Sebastian Swanson
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Sebastian Swanson has a total of 16 co-authors. Here are the top co-authors by number of shared publications:
- Aaron Lee Feller
University of Texas at Austin
2 shared publications - Amy E. Keating
Department of Biology, Massachusetts Institute of Technology
2 shared publications Kristine Deibler
Novo Nordisk 33 Hayden Avenue
2 shared publications- Kristine Deibler
Novo Nordisk, Lexington, MA, USA
2 shared publications - Andrew Savinov
Department of Biology, Massachusetts Institute of Technology, Cambridge, MA, USA
1 shared publication Claus O. Wilke
The University of Texas at Austin 2500 Speedway
1 shared publication
Latest Publications
- 13 July 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (14) , 8110-8122
Scaling SMILES-Based Chemical Language Models for Therapeutic Peptide Engineering
Therapeutic peptides occupy a unique middle ground in drug discovery, offering the high specificity of protein interactions with the chemical diversity of small molecules, yet they currently fall in a computational blind spot. Existing foundation models cannot handle them effectively: protein models are restricted to natural amino acids, while chemical models struggle to process large, polymer-like sequences. This disconnect has forced the field to rely on static chemical descriptors that fail to capture subtle chemical details or on complex multiembedding pipelines that are custom tailored to specific data sets. To bridge this gap, we present PeptideCLM-2, a suite of chemical language models trained on over 100 million molecules to natively represent complex peptide chemistry. This modeling approach expands the available toolkit of machine learning models for therapeutic peptides. Benchmarking results show strong performance versus prior methods for predicting development end points including membrane diffusion, biological function, and half-life.
- 13 July 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (14) , 8110-8122
Scaling SMILES-Based Chemical Language Models for Therapeutic Peptide Engineering
Therapeutic peptides occupy a unique middle ground in drug discovery, offering the high specificity of protein interactions with the chemical diversity of small molecules, yet they currently fall in a computational blind spot. Existing foundation models cannot handle them effectively: protein models are restricted to natural amino acids, while chemical models struggle to process large, polymer-like sequences. This disconnect has forced the field to rely on static chemical descriptors that fail to capture subtle chemical details or on complex multiembedding pipelines that are custom tailored to specific data sets. To bridge this gap, we present PeptideCLM-2, a suite of chemical language models trained on over 100 million molecules to natively represent complex peptide chemistry. This modeling approach expands the available toolkit of machine learning models for therapeutic peptides. Benchmarking results show strong performance versus prior methods for predicting development end points including membrane diffusion, biological function, and half-life.
- 28 April 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (9) , 5206-5216
Predicting Peptide Aggregation with Protein Language Model Embeddings
Amyloid fibrils, a form of peptide aggregates, are associated with multiple diseases and hinder the development of therapeutics. The experimental characterization of aggregating peptides is resource-intensive, and data are scarce, limiting the development of accurate models. We present a deep-learning model, PALM (Predicting Aggregation with Language Model embeddings), which uses transfer learning to predict aggregation from embeddings extracted from a pretrained protein language model (pLM). PALM is trained on the WaltzDB-2.0 dataset to classify peptides and identify aggregation-prone regions within a sequence at single-residue resolution. Compared to existing models, it exhibits competitive performance on diverse held-out experimental datasets. We find that PALM fails to identify single mutations that increase the rate of aggregation of the amyloid beta peptide; however, training the PALM architecture on a larger dataset, CANYA NNK1-3, substantially improves performance in this task. These results show that transfer learning with pLM embeddings improves performance when training on small datasets, but highlight that challenging tasks, such as predicting the effect of single mutations, require more experimental data.
