Oliver Marcus Turnbull
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Oliver Marcus Turnbull has a total of 64 co-authors. Here are the top co-authors by number of shared publications:
- Charlotte Mary Deane
Oxford Protein Informatics Group, Department of Statistics, University of Oxford, Oxford OX1 3LB, UK
9 shared publications - Anna Carbery
University of Oxford, Diamond Light Source
2 shared publications Annabel Suter
Oxford Protein Informatics Group, Department of Statistics, University of Oxford, Oxford, UK
2 shared publicationsDino Oglic
Centre for AI, Biopharmaceuticals R&D, AstraZeneca , Cambridge, CB2 0AA
2 shared publications- Frank von Delft
Research Complex at Harwell, Harwell Science and Innovation Campus, Didcot OX11 0FA, UK
2 shared publications - Harold Grosjean
Structural Bioinformatics and Computational Biochemistry, Department of Biochemistry, University of Oxford, South Parks Road, Oxford, UK
2 shared publications
Latest Publications
- Open Access
Generative Language Modeling for Antibody CDR Grafting and Alignment-driven De Novo Design
Antibodies recognise their targets through hypervariable complementarity-determining regions (CDRs), which are interleaved with conserved frameworks in sequence space, making de novo CDR design an infilling problem. Autoregressive models generate residues left-to-right, which precludes full framework context during CDR generation and conflates framework and CDR likelihoods, leaving no natural prompt–response interface for feedback to steer generation. We present GenCDR, a family of LLaMa-based autoregressive language models that read all frameworks as a conditioning prompt and generate all CDRs jointly as a variable-length response, making CDR likelihoods a clean, separable target for reward attribution. The family comprises IgGenCDR , p-IgGenCDR , and NanoGenCDR , trained on unpaired, paired, and nanobody chains, respectively. GenCDR achieves the highest CDR recovery among autoregressive models and produces natural, diverse, human-like CDRs whose likelihoods correlate with fitness and developability assays. The prompt–response boundary also enables principled alignment: reward signals for binding affinity, expression, or developability can be composed to steer CDR generation. Over four rounds of alignment against antibody–antigen co-folding and developability objectives, we find that NanoGenCDR, which uses no explicit antigen encoding, can reach in silico structural interface metrics competitive with those of a structure-conditioned diffusion pipeline at roughly half the sampling budget, with more natural, developable designs. The same interface can be extended to integrate experimental feedback, opening a path to closed-loop antibody de novo design.
Open AccessGenerative Language Modeling for Antibody CDR Grafting and Alignment-driven De Novo Design
Antibodies recognise their targets through hypervariable complementarity-determining regions (CDRs), which are interleaved with conserved frameworks in sequence space, making de novo CDR design an infilling problem. Autoregressive models generate residues left-to-right, which precludes full framework context during CDR generation and conflates framework and CDR likelihoods, leaving no natural prompt–response interface for feedback to steer generation. We present GenCDR, a family of LLaMa-based autoregressive language models that read all frameworks as a conditioning prompt and generate all CDRs jointly as a variable-length response, making CDR likelihoods a clean, separable target for reward attribution. The family comprises IgGenCDR , p-IgGenCDR , and NanoGenCDR , trained on unpaired, paired, and nanobody chains, respectively. GenCDR achieves the highest CDR recovery among autoregressive models and produces natural, diverse, human-like CDRs whose likelihoods correlate with fitness and developability assays. The prompt–response boundary also enables principled alignment: reward signals for binding affinity, expression, or developability can be composed to steer CDR generation. Over four rounds of alignment against antibody–antigen co-folding and developability objectives, we find that NanoGenCDR, which uses no explicit antigen encoding, can reach in silico structural interface metrics competitive with those of a structure-conditioned diffusion pipeline at roughly half the sampling budget, with more natural, developable designs. The same interface can be extended to integrate experimental feedback, opening a path to closed-loop antibody de novo design.
- Top Cited PapersOpen Access
- 22 February 2026
- Published by Taylor & Francis in mAbs
- Vol. 18 (1) , 2634216
2025 Ginkgo Datapoints Antibody Developability Competition outcomes: limited model performance and a call for data standardization
Monoclonal antibodies (mAbs) are among the most successful therapeutic modalities, with over 200 marketed products and nearly 1400 candidates in clinical development.1 While high-affinity binding r...
