Zachary Fralish
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Zachary Fralish has a total of 19 co-authors. Here are the top co-authors by number of shared publications:
- Daniel Reker
Duke University
9 shared publications - Chloe Markey
Duke University
3 shared publications - Hannah Lee
Duke University
3 shared publications Ashley Chen
Department of Computer Science, Duke University, Durham, USA
1 shared publication- Ashley Chen
Duke University
1 shared publication - Hong A. Chung
Duke
1 shared publication
Latest Publications
- Open Access
- 25 October 2026
- Published by American Chemical Society (ACS) in ChemRxiv
PCAN-PIE: A Machine Learning Tool for Prioritizing Prodrug Candidates by Predicted Bioactivation
Prodrug design enhances critical drug properties by creating drug derivatives with improved properties while harnessing metabolic activation to regenerate the active pharmaceutical ingredient (API) in vivo . However, reliably predicting whether a prodrug converts to the intended API remains challenging, causing project failures and limiting systematic exploration of prodrug design spaces. Here, we developed PCAN-PIE (Prodrug CAndidate Navigator using a Paired-Input Ensemble), a machine learning tool to de-risk prodrug design by distinguishing prodrug candidates likely to release their intended API from those unlikely to activate. PCAN-PIE demonstrates strong performance across different approved prodrug activation chemistries, can identify competing metabolic pathways that may compromise activation, and correctly triages chemically plausible but unreported transformation products. Attribution analysis suggests that PCAN-PIE bases its predictions on chemically meaningful reactive substructures, providing medicinal chemists with an interpretable rationale for candidate prioritization. In a prospective four-candidate prodrug design campaign, PCAN-PIE successfully prioritized the two candidates that ultimately demonstrated measurable API formation and cellular activity, while deprioritizing two candidates that showed negligible conversion and activity. Taken together, this work establishes PCAN-PIE as a focused computational tool for prioritizing prodrug candidates and provides a framework for integrating structure-based prediction into rational prodrug design to support the development of safer and more effective medicines.
- 25 October 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessPCAN-PIE: A Machine Learning Tool for Prioritizing Prodrug Candidates by Predicted Bioactivation
Prodrug design enhances critical drug properties by creating drug derivatives with improved properties while harnessing metabolic activation to regenerate the active pharmaceutical ingredient (API) in vivo . However, reliably predicting whether a prodrug converts to the intended API remains challenging, causing project failures and limiting systematic exploration of prodrug design spaces. Here, we developed PCAN-PIE (Prodrug CAndidate Navigator using a Paired-Input Ensemble), a machine learning tool to de-risk prodrug design by distinguishing prodrug candidates likely to release their intended API from those unlikely to activate. PCAN-PIE demonstrates strong performance across different approved prodrug activation chemistries, can identify competing metabolic pathways that may compromise activation, and correctly triages chemically plausible but unreported transformation products. Attribution analysis suggests that PCAN-PIE bases its predictions on chemically meaningful reactive substructures, providing medicinal chemists with an interpretable rationale for candidate prioritization. In a prospective four-candidate prodrug design campaign, PCAN-PIE successfully prioritized the two candidates that ultimately demonstrated measurable API formation and cellular activity, while deprioritizing two candidates that showed negligible conversion and activity. Taken together, this work establishes PCAN-PIE as a focused computational tool for prioritizing prodrug candidates and provides a framework for integrating structure-based prediction into rational prodrug design to support the development of safer and more effective medicines.
- 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.
- Open Access
- 3 April 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Ensemble Siamese Neural Networks for Prodrug Activation Prediction
Prodrug design enhances critical drug properties by harnessing metabolic activation to release the active pharmaceutical ingredient. However, predicting prodrug activation remains challenging, causing design failures and limiting systematic exploration of prodrug design spaces. Here, we developed an ensemble Siamese neural network model to predict prodrug activation. We demonstrate that our model accurately captures prodrug activations from molecular structure, including in a prospective, in-house prodrug development effort. The model reliably identifies competing metabolic pathways and distinguishes true activations from metabolic counterfactuals. Attribution analysis reveals that predictions are driven by chemically meaningful substructures associated with site-specific reactivity. Taken together, this work establishes a generalizable framework for prodrug activation prediction and suggests that key determinants of metabolic reactivity can be learned directly from molecular structure, which could enable more rational prodrug design to accelerate the development of safer and more effective medicines.
