug–excipient combinations without relying on experimental data for model training. Our sequential-pool yoked deep learning strategy utilizes adaptive machine learning to reduce the number of expensive simulations required while maintaining predictive accuracy, enabling screening of 18 million molecular pairs. Prospective validations identified 24 drug nanoparticles using eight previously unknown stabilizing molecules. The new nanoparticles exhibit ultra-high drug loading, effectively improve the aqueous solubility of drugs, and maintain in vitro drug efficacy. Together, this work establishes a generalizable framework coupling molecular simulations and yoked deep learning to accelerate discovery in self-assembling molecular systems.
- 6 April 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessPhysics-informed design of drug-excipient nanoparticles via free energy calculation and yoked deep learning
The discovery of self-assembling molecular systems is often constrained by the high computational cost of physics-based simulations and the limited generalizability of data-driven models trained on experimental datasets. This limits our ability to efficiently navigate massive design spaces. Here, we introduce a physics-informed discovery framework that integrates free-energy calculations with sequential-pool active learning and adaptive yoked deep learning to effectively navigate vast chemical coassembly spaces. Using pairwise free energy as a physically grounded descriptor of co-assembly propensity and excipient net charge as a proxy for interfacial stabilization, our approach predicts nanoparticle-forming drug–excipient combinations without relying on experimental data for model training. Our sequential-pool yoked deep learning strategy utilizes adaptive machine learning to reduce the number of expensive simulations required while maintaining predictive accuracy, enabling screening of 18 million molecular pairs. Prospective validations identified 24 drug nanoparticles using eight previously unknown stabilizing molecules. The new nanoparticles exhibit ultra-high drug loading, effectively improve the aqueous solubility of drugs, and maintain in vitro drug efficacy. Together, this work establishes a generalizable framework coupling molecular simulations and yoked deep learning to accelerate discovery in self-assembling molecular systems.
- Open Access
- 15 October 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Genetically encoded sterol-modification of a synthetic intrinsically disordered protein leads to diverse self-assembly behavior
Post translational modifications (PTMs) of proteins are used by natural systems to expand beyond the 20 canonical amino acids. The variation introduced at the sequence level by PTMs after expression leads to changes in both the structure and function of proteins. PTMs expand the chemical repertoire from which new biomaterials can be constructed. Inspired by the post-translational addition of cholesterol to proteins, we have synthesized five new hybrid lipid-protein biomaterials called Sterol modified polypeptides (STaMPs). These STaMPs consist of an elastin like polypeptide (ELP) conjugated to a sterol, namely coprostanol, epicoprostanol, androstanol, galeterone, or dehydroepiandrosterone. We show that the STaMPs form self-assembled nanostructures such as spherical and cylindrical micelles, and vesicles. Furthermore, the sterols modify the typical LCST behavior of ELPs in a predictable fashion depending on the hydrophobicity of the sterol appended. STaMPs could expand the possibilities for drug delivery by enabling the physical encapsulation of hydrophobic drugs, increasing the solubility and half-life of sterol-based chemotherapeutics, and forming thermosensitive drug depots.
- 15 October 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessGenetically encoded sterol-modification of a synthetic intrinsically disordered protein leads to diverse self-assembly behavior
Post translational modifications (PTMs) of proteins are used by natural systems to expand beyond the 20 canonical amino acids. The variation introduced at the sequence level by PTMs after expression leads to changes in both the structure and function of proteins. PTMs expand the chemical repertoire from which new biomaterials can be constructed. Inspired by the post-translational addition of cholesterol to proteins, we have synthesized five new hybrid lipid-protein biomaterials called Sterol modified polypeptides (STaMPs). These STaMPs consist of an elastin like polypeptide (ELP) conjugated to a sterol, namely coprostanol, epicoprostanol, androstanol, galeterone, or dehydroepiandrosterone. We show that the STaMPs form self-assembled nanostructures such as spherical and cylindrical micelles, and vesicles. Furthermore, the sterols modify the typical LCST behavior of ELPs in a predictable fashion depending on the hydrophobicity of the sterol appended. STaMPs could expand the possibilities for drug delivery by enabling the physical encapsulation of hydrophobic drugs, increasing the solubility and half-life of sterol-based chemotherapeutics, and forming thermosensitive drug depots.