- 22 February 2026
- Published by Taylor & Francis in mAbs
- Vol. 18 (1) , 2634216
Top Cited PapersOpen Access2025 Ginkgo Datapoints Antibody Developability Competition outcomes: limited model performance and a call for data standardization
Monoclonal antibodies (mAbs) are among the most successful therapeutic modalities, with over 200 marketed products and nearly 1400 candidates in clinical development.1 While high-affinity binding r...
- Open Access
- 1 November 2024
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 40 (11)
p-IgGen: a paired antibody generative language model
A key challenge in antibody drug discovery is designing novel sequences that are free from developability issues—such as aggregation, polyspecificity, poor expression, or low solubility. Here we present p-IgGen, a protein language model for paired heavy-light chain antibody generation. The model generates diverse, antibody-like sequences with pairing properties found in natural antibodies. We also create a finetuned version of p-IgGen that biases the model to generate antibodies with 3D biophysical properties that fall within distributions seen in clinical-stage therapeutic antibodies. The model and inference code are freely available at www.github.com/oxpig/p-IgGen. Cleaned training data is deposited at doi.org/10.5281/zenodo.13880874. Supplementary data are available at Bioinformatics online.
- 1 November 2024
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 40 (11)
Open Accessp-IgGen: a paired antibody generative language model
A key challenge in antibody drug discovery is designing novel sequences that are free from developability issues—such as aggregation, polyspecificity, poor expression, or low solubility. Here we present p-IgGen, a protein language model for paired heavy-light chain antibody generation. The model generates diverse, antibody-like sequences with pairing properties found in natural antibodies. We also create a finetuned version of p-IgGen that biases the model to generate antibodies with 3D biophysical properties that fall within distributions seen in clinical-stage therapeutic antibodies. The model and inference code are freely available at www.github.com/oxpig/p-IgGen. Cleaned training data is deposited at doi.org/10.5281/zenodo.13880874. Supplementary data are available at Bioinformatics online.
- Open Access
p-IgGen: A Paired Antibody Generative Language Model
Summary: A key challenge in antibody drug discovery is designing novel sequences that are free from developability issues - such as aggregation, polyspecificity, poor expression, or low solubility. Here, we present p-IgGen, a protein language model for paired heavylight chain antibody generation. The model generates diverse, antibody-like sequences with pairing properties found in natural antibodies. We also create a finetuned version of p-IgGen that biases the model to generate antibodies with 3D biophysical properties that fall within distributions seen in clinical-stage therapeutic antibodies.
Open Accessp-IgGen: A Paired Antibody Generative Language Model
Summary: A key challenge in antibody drug discovery is designing novel sequences that are free from developability issues - such as aggregation, polyspecificity, poor expression, or low solubility. Here, we present p-IgGen, a protein language model for paired heavylight chain antibody generation. The model generates diverse, antibody-like sequences with pairing properties found in natural antibodies. We also create a finetuned version of p-IgGen that biases the model to generate antibodies with 3D biophysical properties that fall within distributions seen in clinical-stage therapeutic antibodies.
- Open Access
The Observed T cell receptor Space database enables paired-chain repertoire mining, coherence analysis and language modelling
T cell activation is governed through T cell receptors (TCRs), heterodimers of two sequence-variable chains (often an alpha [ α ] and beta [ β ] chain) that recognise linear antigen fragments presented on the cell surfaces. Early sequencing technologies limited the study of immune repertoire TCRs to unpaired transcripts, leading to extensive analysis of β -chain data alone as its greater sequence diversity suggested it should dominate antigen recognition. Over time, structural data has revealed that both α and β chains contribute to binding most antigens and highthroughput single-cell handling technologies have been increasingly applied to obtain samples of complete TCR variable region sequences from repertoires. Despite this, there is currently no repository dedicated to the curation of publicly available paired TCR sequence data. We have addressed this gap by creating the Observed T cell receptor Space (OTS) database, a source of consistently processed and annotated, full-length, paired-chain TCR sequencing data from 50 studies and at least 75 individuals. Currently, OTS contains 5.35M redundant (1.63M nonredundant) predominantly human TCR sequences and, based on recent data availability trends, will grow rapidly. We perform an initial analysis of OTS, leading to the identification of pairing biases, public TCRs, and distinct chain coherence patterns relative to antibodies. We also harness the data to build a publicly available paired-chain TCR language model, providing paired embedding representations and a method for residue in-filling that is conditional on the partner chain. OTS will be updated and maintained as a central community resource and is freely downloadable and available as a web application at https://opig.stats.ox.ac.uk/webapps/ots .