- 28 April 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (9) , 5206-5216
Predicting Peptide Aggregation with Protein Language Model Embeddings
Amyloid fibrils, a form of peptide aggregates, are associated with multiple diseases and hinder the development of therapeutics. The experimental characterization of aggregating peptides is resource-intensive, and data are scarce, limiting the development of accurate models. We present a deep-learning model, PALM (Predicting Aggregation with Language Model embeddings), which uses transfer learning to predict aggregation from embeddings extracted from a pretrained protein language model (pLM). PALM is trained on the WaltzDB-2.0 dataset to classify peptides and identify aggregation-prone regions within a sequence at single-residue resolution. Compared to existing models, it exhibits competitive performance on diverse held-out experimental datasets. We find that PALM fails to identify single mutations that increase the rate of aggregation of the amyloid beta peptide; however, training the PALM architecture on a larger dataset, CANYA NNK1-3, substantially improves performance in this task. These results show that transfer learning with pLM embeddings improves performance when training on small datasets, but highlight that challenging tasks, such as predicting the effect of single mutations, require more experimental data.
- Open Access
Scaling SMILES-based chemical language models for therapeutic peptide engineering
Therapeutic peptides occupy a unique middle ground in drug discovery, offering the high specificity of protein interactions with the chemical diversity of small molecules, yet they currently fall in a computational blind spot. Existing foundation models cannot handle them effectively: protein models are restricted to natural amino acids, while chemical models struggle to process large, polymer-like sequences. This disconnect has forced the field to rely on static chemical descriptors that fail to capture subtle chemical details or on complex multi-embedding pipelines that are custom tailored to specific datasets. To bridge this gap, we present PeptideCLM-2, a suite of chemical language models trained on over 100 million molecules to natively represent complex peptide chemistry. This modeling approach expands the available toolkit of machine learning models for therapeutic peptides. Benchmarking results show strong performance versus prior methods for predicting development endpoints including membrane diffusion, biological function, and half life.
Open AccessScaling SMILES-based chemical language models for therapeutic peptide engineering
Therapeutic peptides occupy a unique middle ground in drug discovery, offering the high specificity of protein interactions with the chemical diversity of small molecules, yet they currently fall in a computational blind spot. Existing foundation models cannot handle them effectively: protein models are restricted to natural amino acids, while chemical models struggle to process large, polymer-like sequences. This disconnect has forced the field to rely on static chemical descriptors that fail to capture subtle chemical details or on complex multi-embedding pipelines that are custom tailored to specific datasets. To bridge this gap, we present PeptideCLM-2, a suite of chemical language models trained on over 100 million molecules to natively represent complex peptide chemistry. This modeling approach expands the available toolkit of machine learning models for therapeutic peptides. Benchmarking results show strong performance versus prior methods for predicting development endpoints including membrane diffusion, biological function, and half life.
- Open Access
Predicting peptide aggregation with protein language model embeddings
Amyloid fibrils, a form of peptide aggregate, are associated with multiple diseases and hinder the development of therapeutics. The experimental characterization of aggregating peptides is resource-intensive and data are scarce, limiting the development of accurate models. We present a deep-learning model, PALM (Predicting Aggregation with Language Model embeddings), which uses transfer learning to predict aggregation from embeddings extracted from a pretrained protein language model (pLM). PALM is trained on the WaltzDB-2.0 dataset to classify peptides and identify aggregation-prone regions within a sequence at single-residue resolution. Compared to existing models, it exhibits competitive performance on diverse held-out experimental datasets. We find that PALM fails to identify single mutations that increase the rate of aggregation of amyloid beta peptide; however, training the PALM architecture on a larger dataset, CANYA NNK1-3, substantially improves performance in this task. These results show that transfer learning with pLM embeddings improves performance when training on small datasets, but highlight that challenging tasks, such as predicting the effect of single mutations, require more experimental data.
