Thomas M. Whitehead
Research Subjects
Thomas M. Whitehead has a total of 74 co-authors. Here are the top co-authors by number of shared publications:
- Benedict Irwin
Optibrium Ltd
4 shared publications G. J. Conduit
Cavendish Laboratory, J.J. Thomson Avenue, Cambridge CB3 0HE, United Kingdom
4 shared publications- Matthew D. Segall
Optibrium Limited
3 shared publications Gareth J. Conduit
Intellegens Ltd., The Studio, Chesterton Mill, Cambridge CB4 3NP, United Kingdom
2 shared publications- Gareth J. Conduit
Theory of Condensed Matter Group, Cavendish Laboratory, University of Cambridge, Cambridge CB3 0HE, UK
2 shared publications - Adele M. Lehane
Research School of Biology Australian National University Canberra 2601 Australia
1 shared publication
Latest Publications
- Open Access
- 11 December 2025
- Published by American Chemical Society (ACS) in ChemRxiv
The Potential of Machine Learning in Oligonucleotide Therapeutics Manufacturing
Approvals of synthetic oligonucleotide therapeutics are sharply rising and with it the challenges to achieve sustainable manufacturing strategies. These chain-like molecules are commonly assembled via solid phase oligonucleotide synthesis (SPOS), a method that allows for quick and efficient sequential additions of monomers up until the desired molecule length. A primary challenge in SPOS is the meticulous control of the multitude of variables at each additional step to avoid formation of complex and unwanted impurities. The analysis, identification, and quantification of these impurities can also be troublesome and time consuming. Optimisation of SPOS and related analytical methods is essential to guarantee reproducibility and robustness during manufacture. Furthermore, SPOS is heavily reliant on solvent washes that drastically increase the process mass intensity (PMI) and environmental impact of the synthesis which should be addressed during optimisation. In this perspective, we discuss the potential of machine learning (ML) models in oligonucleotide development and manufacturing to identify sequence-specific optimal process parameters, downstream processes, and impurity analysis; and recognise complex patterns that are often missed by existing development methods. The expected benefit of adopting ML is improved product yield, improved productivity and reduced time and cost through reduction in the number of experiments required for process development.
- 11 December 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessThe Potential of Machine Learning in Oligonucleotide Therapeutics Manufacturing
Approvals of synthetic oligonucleotide therapeutics are sharply rising and with it the challenges to achieve sustainable manufacturing strategies. These chain-like molecules are commonly assembled via solid phase oligonucleotide synthesis (SPOS), a method that allows for quick and efficient sequential additions of monomers up until the desired molecule length. A primary challenge in SPOS is the meticulous control of the multitude of variables at each additional step to avoid formation of complex and unwanted impurities. The analysis, identification, and quantification of these impurities can also be troublesome and time consuming. Optimisation of SPOS and related analytical methods is essential to guarantee reproducibility and robustness during manufacture. Furthermore, SPOS is heavily reliant on solvent washes that drastically increase the process mass intensity (PMI) and environmental impact of the synthesis which should be addressed during optimisation. In this perspective, we discuss the potential of machine learning (ML) models in oligonucleotide development and manufacturing to identify sequence-specific optimal process parameters, downstream processes, and impurity analysis; and recognise complex patterns that are often missed by existing development methods. The expected benefit of adopting ML is improved product yield, improved productivity and reduced time and cost through reduction in the number of experiments required for process development.