Open AccessThe Observed T cell receptor Space database enables paired-chain repertoire mining, coherence analysis and language modelling
T cell activation is governed through T cell receptors (TCRs), heterodimers of two sequence-variable chains (often an alpha [ α ] and beta [ β ] chain) that recognise linear antigen fragments presented on the cell surfaces. Early sequencing technologies limited the study of immune repertoire TCRs to unpaired transcripts, leading to extensive analysis of β -chain data alone as its greater sequence diversity suggested it should dominate antigen recognition. Over time, structural data has revealed that both α and β chains contribute to binding most antigens and highthroughput single-cell handling technologies have been increasingly applied to obtain samples of complete TCR variable region sequences from repertoires. Despite this, there is currently no repository dedicated to the curation of publicly available paired TCR sequence data. We have addressed this gap by creating the Observed T cell receptor Space (OTS) database, a source of consistently processed and annotated, full-length, paired-chain TCR sequencing data from 50 studies and at least 75 individuals. Currently, OTS contains 5.35M redundant (1.63M nonredundant) predominantly human TCR sequences and, based on recent data availability trends, will grow rapidly. We perform an initial analysis of OTS, leading to the identification of pairing biases, public TCRs, and distinct chain coherence patterns relative to antibodies. We also harness the data to build a publicly available paired-chain TCR language model, providing paired embedding representations and a method for residue in-filling that is conditional on the partner chain. OTS will be updated and maintained as a central community resource and is freely downloadable and available as a web application at https://opig.stats.ox.ac.uk/webapps/ots .
- Open Access
- 8 January 2024
- Published by Springer Nature in Communications Biology
- Vol. 7 (1) , 1-13
Contextualising the developability risk of antibodies with lambda light chains using enhanced therapeutic antibody profiling
Antibodies with lambda light chains (λ-antibodies) are generally considered to be less developable than those with kappa light chains (κ-antibodies). Though this hypothesis has not been formally established, it has led to substantial systematic biases in drug discovery pipelines and thus contributed to kappa dominance amongst clinical-stage therapeutics. However, the identification of increasing numbers of epitopes preferentially engaged by λ-antibodies shows there is a functional cost to neglecting to consider them as potential lead candidates. Here, we update our Therapeutic Antibody Profiler (TAP) tool to use the latest data and machine learning-based structure prediction, and apply it to evaluate developability risk profiles for κ-antibodies and λ-antibodies based on their surface physicochemical properties. We find that while human λ-antibodies on average have a higher risk of developability issues than κ-antibodies, a sizeable proportion are assigned lower-risk profiles by TAP and should represent more tractable candidates for therapeutic development. Through a comparative analysis of the low- and high-risk populations, we highlight opportunities for strategic design that TAP suggests would enrich for more developable λ-antibodies. Overall, we provide context to the differing developability of κ- and λ-antibodies, enabling a rational approach to incorporate more diversity into the initial pool of immunotherapeutic candidates.
- 8 January 2024
- Published by Springer Nature in Communications Biology
- Vol. 7 (1) , 1-13
Open AccessContextualising the developability risk of antibodies with lambda light chains using enhanced therapeutic antibody profiling
Antibodies with lambda light chains (λ-antibodies) are generally considered to be less developable than those with kappa light chains (κ-antibodies). Though this hypothesis has not been formally established, it has led to substantial systematic biases in drug discovery pipelines and thus contributed to kappa dominance amongst clinical-stage therapeutics. However, the identification of increasing numbers of epitopes preferentially engaged by λ-antibodies shows there is a functional cost to neglecting to consider them as potential lead candidates. Here, we update our Therapeutic Antibody Profiler (TAP) tool to use the latest data and machine learning-based structure prediction, and apply it to evaluate developability risk profiles for κ-antibodies and λ-antibodies based on their surface physicochemical properties. We find that while human λ-antibodies on average have a higher risk of developability issues than κ-antibodies, a sizeable proportion are assigned lower-risk profiles by TAP and should represent more tractable candidates for therapeutic development. Through a comparative analysis of the low- and high-risk populations, we highlight opportunities for strategic design that TAP suggests would enrich for more developable λ-antibodies. Overall, we provide context to the differing developability of κ- and λ-antibodies, enabling a rational approach to incorporate more diversity into the initial pool of immunotherapeutic candidates.