Open AccessPredicting peptide aggregation with protein language model embeddings
Amyloid fibrils, a form of peptide aggregate, are associated with multiple diseases and hinder the development of therapeutics. The experimental characterization of aggregating peptides is resource-intensive and data are scarce, limiting the development of accurate models. We present a deep-learning model, PALM (Predicting Aggregation with Language Model embeddings), which uses transfer learning to predict aggregation from embeddings extracted from a pretrained protein language model (pLM). PALM is trained on the WaltzDB-2.0 dataset to classify peptides and identify aggregation-prone regions within a sequence at single-residue resolution. Compared to existing models, it exhibits competitive performance on diverse held-out experimental datasets. We find that PALM fails to identify single mutations that increase the rate of aggregation of amyloid beta peptide; however, training the PALM architecture on a larger dataset, CANYA NNK1-3, substantially improves performance in this task. These results show that transfer learning with pLM embeddings improves performance when training on small datasets, but highlight that challenging tasks, such as predicting the effect of single mutations, require more experimental data.
- Open Access
High-throughput discovery of inhibitory protein fragments with AlphaFold
Peptides can bind to specific sites on larger proteins and thereby function as inhibitors and regulatory elements. Peptide fragments of larger proteins are particularly attractive for achieving these functions due to their inherent potential to form native-like binding interactions. Recently developed experimental approaches allow for high-throughput measurement of protein fragment inhibitory activity in living cells. However, it has thus far not been possible to predict de novo which of the many possible protein fragments bind to protein targets, let alone act as inhibitors. We have developed a computational method, FragFold, that employs AlphaFold to predict protein fragment binding to full-length proteins in a high-throughput manner. Applying FragFold to thousands of fragments tiling across diverse proteins revealed peaks of predicted binding along each protein sequence. Comparisons with experimental measurements establish that our approach is a sensitive predictor of fragment function: Evaluating inhibitory fragments from known protein-protein interaction interfaces, we find 87% are predicted by FragFold to bind in a native-like mode. Across full protein sequences, 68% of FragFold-predicted binding peaks match experimentally measured inhibitory peaks. Deep mutational scanning experiments support the predicted binding modes and uncover superior inhibitory peptides in high throughput. Further, FragFold is able to predict previously unknown protein binding modes, explaining prior genetic and biochemical data. The success rate of FragFold demonstrates that this computational approach should be broadly applicable for discovering inhibitory protein fragments across proteomes. Significance Statement: Peptides can regulate protein interactions by binding to specific interfaces, and fragments of larger proteins have high potential to function in this manner. Recently developed experimental methods allow massively parallel measurement of protein fragment-based inhibition in vivo . However, we have lacked comparable computational methods to predict which protein fragments act as inhibitors and how they bind. Here we report a new approach, FragFold, which leverages high-throughput AlphaFold predictions of protein – fragment binding to tackle these problems at scale. FragFold is successful at predicting inhibitory protein fragments and their binding modes across diverse protein structures and functions. This new approach stands to enable proteome-wide discovery of inhibitory protein fragments and aid the interpretation of high-throughput experimental measurements of inhibitory activity. Classification: Biological Sciences / Biophysics and Computational Biology
Open AccessHigh-throughput discovery of inhibitory protein fragments with AlphaFold
Peptides can bind to specific sites on larger proteins and thereby function as inhibitors and regulatory elements. Peptide fragments of larger proteins are particularly attractive for achieving these functions due to their inherent potential to form native-like binding interactions. Recently developed experimental approaches allow for high-throughput measurement of protein fragment inhibitory activity in living cells. However, it has thus far not been possible to predict de novo which of the many possible protein fragments bind to protein targets, let alone act as inhibitors. We have developed a computational method, FragFold, that employs AlphaFold to predict protein fragment binding to full-length proteins in a high-throughput manner. Applying FragFold to thousands of fragments tiling across diverse proteins revealed peaks of predicted binding along each protein sequence. Comparisons with experimental measurements establish that our approach is a sensitive predictor of fragment function: Evaluating inhibitory