- Open Access
- 13 March 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Comparing Massively-Multitask Regression Algorithms for Drug Discovery
Massively-Multitask Regression Models (MMRMs) trained on millions of compounds and many thousands of assays can predict bioactivity with accuracy comparable to 4-concentration IC50 experiments. Recent advances in hardware and algorithms have produced a variety of methods for multitask modeling. This report compares the performance of six MMRM algorithms: Profile-QSAR (pQSAR), Alchemite, a meta learner (MetaNN), a multitask feed-forward neural network (MT-DNN), Bayesian factorization with side information (Macau) and Inductive Matrix Completion (IMC). To ensure a fair comparison, each was trained by an expert, in several cases an author, of each method. All used the same sets of 159 kinase and 4276 diverse ChEMBL assays, employing the same realistically novel training/test set splits. MMRMs generally performed much better than a benchmark of single-task random forest regression models for our use case of virtually screening the compound collection on which the models were trained. The comparison was complicated, because methods that train all models simultaneously must leave out the test-set measurements for all assays to avoid test-set leakage, here 75% of measurements. MMRMs which train models one-at-a-time need only leave out data for each assay as it is trained, training on 99+% of data. This does not affect the accuracy of the final production models trained on 100% of data but does affect evaluation of how the final models will perform. The comparisons, therefore, included 3 training/test set collections: “all-out” models that leave out all test sets during training; “one-out” models where practical; and “subset-out” models, which only built models for about 10% of kinase assays or 1% of diverse assays, but could thus train evaluation models on about 90% or 99% of the measurements respectively. Many methods achieve similar accuracy. However, models trained on only 75% of the data performed much worse than those trained on 99+%. This indicates that all-out models seriously underestimate the performance of the final production models. Subset-out models were closer to one-out. A compromise method is to assess performance of the final models by multiple subset-out models, a more practicable computation for 1000s of assays. MMRMs demonstrated little advantage over single-task models for “cold-start” predictions on our novel test-set compounds not only unlike the specific assay’s training set, but also never tested on any of the other multitask supporting assays. Instead, the accuracy advantage was mainly from imputations within these sparse assay collections, compounds unlike the training set for the assay of interest, but with some measurements on other assays. This implies that MMRMs are best suited for hit-finding, off-target, promiscuity, mechanism-of-action, polypharmacology or drug repurposing prediction of compounds from the source used to train the overall multitask model. They have little advantage over single-task models, at much higher cost, for virtual screening of vendor archives or exploratory generative chemistry. Given that accuracy of the final models is often comparable between several of the algorithms, the paper concludes with a detailed discussion of other practical pros and cons of each method that might help choose which method to employ.
- 13 March 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessComparing Massively-Multitask Regression Algorithms for Drug Discovery
Massively-Multitask Regression Models (MMRMs) trained on millions of compounds and many thousands of assays can predict bioactivity with accuracy comparable to 4-concentration IC50 experiments. Recent advances in hardware and algorithms have produced a variety of methods for multitask modeling. This report compares the performance of six MMRM algorithms: Profile-QSAR (pQSAR), Alchemite, a meta learner (MetaNN), a multitask feed-forward neural network (MT-DNN), Bayesian factorization with side information (Macau) and Inductive Matrix Completion (IMC). To ensure a fair comparison, each was trained by an expert, in several cases an author, of each method. All used the same sets of 159 kinase and 4276 diverse ChEMBL assays, employing the same realistically novel training/test set splits. MMRMs generally performed much better than a benchmark of single-task random forest regression models for our use case of virtually screening the compound collection on which the models were trained. The comparison was complicated, because methods that train all models simultaneously must leave out the test-set measurements for all assays to avoid test-set leakage, here 75% of measurements. MMRMs which train models one-at-a-time need only leave out data for each assay as it is trained, training on 99+% of data. This does not affect the accuracy of the final production models trained on 100% of data but does affect evaluation of how the final models will perform. The comparisons, therefore, included 3 training/test set collections: “all-out” models that leave out all test sets during training; “one-out” models where practical; and “subset-out” models, which only built models for about 10% of kinase assays or 1% of diverse assays, but could thus train evaluation models on about 90% or 99% of the measurements respectively. Many methods achieve similar accuracy. However, models trained on only 75% of the data performed much worse than those trained on 99+%. This indicates that all-out models seriously underestimate the performance of the final production models. Subset-out models were closer to one-out. A compromise method is to assess performance of the final models by multiple subset-out models, a more practicable computation for 1000s of assays. MMRMs demonstrated little advantage over single-task models for “cold-start” predictions on our novel test-set compounds not only unlike the specific assay’s training set, but also never tested on any of the other multitask supporting assays. Instead, the accuracy advantage was mainly from imputations within these sparse assay collections, compounds unlike the training set for the assay of interest, but with some measurements on other assays. This implies that MMRMs are best suited for hit-finding, off-target, promiscuity, mechanism-of-action, polypharmacology or drug repurposing prediction of compounds from the source used to train the overall multitask model. They have little advantage over single-task models, at much higher cost, for virtual screening of vendor archives or exploratory generative chemistry. Given that accuracy of the final models is often comparable between several of the algorithms, the paper concludes with a detailed discussion of other practical pros and cons of each method that might help choose which method to employ.