- Open Access
Contextualising the developability risk of antibodies with lambda light chains using enhanced therapeutic antibody profiling
Antibodies with lambda light chains ( λ -antibodies) are generally considered to be less developable than those with kappa light chains ( κ -antibodies), leading to substantial systematic biases in drug discovery pipelines. This has contributed to kappa dominance amongst clinical-stage therapeutics. However, the identification of increasing numbers of epitopes preferentially engaged by λ -antibodies shows there is a functional cost to neglecting them as potential lead candidates during discovery campaigns. Here, we update our Therapeutic Antibody Profiler (TAP) tool to use the latest data and machine learning-based structure prediction methods, and apply this new protocol to evaluate developability risk profiles for κ -antibodies and λ -antibodies based on their surface physicochemical properties. We find that λ -antibodies are on average at a higher risk of poor developability — as an indication, over 40% of single-cell sequenced human λ -antibodies are flagged by TAP for risk-prone patches of surface hydrophobicity (PSH), compared to around 11% of human κ -antibodies. Nonetheless, a substantial proportion of natural λ -antibodies are assigned more moderate risk profiles by TAP and should therefore represent more tractable candidates for therapeutic development. We also analyse the populations of high and low risk antibodies, highlighting opportunities for strategic design that TAP suggests would enrich for more developable λ -based candidates. Overall, we provide context to the differing developability of κ - and λ -antibodies, enabling a rational approach to incorporate more diversity into the initial pool of immunotherapeutic candidates.
Open AccessContextualising the developability risk of antibodies with lambda light chains using enhanced therapeutic antibody profiling
Antibodies with lambda light chains ( λ -antibodies) are generally considered to be less developable than those with kappa light chains ( κ -antibodies), leading to substantial systematic biases in drug discovery pipelines. This has contributed to kappa dominance amongst clinical-stage therapeutics. However, the identification of increasing numbers of epitopes preferentially engaged by λ -antibodies shows there is a functional cost to neglecting them as potential lead candidates during discovery campaigns. Here, we update our Therapeutic Antibody Profiler (TAP) tool to use the latest data and machine learning-based structure prediction methods, and apply this new protocol to evaluate developability risk profiles for κ -antibodies and λ -antibodies based on their surface physicochemical properties. We find that λ -antibodies are on average at a higher risk of poor developability — as an indication, over 40% of single-cell sequenced human λ -antibodies are flagged by TAP for risk-prone patches of surface hydrophobicity (PSH), compared to around 11% of human κ -antibodies. Nonetheless, a substantial proportion of natural λ -antibodies are assigned more moderate risk profiles by TAP and should therefore represent more tractable candidates for therapeutic development. We also analyse the populations of high and low risk antibodies, highlighting opportunities for strategic design that TAP suggests would enrich for more developable λ -based candidates. Overall, we provide context to the differing developability of κ - and λ -antibodies, enabling a rational approach to incorporate more diversity into the initial pool of immunotherapeutic candidates.
- Open Access
- 11 May 2023
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 63 (10) , 2960-2974
A Small Step Toward Generalizability: Training a Machine Learning Scoring Function for Structure-Based Virtual Screening
Over the past few years, many machine learning-based scoring functions for predicting the binding of small molecules to proteins have been developed. Their objective is to approximate the distribution which takes two molecules as input and outputs the energy of their interaction. Only a scoring function that accounts for the interatomic interactions involved in binding can accurately predict binding affinity on unseen molecules. However, many scoring functions make predictions based on data set biases rather than an understanding of the physics of binding. These scoring functions perform well when tested on similar targets to those in the training set but fail to generalize to dissimilar targets. To test what a machine learning-based scoring function has learned, input attribution, a technique for learning which features are important to a model when making a prediction on a particular data point, can be applied. If a model successfully learns something beyond data set biases, attribution should give insight into the important binding interactions that are taking place. We built a machine learning-based scoring function that aimed to avoid the influence of bias via thorough train and test data set filtering and show that it achieves comparable performance on the Comparative Assessment of Scoring Functions, 2016 (CASF-2016) benchmark to other leading methods. We then use the CASF-2016 test set to perform attribution and find that the bonds identified as important by PointVS, unlike those extracted from other scoring functions, have a high correlation with those found by a distance-based interaction profiler. We then show that attribution can be used to extract important binding pharmacophores from a given protein target when supplied with a number of bound structures. We use this information to perform fragment elaboration and see improvements in docking scores compared to using structural information from a traditional, data-based approach. This not only provides definitive proof that the scoring function has learned to identify some important binding interactions but also constitutes the first deep learning-based method for extracting structural information from a target for molecule design.