- 11 September 2025
- Published by American Chemical Society (ACS) in ACS Nano
- Vol. 19 (37) , 33288-33296
TuNa-AI: A Hybrid Kernel Machine To Design Tunable Nanoparticles for Drug Delivery
Artificial intelligence (AI) has the potential to transform nanoparticle development for drug delivery; however, existing strategies typically optimize either material selection or component ratios in isolation. To enable simultaneous optimization of both, we integrated an automated liquid handling platform with machine learning to systematically explore the nanoparticle formulation space. A data set comprising 1275 distinct formulations (spanning drug molecules, excipients, and synthesis molar ratios) was generated, resulting in a 42.9% increase in successful nanoparticle formation through composition optimization. We developed a bespoke hybrid kernel machine that couples molecular feature learning with relative compositional inference, enhancing the modeling of formulation outcomes across chemical spaces. This hybrid kernel significantly improved prediction performance across three kernel-based algorithms, with a support vector machine (SVM) achieving superior performance when using our kernel compared to standard kernels and outperforming all other machine learning architectures, including transformer-based deep neural networks. Using SVM-guided predictions, we successfully formulated the difficult-to-encapsulate venetoclax with optimized taurocholic acid ratios, yielding enhanced in vitro efficacy against Kasumi-1 leukemia cells. In a second case study, our AI-guided platform reduced excipient usage by 75% in a trametinib formulation while preserving the in vitro efficacy and in vivo pharmacokinetics relative to the standard formulation. Taken together, this study establishes a generalizable framework that combines robotic experimentation, kernel machine learning, and experimental validation to accelerate nanoparticle composition optimization for drug delivery.
- 11 September 2025
- Published by American Chemical Society (ACS) in ACS Nano
- Vol. 19 (37) , 33288-33296
TuNa-AI: A Hybrid Kernel Machine To Design Tunable Nanoparticles for Drug Delivery
Artificial intelligence (AI) has the potential to transform nanoparticle development for drug delivery; however, existing strategies typically optimize either material selection or component ratios in isolation. To enable simultaneous optimization of both, we integrated an automated liquid handling platform with machine learning to systematically explore the nanoparticle formulation space. A data set comprising 1275 distinct formulations (spanning drug molecules, excipients, and synthesis molar ratios) was generated, resulting in a 42.9% increase in successful nanoparticle formation through composition optimization. We developed a bespoke hybrid kernel machine that couples molecular feature learning with relative compositional inference, enhancing the modeling of formulation outcomes across chemical spaces. This hybrid kernel significantly improved prediction performance across three kernel-based algorithms, with a support vector machine (SVM) achieving superior performance when using our kernel compared to standard kernels and outperforming all other machine learning architectures, including transformer-based deep neural networks. Using SVM-guided predictions, we successfully formulated the difficult-to-encapsulate venetoclax with optimized taurocholic acid ratios, yielding enhanced in vitro efficacy against Kasumi-1 leukemia cells. In a second case study, our AI-guided platform reduced excipient usage by 75% in a trametinib formulation while preserving the in vitro efficacy and in vivo pharmacokinetics relative to the standard formulation. Taken together, this study establishes a generalizable framework that combines robotic experimentation, kernel machine learning, and experimental validation to accelerate nanoparticle composition optimization for drug delivery.
- Open Access
- 11 June 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Identifying Antibiotic Effects of Investigational Drugs on Commensal Bacteria with Machine Learning
Many human-targeted medications have been found to impact the composition of patients’ gastrointestinal microbiomes, which has been proposed as an unrecognized source of drug side effects, comorbidities, and reduced treatment efficiencies. However, current methods to detect such drug effects on the microbiome largely rely on patient sample analysis or in vitro high-throughput screening - which are both laborious and resource intensive. To accelerate the systematic discovery of drug effects on the microbiome, we designed machine learning (ML) workflows on existing data from high-throughput experiments to predict which drugs can inhibit the growth of 40 commensal microbes in monocultures. We employed these models to predict potential effects of thousands of investigational drugs and found that up to 60% of these drugs are predicted to inhibit the growth of at least one commensal microbe. Using prospective in vitro validations, our workflows enabled us to uncover two non-antibiotic drugs, the recently approved chemotherapeutic Entrectinib and the clinical drug candidate PSI-697, to have previously unknown growth inhibition effects on multiple commensal microbes. Furthermore, we show that resistance to the effects of these drugs is mediated by BamB and TolC, implicating these drugs as potential agents to exacerbate antibiotic resistance. We further validated that Entrectinib significantly reduces microbial richness in a synthetic microbial model community. Taken together, our ML-assisted workflow and future extensions can accelerate the systematic discovery of microbiome-drug interactions, with implications for safer drug development and molecular microbiome engineering.