- 21 June 2024
- Published by Taylor & Francis
- p. 236-265
Adaptive Experimental Design
Product formulation is a complex and challenging process involving the careful balancing of multiple factors to achieve the desired product characteristics. This chapter explores the data-driven Adaptive Experimental Design (AED) approach to product formulation, illustrating how agility and responsiveness to new information and experimental results lead to a more efficient and effective process. We break the AED paradigm into its component parts, focusing in turn on acquisition functions, optimizers, and the use of surrogate models, before reuniting them to discuss their synergies and how they support an improved product formulation process. This chapter follows a focused case study from Domino Printing Sciences, where AED’s utility in developing cutting-edge ink formulations is clearly demonstrated. This example serves as a blueprint for leveraging AED in a variety of sectors, underscoring its value in dynamic data-driven environments.
- 21 June 2024
- Published by Taylor & Francis
- p. 236-265
Adaptive Experimental Design
Product formulation is a complex and challenging process involving the careful balancing of multiple factors to achieve the desired product characteristics. This chapter explores the data-driven Adaptive Experimental Design (AED) approach to product formulation, illustrating how agility and responsiveness to new information and experimental results lead to a more efficient and effective process. We break the AED paradigm into its component parts, focusing in turn on acquisition functions, optimizers, and the use of surrogate models, before reuniting them to discuss their synergies and how they support an improved product formulation process. This chapter follows a focused case study from Domino Printing Sciences, where AED’s utility in developing cutting-edge ink formulations is clearly demonstrated. This example serves as a blueprint for leveraging AED in a variety of sectors, underscoring its value in dynamic data-driven environments.
- 13 December 2023
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 64 (7) , 2624-2636
Quantifying the Benefits of Imputation over QSAR Methods in Toxicology Data Modeling
Imputation machine learning (ML) surpasses traditional approaches in modeling toxicity data. The method was tested on an open-source data set comprising approximately 2500 ingredients with limited in vitro and in vivo data obtained from the OECD QSAR Toolbox. By leveraging the relationships between different toxicological end points, imputation extracts more valuable information from each data point compared to well-established single end point methods, such as ML-based Quantitative Structure Activity Relationship (QSAR) approaches, providing a final improvement of up to around 0.2 in the coefficient of determination. A significant aspect of this methodology is its resilience to the inclusion of extraneous chemical or experimental data. While additional data typically introduces a considerable level of noise and can hinder performance of single end point QSAR modeling, imputation models remain unaffected. This implies a reduction in the need for laborious manual preprocessing tasks such as feature selection, thereby making data preparation for ML analysis more efficient. This successful test, conducted on open-source data, validates the efficacy of imputation approaches in toxicity data analysis. This work opens the way for applying similar methods to other types of sparse toxicological data matrices, and so we discuss the development of regulatory authority guidelines to accept imputation models, a key aspect for the wider adoption of these methods.
- 13 December 2023
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 64 (7) , 2624-2636
Quantifying the Benefits of Imputation over QSAR Methods in Toxicology Data Modeling
Imputation machine learning (ML) surpasses traditional approaches in modeling toxicity data. The method was tested on an open-source data set comprising approximately 2500 ingredients with limited in vitro and in vivo data obtained from the OECD QSAR Toolbox. By leveraging the relationships between different toxicological end points, imputation extracts more valuable information from each data point compared to well-established single end point methods, such as ML-based Quantitative Structure Activity Relationship (QSAR) approaches, providing a final improvement of up to around 0.2 in the coefficient of determination. A significant aspect of this methodology is its resilience to the inclusion of extraneous chemical or experimental data. While additional data typically introduces a considerable level of noise and can hinder performance of single end point QSAR modeling, imputation models remain unaffected. This implies a reduction in the need for laborious manual preprocessing tasks such as feature selection, thereby making data preparation for ML analysis more efficient. This successful test, conducted on open-source data, validates the efficacy of imputation approaches in toxicity data analysis. This work opens the way for applying similar methods to other types of sparse toxicological data matrices, and so we discuss the development of regulatory authority guidelines to accept imputation models, a key aspect for the wider adoption of these methods.