- 3 April 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessEnsemble Siamese Neural Networks for Prodrug Activation Prediction
Prodrug design enhances critical drug properties by harnessing metabolic activation to release the active pharmaceutical ingredient. However, predicting prodrug activation remains challenging, causing design failures and limiting systematic exploration of prodrug design spaces. Here, we developed an ensemble Siamese neural network model to predict prodrug activation. We demonstrate that our model accurately captures prodrug activations from molecular structure, including in a prospective, in-house prodrug development effort. The model reliably identifies competing metabolic pathways and distinguishes true activations from metabolic counterfactuals. Attribution analysis reveals that predictions are driven by chemically meaningful substructures associated with site-specific reactivity. Taken together, this work establishes a generalizable framework for prodrug activation prediction and suggests that key determinants of metabolic reactivity can be learned directly from molecular structure, which could enable more rational prodrug design to accelerate the development of safer and more effective medicines.
- Open Access
- 27 November 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Ensemble Siamese Neural Networks for Prodrug Activation Prediction
Prodrug design enhances critical drug properties without affecting potency by harnessing metabolic activation to release the active pharmaceutical ingredient (API). However, predicting prodrug activation remains challenging, causing design failures and limiting systematic exploration of the prodrug design space. Here, we developed an ensemble Siamese neural network model to predict prodrug activation. We demonstrate that our model accurately predicts 95% of all approved prodrugs, compared to less than 75% for any of the currently established metabolism models. Prospective validation on in-house designed prodrugs further confirmed the model’s ability to rank prodrug candidates by their observed release profiles. This work establishes a generalizable framework for prodrug activation prediction and provides a foundation for rational prodrug design, with the potential to accelerate the development of safer and more effective medicines.
- 27 November 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessEnsemble Siamese Neural Networks for Prodrug Activation Prediction
Prodrug design enhances critical drug properties without affecting potency by harnessing metabolic activation to release the active pharmaceutical ingredient (API). However, predicting prodrug activation remains challenging, causing design failures and limiting systematic exploration of the prodrug design space. Here, we developed an ensemble Siamese neural network model to predict prodrug activation. We demonstrate that our model accurately predicts 95% of all approved prodrugs, compared to less than 75% for any of the currently established metabolism models. Prospective validation on in-house designed prodrugs further confirmed the model’s ability to rank prodrug candidates by their observed release profiles. This work establishes a generalizable framework for prodrug activation prediction and provides a foundation for rational prodrug design, with the potential to accelerate the development of safer and more effective medicines.
- Open Access
- 4 August 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Profiling Biological Effects of Microbiome Metabolites via Machine Learning
Human microbiome-derived metabolites are key mediators of host physiology. However, their biological effects remain largely uncharacterized due to limitations of current low-throughput and untargeted experimental approaches that are time-intensive and costly. This has hindered the systematic biological characterization of microbiome metabolites. To address this gap and accelerate the identification of biological effects of microbiome metabolites, we developed a machine learning platform trained on publicly available drug development data to rapidly predict a wide array of chemical and biological properties of microbiome metabolites. Prospective experimental validation confirmed the predictive accuracy of our models and uncovered previously unknown effects of several metabolites. For example, we identified previously unknown Interleukin-8 secretion stimulation by the metabolites spermine and spermidine, which have been regarded anti-inflammatory thus far. Our findings demonstrate the power of machine learning to accelerate the functional annotations of microbiome-derived metabolites, paving the way for new biomarker discovery and therapeutic development.