fragments from known protein-protein interaction interfaces, we find 87% are predicted by FragFold to bind in a native-like mode. Across full protein sequences, 68% of FragFold-predicted binding peaks match experimentally measured inhibitory peaks. Deep mutational scanning experiments support the predicted binding modes and uncover superior inhibitory peptides in high throughput. Further, FragFold is able to predict previously unknown protein binding modes, explaining prior genetic and biochemical data. The success rate of FragFold demonstrates that this computational approach should be broadly applicable for discovering inhibitory protein fragments across proteomes. Significance Statement: Peptides can regulate protein interactions by binding to specific interfaces, and fragments of larger proteins have high potential to function in this manner. Recently developed experimental methods allow massively parallel measurement of protein fragment-based inhibition in vivo . However, we have lacked comparable computational methods to predict which protein fragments act as inhibitors and how they bind. Here we report a new approach, FragFold, which leverages high-throughput AlphaFold predictions of protein – fragment binding to tackle these problems at scale. FragFold is successful at predicting inhibitory protein fragments and their binding modes across diverse protein structures and functions. This new approach stands to enable proteome-wide discovery of inhibitory protein fragments and aid the interpretation of high-throughput experimental measurements of inhibitory activity. Classification: Biological Sciences / Biophysics and Computational Biology
- Open Access
- 10 May 2022
- Published by Wiley in Protein Science
- Vol. 31 (6) , e4322
Tertiary motifs as building blocks for the design of protein‐binding peptides
Despite advances in protein engineering, the de novo design of small proteins or peptides that bind to a desired target remains a difficult task. Most computational methods search for binder structures in a library of candidate scaffolds, which can lead to designs with poor target complementarity and low success rates. Instead of choosing from pre‐defined scaffolds, we propose that custom peptide structures can be constructed to complement a target surface. Our method mines tertiary motifs (TERMs) from known structures to identify surface‐complementing fragments or “seeds.” We combine seeds that satisfy geometric overlap criteria to generate peptide backbones and score the backbones to identify the most likely binding structures. We found that TERM‐based seeds can describe known binding structures with high resolution: the vast majority of peptide binders from 486 peptide‐protein complexes can be covered by seeds generated from single‐chain structures. Furthermore, we demonstrate that known peptide structures can be reconstructed with high accuracy from peptide‐covering seeds. As a proof of concept, we used our method to design 100 peptide binders of TRAF6, seven of which were predicted by Rosetta to form higher‐quality interfaces than a native binder. The designed peptides interact with distinct sites on TRAF6, including the native peptide‐binding site. These results demonstrate that known peptide‐binding structures can be constructed from TERMs in single‐chain structures and suggest that TERM information can be applied to efficiently design novel target‐complementing binders.
- 10 May 2022
- Published by Wiley in Protein Science
- Vol. 31 (6) , e4322
Open AccessTertiary motifs as building blocks for the design of protein‐binding peptides
Despite advances in protein engineering, the de novo design of small proteins or peptides that bind to a desired target remains a difficult task. Most computational methods search for binder structures in a library of candidate scaffolds, which can lead to designs with poor target complementarity and low success rates. Instead of choosing from pre‐defined scaffolds, we propose that custom peptide structures can be constructed to complement a target surface. Our method mines tertiary motifs (TERMs) from known structures to identify surface‐complementing fragments or “seeds.” We combine seeds that satisfy geometric overlap criteria to generate peptide backbones and score the backbones to identify the most likely binding structures. We found that TERM‐based seeds can describe known binding structures with high resolution: the vast majority of peptide binders from 486 peptide‐protein complexes can be covered by seeds generated from single‐chain structures. Furthermore, we demonstrate that known peptide structures can be reconstructed with high accuracy from peptide‐covering seeds. As a proof of concept, we used our method to design 100 peptide binders of TRAF6, seven of which were predicted by Rosetta to form higher‐quality interfaces than a native binder. The designed peptides interact with distinct sites on TRAF6, including the native peptide‐binding site. These results demonstrate that known peptide‐binding structures can be constructed from TERMs in single‐chain structures and suggest that TERM information can be applied to efficiently design novel target‐complementing binders.