- Open Access
- 1 December 2021
- Published by Elsevier in Cell Reports Physical Science
- Vol. 2 (12)
Formulation and manufacturing optimization of lithium-ion graphite-based electrodes via machine learning
- 1 December 2021
- Published by Elsevier in Cell Reports Physical Science
- Vol. 2 (12)
Open AccessFormulation and manufacturing optimization of lithium-ion graphite-based electrodes via machine learning
- 26 March 2021
- Published by World Scientific Pub Co Pte Ltd in International Journal on Artificial Intelligence Tools
- Vol. 30 (02)
Hyperparameter-free Regularization by Sampling from an Infinite Space of Neural Networks
Stochastic activation functions, where the output of an activation function is a hyper-parameter-free random function of the inputs, generalize the concept of dropout to sampling from an infinitely large space of related networks. Stochastic activation functions provide intrinsic regularization and sparsification of artificial neural networks, along with cheap and accurate estimates of the uncertainty in the predictions from a network. Examples are presented against standard benchmarking datasets.
- 26 March 2021
- Published by World Scientific Pub Co Pte Ltd in International Journal on Artificial Intelligence Tools
- Vol. 30 (02)
Hyperparameter-free Regularization by Sampling from an Infinite Space of Neural Networks
Stochastic activation functions, where the output of an activation function is a hyper-parameter-free random function of the inputs, generalize the concept of dropout to sampling from an infinitely large space of related networks. Stochastic activation functions provide intrinsic regularization and sparsification of artificial neural networks, along with cheap and accurate estimates of the uncertainty in the predictions from a network. Examples are presented against standard benchmarking datasets.
- Open Access
- 6 November 2020
- Published by American Chemical Society (ACS) in ChemRxiv
An Open Drug Discovery Competition: Experimental Validation of Predictive Models in a Series of Novel Antimalarials
The discovery of new antimalarial medicines with novel mechanisms of action is key to combating the problem of increasing resistance to our frontline treatments. The Open Source Malaria (OSM) consortium has been developing compounds ("Series 4") that have potent activity against Plasmodium falciparum in vitro and in vivo and that have been suggested to act through the inhibition of PfATP4, an essential membrane ion pump that regulates the parasite’s intracellular Na+ concentration. The structure of PfATP4 is yet to be determined. In the absence of structural information about this target, a public competition was created to develop a model that would allow the prediction of anti-PfATP4 activity among Series 4 compounds, thereby reducing project costs associated with the unnecessary synthesis of inactive compounds.In the first round, in 2016, six participants used the open data collated by OSM to develop moderately predictive models using diverse methods. Notably, all submitted models were available to all other participants in real time. Since then further bioactivity data have been acquired and machine learning methods have rapidly developed, so a second round of the competition was undertaken, in 2019, again with freely-donated models that other participants could see. The best-performing models from this second round were used to predict novel inhibitory molecules, of which several were synthesised and evaluated against the parasite. One such compound, containing a motif that the human chemists familiar with this series would have dismissed as ill-advised, was active. The project demonstrated the abilities of new machine learning methods in the prediction of active compounds where there is no biological target structure, frequently the central problem in phenotypic drug discovery. Since all data and participant interactions remain in the public domain, this research project “lives” and may be improved by others.
- 6 November 2020
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessAn Open Drug Discovery Competition: Experimental Validation of Predictive Models in a Series of Novel Antimalarials
The discovery of new antimalarial medicines with novel mechanisms of action is key to combating the problem of increasing resistance to our frontline treatments. The Open Source Malaria (OSM) consortium has been developing compounds ("Series 4") that have potent activity against Plasmodium falciparum in vitro and in vivo and that have been suggested to act through the inhibition of PfATP4, an essential membrane ion pump that regulates the parasite’s intracellular Na+ concentration. The structure of PfATP4 is yet to be determined. In the absence of structural information about this target, a public competition was created to develop a model that would allow the prediction of anti-PfATP4 activity among Series 4 compounds, thereby reducing project costs associated with the unnecessary synthesis of inactive compounds.In the first round, in 2016, six participants used the open data collated by OSM to develop moderately predictive models using diverse methods. Notably, all submitted models were available to all other participants in real time. Since then further bioactivity data have been acquired and machine learning methods have rapidly developed, so a second round of the competition was undertaken, in 2019, again with freely-donated models that other participants could see. The best-performing models from this second round were used to predict novel inhibitory molecules, of which several were synthesised and evaluated against the parasite. One such compound, containing a motif that the human chemists familiar with this series would have dismissed as ill-advised, was active. The project demonstrated the abilities of new machine learning methods in the prediction of active compounds where there is no biological target structure, frequently the central problem in phenotypic drug discovery. Since all data and participant interactions remain in the public domain, this research project “lives” and may be improved by others.