- 11 June 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessIdentifying Antibiotic Effects of Investigational Drugs on Commensal Bacteria with Machine Learning
Many human-targeted medications have been found to impact the composition of patients’ gastrointestinal microbiomes, which has been proposed as an unrecognized source of drug side effects, comorbidities, and reduced treatment efficiencies. However, current methods to detect such drug effects on the microbiome largely rely on patient sample analysis or in vitro high-throughput screening - which are both laborious and resource intensive. To accelerate the systematic discovery of drug effects on the microbiome, we designed machine learning (ML) workflows on existing data from high-throughput experiments to predict which drugs can inhibit the growth of 40 commensal microbes in monocultures. We employed these models to predict potential effects of thousands of investigational drugs and found that up to 60% of these drugs are predicted to inhibit the growth of at least one commensal microbe. Using prospective in vitro validations, our workflows enabled us to uncover two non-antibiotic drugs, the recently approved chemotherapeutic Entrectinib and the clinical drug candidate PSI-697, to have previously unknown growth inhibition effects on multiple commensal microbes. Furthermore, we show that resistance to the effects of these drugs is mediated by BamB and TolC, implicating these drugs as potential agents to exacerbate antibiotic resistance. We further validated that Entrectinib significantly reduces microbial richness in a synthetic microbial model community. Taken together, our ML-assisted workflow and future extensions can accelerate the systematic discovery of microbiome-drug interactions, with implications for safer drug development and molecular microbiome engineering.
- Open Access
- 18 March 2025
- Published by American Chemical Society (ACS) in ChemRxiv
TuNa-AI: a hybrid kernel machine to design tunable nanoparticles for drug delivery
Artificial intelligence (AI) stands to accelerate the development of nanoparticles for drug delivery, but current methodologies either focus on the identification of materials or adjusting of relative ratios of multi-component systems. Here, we developed a bespoke hybrid kernel machine integrating molecular learning and relative composition inference to engineer nanoparticles with new components and tunable composition. Our approach identified nanoformulations that encapsulate previously inaccessible drugs and can also guide excipient reduction.
- 18 March 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessTuNa-AI: a hybrid kernel machine to design tunable nanoparticles for drug delivery
Artificial intelligence (AI) stands to accelerate the development of nanoparticles for drug delivery, but current methodologies either focus on the identification of materials or adjusting of relative ratios of multi-component systems. Here, we developed a bespoke hybrid kernel machine integrating molecular learning and relative composition inference to engineer nanoparticles with new components and tunable composition. Our approach identified nanoformulations that encapsulate previously inaccessible drugs and can also guide excipient reduction.
- Top Cited Papers
- 17 October 2024
- Published by American Chemical Society (ACS) in Journal of the American Chemical Society
- Vol. 146 (43) , 29325-29334
Active Learning Guided Discovery of High Entropy Oxides Featuring High H2-production
High entropy oxides (HEOs) represent a class of solid solutions comprising multiple elements, offering significant scientific potential. Due to the enormous combination types of elements, the design of HEOs with desirable properties within high-dimensional composition spaces has traditionally relied heavily on knowledge and intuition. In this study, we present an active learning (AL) strategy tailored to efficiently explore the vast compositional space of HEOs. Our approach operates as a closed-loop system, iteratively cycling through "Training, Prediction, and Experiment" stages. Across multiple AL iterations, we have successfully identified four novel HEOs from a vast array of potential compositions. These newly discovered materials exhibit exceptional stability and demonstrate outstanding performance in H2 evolution rate (251 μmol gcat-1 min-1) during the water-gas shift reaction, surpassing benchmarks set by established catalysts such as Pt/γ-Al2O3 (135 μmol gcat-1 min-1) and Cu/ZnO/Al2O3 (81 μmol gcat-1 min-1). X-ray photoelectron spectroscopy and density functional theory calculations revealed a loss of elemental identity in the selected HEOs. This catalyst discovery process underscores the efficacy of Machine Learning in accelerating the identification of HEOs with unique characteristics by effectively leveraging insights from limited experimental data.