- 4 August 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessProfiling Biological Effects of Microbiome Metabolites via Machine Learning
Human microbiome-derived metabolites are key mediators of host physiology. However, their biological effects remain largely uncharacterized due to limitations of current low-throughput and untargeted experimental approaches that are time-intensive and costly. This has hindered the systematic biological characterization of microbiome metabolites. To address this gap and accelerate the identification of biological effects of microbiome metabolites, we developed a machine learning platform trained on publicly available drug development data to rapidly predict a wide array of chemical and biological properties of microbiome metabolites. Prospective experimental validation confirmed the predictive accuracy of our models and uncovered previously unknown effects of several metabolites. For example, we identified previously unknown Interleukin-8 secretion stimulation by the metabolites spermine and spermidine, which have been regarded anti-inflammatory thus far. Our findings demonstrate the power of machine learning to accelerate the functional annotations of microbiome-derived metabolites, paving the way for new biomarker discovery and therapeutic development.
- Open Access
- 27 August 2024
- Published by Beilstein Institut in Beilstein Journal of Organic Chemistry
- Vol. 20 (1) , 2152-2162
Finding the most potent compounds using active learning on molecular pairs
Active learning allows algorithms to steer iterative experimentation to accelerate and de-risk molecular optimizations, but actively trained models might still exhibit poor performance during early project stages where the training data is limited and model exploitation might lead to analog identification with limited scaffold diversity. Here, we present ActiveDelta, an adaptive approach that leverages paired molecular representations to predict improvements from the current best training compound to prioritize further data acquisition. We apply the ActiveDelta concept to both graph-based deep (Chemprop) and tree-based (XGBoost) models during exploitative active learning for 99 Ki benchmarking datasets. We show that both ActiveDelta implementations excel at identifying more potent inhibitors compared to the standard exploitative active learning implementations of Chemprop, XGBoost, and Random Forest. The ActiveDelta approach is also able to identify more chemically diverse inhibitors in terms of their Murcko scaffolds. Finally, deep models such as Chemprop trained on data selected through ActiveDelta approaches can more accurately identify inhibitors in test data created through simulated time-splits. Overall, this study highlights the large potential for molecular pairing approaches to further improve popular active learning strategies in low data regimes by enabling faster and more accurate identification of more diverse molecular hits against critical drug targets.
- 27 August 2024
- Published by Beilstein Institut in Beilstein Journal of Organic Chemistry
- Vol. 20 (1) , 2152-2162
Open AccessFinding the most potent compounds using active learning on molecular pairs
Active learning allows algorithms to steer iterative experimentation to accelerate and de-risk molecular optimizations, but actively trained models might still exhibit poor performance during early project stages where the training data is limited and model exploitation might lead to analog identification with limited scaffold diversity. Here, we present ActiveDelta, an adaptive approach that leverages paired molecular representations to predict improvements from the current best training compound to prioritize further data acquisition. We apply the ActiveDelta concept to both graph-based deep (Chemprop) and tree-based (XGBoost) models during exploitative active learning for 99 Ki benchmarking datasets. We show that both ActiveDelta implementations excel at identifying more potent inhibitors compared to the standard exploitative active learning implementations of Chemprop, XGBoost, and Random Forest. The ActiveDelta approach is also able to identify more chemically diverse inhibitors in terms of their Murcko scaffolds. Finally, deep models such as Chemprop trained on data selected through ActiveDelta approaches can more accurately identify inhibitors in test data created through simulated time-splits. Overall, this study highlights the large potential for molecular pairing approaches to further improve popular active learning strategies in low data regimes by enabling faster and more accurate identification of more diverse molecular hits against critical drug targets.