- 1 June 2020
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 60 (6) , 2848-2857
Practical Applications of Deep Learning To Impute Heterogeneous Drug Discovery Data
Contemporary deep learning approaches still struggle to bring a useful improvement in the field of drug discovery due to the challenges of sparse, noisy and heterogeneous data that are typically encountered in this context. We use a state-of-the-art deep learning method, Alchemite™, to impute data from drug discovery projects, including multi-target biochemical activities, phenotypic activities in cell-based assays, and a variety of absorption, distribution, metabolism, and excretion (ADME) endpoints. The resulting model gives excellent predictions for activity and ADME endpoints, offering an average increase in R2 of 0.22 versus quantitative structure-activity relationship methods. The model accuracy is robust to combining data across uncorrelated endpoints and projects with different chemical spaces, enabling a single model to be trained for all compounds and endpoints. We demonstrate improvements in accuracy on the latest chemistry and data when updating models with new data as an ongoing medicinal chemistry project progresses.
- 1 June 2020
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 60 (6) , 2848-2857
Practical Applications of Deep Learning To Impute Heterogeneous Drug Discovery Data
Contemporary deep learning approaches still struggle to bring a useful improvement in the field of drug discovery due to the challenges of sparse, noisy and heterogeneous data that are typically encountered in this context. We use a state-of-the-art deep learning method, Alchemite™, to impute data from drug discovery projects, including multi-target biochemical activities, phenotypic activities in cell-based assays, and a variety of absorption, distribution, metabolism, and excretion (ADME) endpoints. The resulting model gives excellent predictions for activity and ADME endpoints, offering an average increase in R2 of 0.22 versus quantitative structure-activity relationship methods. The model accuracy is robust to combining data across uncorrelated endpoints and projects with different chemical spaces, enabling a single model to be trained for all compounds and endpoints. We demonstrate improvements in accuracy on the latest chemistry and data when updating models with new data as an ongoing medicinal chemistry project progresses.
- Open Access
- 6 April 2020
- Published by Taylor & Francis in Future Drug Discovery
- Vol. 2 (2)
Imputation Versus Prediction: Applications in Machine Learning for Drug Discovery
Imputation is a powerful statistical method that is distinct from the predictive modelling techniques more commonly used in drug discovery. Imputation uses sparse experimental data in an incomplete... Lay abstract Imputation is the process of filling in the gaps in a dataset, where values have not yet been measured, using the limited data that are already present. This may appear simple when onl...
- 6 April 2020
- Published by Taylor & Francis in Future Drug Discovery
- Vol. 2 (2)
Open AccessImputation Versus Prediction: Applications in Machine Learning for Drug Discovery
Imputation is a powerful statistical method that is distinct from the predictive modelling techniques more commonly used in drug discovery. Imputation uses sparse experimental data in an incomplete... Lay abstract Imputation is the process of filling in the gaps in a dataset, where values have not yet been measured, using the limited data that are already present. This may appear simple when onl...
- 12 February 2019
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 59 (3) , 1197-1204
Imputation of Assay Bioactivity Data Using Deep Learning
We describe a novel deep learning neural network method and its application to impute assay pIC50 values. Unlike conventional machine learning approaches, this method is trained on sparse bioactivity data as input, typical of that found in public and commercial databases, enabling it to learn directly from correlations between activities measured in different assays. In two case studies on public domain data sets we show that the neural network method outperforms traditional quantitative structure-activity relationship (QSAR) models and other leading approaches. Furthermore, by focussing on only the most confident predictions the accuracy is increased to R2>0.9 using our method, as compared to R2=0.43 using the leading profile-QSAR approach.
- 12 February 2019
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 59 (3) , 1197-1204
Imputation of Assay Bioactivity Data Using Deep Learning
We describe a novel deep learning neural network method and its application to impute assay pIC50 values. Unlike conventional machine learning approaches, this method is trained on sparse bioactivity data as input, typical of that found in public and commercial databases, enabling it to learn directly from correlations between activities measured in different assays. In two case studies on public domain data sets we show that the neural network method outperforms traditional quantitative structure-activity relationship (QSAR) models and other leading approaches. Furthermore, by focussing on only the most confident predictions the accuracy is increased to R2>0.9 using our method, as compared to R2=0.43 using the leading profile-QSAR approach.