- 17 October 2024
- Published by American Chemical Society (ACS) in Journal of the American Chemical Society
- Vol. 146 (43) , 29325-29334
Top Cited PapersActive Learning Guided Discovery of High Entropy Oxides Featuring High H2-production
High entropy oxides (HEOs) represent a class of solid solutions comprising multiple elements, offering significant scientific potential. Due to the enormous combination types of elements, the design of HEOs with desirable properties within high-dimensional composition spaces has traditionally relied heavily on knowledge and intuition. In this study, we present an active learning (AL) strategy tailored to efficiently explore the vast compositional space of HEOs. Our approach operates as a closed-loop system, iteratively cycling through "Training, Prediction, and Experiment" stages. Across multiple AL iterations, we have successfully identified four novel HEOs from a vast array of potential compositions. These newly discovered materials exhibit exceptional stability and demonstrate outstanding performance in H2 evolution rate (251 μmol gcat-1 min-1) during the water-gas shift reaction, surpassing benchmarks set by established catalysts such as Pt/γ-Al2O3 (135 μmol gcat-1 min-1) and Cu/ZnO/Al2O3 (81 μmol gcat-1 min-1). X-ray photoelectron spectroscopy and density functional theory calculations revealed a loss of elemental identity in the selected HEOs. This catalyst discovery process underscores the efficacy of Machine Learning in accelerating the identification of HEOs with unique characteristics by effectively leveraging insights from limited experimental data.
- 19 October 2023
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 63 (21) , 6515-6524
Efficient Exploration of Chemical Compound Space Using Active Learning for Prediction of Thermodynamic Properties of Alkane Molecules
We introduce an exploratory active learning (AL) algorithm using Gaussian process regression and marginalized graph kernel (GPR-MGK) to sample chemical compound space (CCS) at minimal cost. Targeting 251,728 enumerated alkane molecules with 4-19 carbon atoms, we applied the AL algorithm to select a diverse and representative set of molecules and then conducted high-throughput molecular simulations on these selected molecules. To demonstrate the power of the AL algorithm, we built directed message-passing neural networks (D-MPNN) using simulation data as the training set to predict liquid densities, heat capacities, and vaporization enthalpies of the CCS. Validations show that D-MPNN models built on the smallest training set considered in this work, which consists of 313 molecules or 0.124% of the original CCS, predict the properties with R2 > 0.99 against the computational data and R2 > 0.94 against the experimental data. The advantage of the presented AL algorithm is that the predicted uncertainty of GPR depends on only the molecular structures, which renders it compatible with high-throughput data generation.
- 19 October 2023
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 63 (21) , 6515-6524
Efficient Exploration of Chemical Compound Space Using Active Learning for Prediction of Thermodynamic Properties of Alkane Molecules
We introduce an exploratory active learning (AL) algorithm using Gaussian process regression and marginalized graph kernel (GPR-MGK) to sample chemical compound space (CCS) at minimal cost. Targeting 251,728 enumerated alkane molecules with 4-19 carbon atoms, we applied the AL algorithm to select a diverse and representative set of molecules and then conducted high-throughput molecular simulations on these selected molecules. To demonstrate the power of the AL algorithm, we built directed message-passing neural networks (D-MPNN) using simulation data as the training set to predict liquid densities, heat capacities, and vaporization enthalpies of the CCS. Validations show that D-MPNN models built on the smallest training set considered in this work, which consists of 313 molecules or 0.124% of the original CCS, predict the properties with R2 > 0.99 against the computational data and R2 > 0.94 against the experimental data. The advantage of the presented AL algorithm is that the predicted uncertainty of GPR depends on only the molecular structures, which renders it compatible with high-throughput data generation.