CoChat
Sign Up
PreprintCited by 9Protein purification and stabilityComputational Drug Discovery MethodsMachine Learning in Bioinformatics

The DOI matches its Crossref record; 1 of the 1 references it cites were found in a registry or an index.

Machine learning driven acceleration of biopharmaceutical formulation development using Excipient Prediction Software (ExPreSo)

Open access
Vidal-Henriquez, Estefania, Holder, Thomas, Lee, Nicholas Franciss, Pompe, Cornelius, Teese, Mark George·2025·Computational and Structural Biotechnology Journal
DOI 10.1101/2025.02.12.637685PMC PMC12597001

Preprint — published as 10.1016/j.csbj.2025.10.026 · Paper page

Open this paper in CoChat — free accountOpen PDF

Abstract

Formulation development of protein biopharmaceuticals has become increasingly challenging due to new modalities and higher target drug substance concentrations. The limited amount of drug substance available during development, coupled with extensive analytical requirements, restrict the number of excipients that can be empirically screened. There is a strong need for in silico tools to optimize excipient pre-selection before wet lab experiments. Here, we introduce Excipient Prediction Software (ExPreSo), a supervised machine learning algorithm that suggests excipients based on the properties of the protein drug substance and target product profile. ExPreSo was trained on a dataset comprising 335 regulatory-approved peptide and protein drug products. Predictive features included protein structural properties, protein language model embeddings, and drug product characteristics. ExPreSo showed good performance for the nine most prevalent excipients in biopharmaceutical formulations and minimal overfitting. A fast variant of ExPreSo using only sequence-based input features showed similar prediction power to slower models that relied on molecular modeling. Notably, an ExPreSo variant using only protein-based input features also showed good performance, indicating resilience to the influence of platform formulations. To our knowledge, this is the first machine learning algorithm to suggest biopharmaceutical excipients based on the dataset of regulatory-approved drug products. Overall, ExPreSo shows great potential to reduce the time, costs, and risks associated with excipient screening during formulation development. Abstract Figure:

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1101/2025.02.12.637685

Cites

1 of 1 references verified0 citation errors · 0 not checkable

Statistics

Figures

Graphical abstract
Graphical abstract CC BY-NC-ND · PubMed Central
Fig. 1 Main characteristics of the dataset used for machine learning. A) Distribution of antibody heavy chain subtypes. B) Companies retained in the input features. The large “Other Company” category enables the use of this feature in ExPreSo as an independent bioinformatic tool. C) Abundance of excipients in the dataset prior to selecting the top excipients used in ExPreSo and removal of duplicate formulations (see Methods). D) Cumulative number of approved formulations and drug substances in the ExPreSo dataset, before duplicate removal (same data as in Fig. 1C).
Fig. 1 Main characteristics of the dataset used for machine learning. A) Distribution of antibody heavy chain subtypes. B) Companies retained in the input features. The large “Other Company” category enables the use of this feature in ExPreSo as an independent bioinformatic tool. C) Abundance of excipients in the dataset prior to selecting the top excipients used in ExPreSo and removal of duplicate formulations (see Methods). D) Cumulative number of approved formulations and drug substances in the ExPreSo dataset, before duplicate removal (same data as in Fig. 1C).CC BY-NC-ND · PubMed Central
Fig. 2 Predictive performance of ExPreSo for nine excipients. Model performance was evaluated using leave-one-group-out cross-validation on the train set with all predictive features (“All Features”). Results were aggregated over 10 Monte Carlo cross-validation (MCCV) runs, each using a different random seed for test set creation. In each boxplot, whiskers indicate min and max, and the central line indicates the median. A) Precision Recall Area Under the Curve (PR-AUC). The baseline (red dash) indicates the precision of a random predictor. B) Receiver Operating Characteristic Area Under the Curve (ROC-AUC). The baseline (dotted line) indicates the performance of a random predictor. C) ROC-AUC prediction power showed correlation to excipient abundance, with the exception of the acetic acid outlier. Each point represents the mean ROC-AUC from 10 MCCV runs.
Fig. 2 Predictive performance of ExPreSo for nine excipients. Model performance was evaluated using leave-one-group-out cross-validation on the train set with all predictive features (“All Features”). Results were aggregated over 10 Monte Carlo cross-validation (MCCV) runs, each using a different random seed for test set creation. In each boxplot, whiskers indicate min and max, and the central line indicates the median. A) Precision Recall Area Under the Curve (PR-AUC). The baseline (red dash) indicates the precision of a random predictor. B) Receiver Operating Characteristic Area Under the Curve (ROC-AUC). The baseline (dotted line) indicates the performance of a random predictor. C) ROC-AUC prediction power showed correlation to excipient abundance, with the exception of the acetic acid outlier. Each point represents the mean ROC-AUC from 10 MCCV runs.CC BY-NC-ND · PubMed Central
Fig. 3 All Features, Fast, Interpretable, and Protein-Only versions of ExPreSo show similar performance. ExPreSo was validated using four models, each with a different subset of input features (descriptors). The “Interpretable” model lacked sequence-based features such as pLM embeddings. The “Fast” model lacked features derived from molecular modeling. The “Protein-Based” model used only features based on the protein sequence and 3D structure. Shown are the area under the curve (AUC) metrics from leave-one-group-out cross-validation (LOGO-CV), averaged over ten Monte Carlo cross-validation (MCCV) runs (see Methods). Boxplots present Area Under the Precision Recall Curve (PR-AUC) and Area Under the Receiver Operator Characteristic Curve (ROC-AUC); the solid middle line indicates the median, whereas the dotted line indicates the mean. For the ROC-AUC, the baseline (thinner dotted line at 0.5) indicates the performance of a random predictor.
Fig. 3 All Features, Fast, Interpretable, and Protein-Only versions of ExPreSo show similar performance. ExPreSo was validated using four models, each with a different subset of input features (descriptors). The “Interpretable” model lacked sequence-based features such as pLM embeddings. The “Fast” model lacked features derived from molecular modeling. The “Protein-Based” model used only features based on the protein sequence and 3D structure. Shown are the area under the curve (AUC) metrics from leave-one-group-out cross-validation (LOGO-CV), averaged over ten Monte Carlo cross-validation (MCCV) runs (see Methods). Boxplots present Area Under the Precision Recall Curve (PR-AUC) and Area Under the Receiver Operator Characteristic Curve (ROC-AUC); the solid middle line indicates the median, whereas the dotted line indicates the mean. For the ROC-AUC, the baseline (thinner dotted line at 0.5) indicates the performance of a random predictor.CC BY-NC-ND · PubMed Central
Fig. 4 Minimal overfitting was observed across all model variants. “Overfitting index”, defined as the relative difference in performance between the LOGO-CV and the blind test set. Higher values indicate larger overfitting. Values above zero (dotted line) indicate better performance for the LOGO-CV than the blind test. The central line in each boxplot corresponds to the median; box edges indicate the interquartile range, and whiskers extend to datapoints within 1.5 times the interquartile range. Results represent the mean values across ten Monte Carlo cross-validation (MCCV) runs, as described in the Methods.
Fig. 4 Minimal overfitting was observed across all model variants. “Overfitting index”, defined as the relative difference in performance between the LOGO-CV and the blind test set. Higher values indicate larger overfitting. Values above zero (dotted line) indicate better performance for the LOGO-CV than the blind test. The central line in each boxplot corresponds to the median; box edges indicate the interquartile range, and whiskers extend to datapoints within 1.5 times the interquartile range. Results represent the mean values across ten Monte Carlo cross-validation (MCCV) runs, as described in the Methods.CC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Research works printed but not linked: 51