- Open Access
- 31 May 2024
- Published by Royal Society of Chemistry (RSC) in RSC Medicinal Chemistry
- Vol. 15 (7) , 2474-2482
Leveraging bounded datapoints to classify molecular potency improvements
Molecular machine learning algorithms are becoming increasingly powerful at predicting the potency of potential drug candidates to guide molecular discovery, lead series prioritization, and structural optimization. However, a substantial amount of inhibition data is bounded and inaccessible to traditional regression algorithms. Here, we develop a novel molecular pairing approach to process this data. This creates a new classification task of predicting which one of two paired molecules is more potent. This novel classification task can be accurately solved by various, established molecular machine learning algorithms, including XGBoost and Chemprop. Across 230 ChEMBL IC50 datasets, both tree-based and neural network-based “DeltaClassifiers” show improvements over traditional regression approaches in correctly classifying molecular potency improvements. The Chemprop-based deep DeltaClassifier outperformed all here evaluated regression approaches for paired molecules with shared and with distinct scaffolds, highlighting the promise of this approach for molecular optimization and scaffold-hopping.
- 31 May 2024
- Published by Royal Society of Chemistry (RSC) in RSC Medicinal Chemistry
- Vol. 15 (7) , 2474-2482
Open AccessLeveraging bounded datapoints to classify molecular potency improvements
Molecular machine learning algorithms are becoming increasingly powerful at predicting the potency of potential drug candidates to guide molecular discovery, lead series prioritization, and structural optimization. However, a substantial amount of inhibition data is bounded and inaccessible to traditional regression algorithms. Here, we develop a novel molecular pairing approach to process this data. This creates a new classification task of predicting which one of two paired molecules is more potent. This novel classification task can be accurately solved by various, established molecular machine learning algorithms, including XGBoost and Chemprop. Across 230 ChEMBL IC50 datasets, both tree-based and neural network-based “DeltaClassifiers” show improvements over traditional regression approaches in correctly classifying molecular potency improvements. The Chemprop-based deep DeltaClassifier outperformed all here evaluated regression approaches for paired molecules with shared and with distinct scaffolds, highlighting the promise of this approach for molecular optimization and scaffold-hopping.
- Top Cited Papers
- 2 April 2024
- Published by Springer Nature in Nature Reviews Drug Discovery
- Vol. 23 (5) , 365-380
The landscape of small-molecule prodrugs
Prodrugs are derivatives with superior properties compared with the parent active pharmaceutical ingredient (API), which undergo biotransformation after administration to generate the API in situ. Although sharing this general characteristic, prodrugs encompass a wide range of different chemical structures, therapeutic indications and properties. Here we provide the first holistic analysis of the current landscape of approved prodrugs using cheminformatics and data science approaches to reveal trends in prodrug development. We highlight rationales that underlie prodrug design, their indications, mechanisms of API release, the chemistry of promoieties added to APIs to form prodrugs and the market impact of prodrugs. On the basis of this analysis, we discuss strengths and limitations of current prodrug approaches and suggest areas for future development.
- 2 April 2024
- Published by Springer Nature in Nature Reviews Drug Discovery
- Vol. 23 (5) , 365-380
Top Cited PapersThe landscape of small-molecule prodrugs
Prodrugs are derivatives with superior properties compared with the parent active pharmaceutical ingredient (API), which undergo biotransformation after administration to generate the API in situ. Although sharing this general characteristic, prodrugs encompass a wide range of different chemical structures, therapeutic indications and properties. Here we provide the first holistic analysis of the current landscape of approved prodrugs using cheminformatics and data science approaches to reveal trends in prodrug development. We highlight rationales that underlie prodrug design, their indications, mechanisms of API release, the chemistry of promoieties added to APIs to form prodrugs and the market impact of prodrugs. On the basis of this analysis, we discuss strengths and limitations of current prodrug approaches and suggest areas for future development.
- Open Access
- 10 April 2023
- Published by American Chemical Society (ACS) in ChemRxiv
DeepDelta: Predicting Pharmacokinetic Improvements of Molecular Derivatives with Deep Learning
Established molecular machine learning models process individual molecules as inputs to predict their biological, chemical, or physical properties. However, such algorithms require large datasets and have not been optimized to predict property differences between molecules. Many drug and material development tasks would benefit from an algorithm that can directly compare two molecules to guide molecular optimization and prioritization. Here, we develop DeepDelta, a pairwise deep learning approach that processes two molecules simultaneously and learns to predict property differences between two molecules from small datasets. On 10 pharmacokinetic benchmark tasks, our DeepDelta approach outperforms two established molecular machine learning algorithms, the message passing neural network (MPNN) ChemProp and Random Forest using radial fingerprints. We further analyze our performance and find that DeepDelta is particularly outperforming established approaches at predicting large differences in molecular properties and can perform scaffold hopping. Furthermore, we derive simple computational tests of our models based on mathematical invariants and show that compliance to these tests correlate with overall model performance – providing an innovative, unsupervised, and easily computable measure of expected model performance and applicability. Taken together, DeepDelta provides an accurate approach to predict molecular property differences and will allow for higher fidelity and transparency in molecular optimization for drug development and the chemical sciences.