- 11 May 2023
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 63 (10) , 2960-2974
Open AccessA Small Step Toward Generalizability: Training a Machine Learning Scoring Function for Structure-Based Virtual Screening
Over the past few years, many machine learning-based scoring functions for predicting the binding of small molecules to proteins have been developed. Their objective is to approximate the distribution which takes two molecules as input and outputs the energy of their interaction. Only a scoring function that accounts for the interatomic interactions involved in binding can accurately predict binding affinity on unseen molecules. However, many scoring functions make predictions based on data set biases rather than an understanding of the physics of binding. These scoring functions perform well when tested on similar targets to those in the training set but fail to generalize to dissimilar targets. To test what a machine learning-based scoring function has learned, input attribution, a technique for learning which features are important to a model when making a prediction on a particular data point, can be applied. If a model successfully learns something beyond data set biases, attribution should give insight into the important binding interactions that are taking place. We built a machine learning-based scoring function that aimed to avoid the influence of bias via thorough train and test data set filtering and show that it achieves comparable performance on the Comparative Assessment of Scoring Functions, 2016 (CASF-2016) benchmark to other leading methods. We then use the CASF-2016 test set to perform attribution and find that the bonds identified as important by PointVS, unlike those extracted from other scoring functions, have a high correlation with those found by a distance-based interaction profiler. We then show that attribution can be used to extract important binding pharmacophores from a given protein target when supplied with a number of bound structures. We use this information to perform fragment elaboration and see improvements in docking scores compared to using structural information from a traditional, data-based approach. This not only provides definitive proof that the scoring function has learned to identify some important binding interactions but also constitutes the first deep learning-based method for extracting structural information from a target for molecule design.
- Open Access
A Step Towards Generalisability: Training a Machine Learning Scoring Function for Structure-Based Virtual Screening
Over the last few years, many machine learning-based scoring functions for predicting the binding of small molecules to proteins have been developed. Their objective is to approximate the distribution which takes two molecules as input and outputs the energy of their interaction. Only a scoring function that accounts for the interatomic interactions involved in binding can accurately predict binding affinity on unseen molecules. However, many scoring functions make predictions based on dataset biases rather than an understanding of the physics of binding. These scoring functions perform well when tested on similar targets to those in the training set, but fail to generalise to dissimilar targets. To test what a machine learning-based scoring function has learnt, input attribution—a technique for learning which features are important to a model when making a prediction on a particular data point—can be applied. If a model successfully learns something beyond dataset biases, attribution should give insight into the important binding interactions that are taking place. We built a machine learning-based scoring function that aimed to avoid the influence of bias via thorough train and test dataset filtering, and show that it achieves comparable performance on the CASF-2016 benchmark to other leading methods. We then use the CASF-2016 test set to perform attribution, and find that the bonds identified as important by PointVS, unlike those extracted from other scoring functions, have a high correlation with those found by a distance-based interaction profiler. We then show that attribution can be used to extract important binding pharmacophores from a given protein target when supplied with a number of bound structures. We use this information to perform fragment elaboration, and see improvements in docking scores compared to using structural information from a traditional, data-based approach. This not only provides definitive proof that the scoring function has learnt to identify some important binding interactions, but also constitutes the first deep learning-based method for extracting structural information from a target for molecule design.
Open AccessA Step Towards Generalisability: Training a Machine Learning Scoring Function for Structure-Based Virtual Screening
Over the last few years, many machine learning-based scoring functions for predicting the binding of small molecules to proteins have been developed. Their objective is to approximate the distribution which takes two molecules as input and outputs the energy of their interaction. Only a scoring function that accounts for the interatomic interactions involved in binding can accurately predict binding affinity on unseen molecules. However, many scoring functions make predictions based on dataset biases rather than an understanding of the physics of binding. These scoring functions perform well when tested on similar targets to those in the training set, but fail to generalise to dissimilar targets. To test what a machine learning-based scoring function has learnt, input attribution—a technique for learning which features are important to a model when making a prediction on a particular data point—can be applied. If a model successfully learns something beyond dataset biases, attribution should give insight into the important binding interactions that are taking place. We built a machine learning-based scoring function that aimed to avoid the influence of bias via thorough train and test dataset filtering, and show that it achieves comparable performance on the CASF-2016 benchmark to other leading methods. We then use the CASF-2016 test set to perform attribution, and find that the bonds identified as important by PointVS, unlike those extracted from other scoring functions, have a high correlation with those found by a distance-based interaction profiler. We then show that attribution can be used to extract important binding pharmacophores from a given protein target when supplied with a number of bound structures. We use this information to perform fragment elaboration, and see improvements in docking scores compared to using structural information from a traditional, data-based approach. This not only provides definitive proof that the scoring function has learnt to identify some important binding interactions, but also constitutes the first deep learning-based method for extracting structural information from a target for molecule design.