- 10 April 2023
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessDeepDelta: Predicting Pharmacokinetic Improvements of Molecular Derivatives with Deep Learning
Established molecular machine learning models process individual molecules as inputs to predict their biological, chemical, or physical properties. However, such algorithms require large datasets and have not been optimized to predict property differences between molecules. Many drug and material development tasks would benefit from an algorithm that can directly compare two molecules to guide molecular optimization and prioritization. Here, we develop DeepDelta, a pairwise deep learning approach that processes two molecules simultaneously and learns to predict property differences between two molecules from small datasets. On 10 pharmacokinetic benchmark tasks, our DeepDelta approach outperforms two established molecular machine learning algorithms, the message passing neural network (MPNN) ChemProp and Random Forest using radial fingerprints. We further analyze our performance and find that DeepDelta is particularly outperforming established approaches at predicting large differences in molecular properties and can perform scaffold hopping. Furthermore, we derive simple computational tests of our models based on mathematical invariants and show that compliance to these tests correlate with overall model performance – providing an innovative, unsupervised, and easily computable measure of expected model performance and applicability. Taken together, DeepDelta provides an accurate approach to predict molecular property differences and will allow for higher fidelity and transparency in molecular optimization for drug development and the chemical sciences.
- Open Access
- 7 May 2021
- Published by Modestum Ltd in International Electronic Journal of Mathematics Education
- Vol. 16 (2) , em0631
Using Differential Equations to Model Phoretic Parasitism as Part of SCUDEM Challenge
The SIMIODE Challenge for Undergraduates in Differential Equation Modelling (SCUDEM) offers students the opportunity to improve their mathematical capacity, ability to think critically, and communication skills through researching, developing, and presenting on a differential equations model for a natural phenomenon. During the fall 2019 SCUDEM, we chose to model population dynamics of phoretic parasitism through chemical espionage of anti-aphrodisiacs by parasitic wasps. Additionally, we modelled the dispersion of anti-aphrodisiac pheromones from impregnated butterflies and the propagation of resultant action potentials in wasps due to stimulation of olfactory receptor neurons from these pheromones. This article displays a summary of our undergraduate student team’s response to a provided prompt in SCUDEM IV. This article additionally shares personal testimonies from three students about the modelling challenge to highlight the benefits of participating in differential equation modelling challenges.
- 7 May 2021
- Published by Modestum Ltd in International Electronic Journal of Mathematics Education
- Vol. 16 (2) , em0631
Open AccessUsing Differential Equations to Model Phoretic Parasitism as Part of SCUDEM Challenge
The SIMIODE Challenge for Undergraduates in Differential Equation Modelling (SCUDEM) offers students the opportunity to improve their mathematical capacity, ability to think critically, and communication skills through researching, developing, and presenting on a differential equations model for a natural phenomenon. During the fall 2019 SCUDEM, we chose to model population dynamics of phoretic parasitism through chemical espionage of anti-aphrodisiacs by parasitic wasps. Additionally, we modelled the dispersion of anti-aphrodisiac pheromones from impregnated butterflies and the propagation of resultant action potentials in wasps due to stimulation of olfactory receptor neurons from these pheromones. This article displays a summary of our undergraduate student team’s response to a provided prompt in SCUDEM IV. This article additionally shares personal testimonies from three students about the modelling challenge to highlight the benefits of participating in differential equation modelling challenges.