  1. 10.1002/jps.10154 · DOI
  2. 10.1016/j.ejpb.2011.03.004 · DOI
  3. 10.1016/j.xphs.2023.03.026 · DOI
  4. 10.1080/19420862.2023.2285277 · DOI
  5. 10.1080/19420862.2023.2211185 · DOI
  6. 10.1016/j.xphs.2021.11.008 · DOI
  7. 10.1080/19420862.2023.2205540 · DOI
  8. 10.1016/j.ijpharm.2024.124877 · DOI
  9. 10.1038/nrd3581 · DOI
  10. 10.1016/j.molimm.2015.01.003 · DOI · Paper page
  11. 10.1038/s41573-023-00709-2 · DOI
  12. 10.1093/bib/bbae307 · DOI
  13. 10.1080/19420862.2023.2185924 · DOI
  14. 10.1016/j.sbi.2022.102343 · DOI
  15. 10.1080/19420862.2023.2171248 · DOI
  16. 10.1007/s43393-021-00024-0 · DOI
  17. 10.3390/ijms17060853 · DOI
  18. 10.1016/j.ejps.2023.106625 · DOI
  19. 10.1007/S11095-019-2703-7/FIGURES/8 · DOI
  20. 10.1021/acs.molpharmaceut.9b00545 · DOI
  21. 10.1080/19420862.2020.1854149 · DOI
  22. 10.1021/acs.molpharmaceut.4c00859 · DOI
  23. 10.1021/acs.molpharmaceut.1c00135 · DOI
  24. 10.3389/fmolb.2021.689400 · DOI
  25. 10.1016/j.xphs.2017.12.015 · DOI
  26. 10.1021/acscentsci.8b00913 · DOI
  27. 10.1021/acs.molpharmaceut.0c00629 · DOI
  28. 10.1021/acs.molpharmaceut.0c00775 · DOI
  29. 10.1080/19420862.2023.2212416 · DOI
  30. 10.26434/chemrxiv-2024-1zv1q · DOI
  31. 10.1016/j.xphs.2020.10.051 · DOI
  32. 10.1016/j.ejpb.2020.04.002 · DOI
  33. 10.1093/bib/bbad468 · DOI
  34. 10.1016/j.ijpharm.2023.122839 · DOI
  35. 10.1007/s11095-020-02919-4 · DOI
  36. 10.1016/j.xphs.2021.03.017 · DOI
  37. 10.1093/abt/tbab027 · DOI
  38. 10.1186/s12859-019-2669-9 · DOI
  39. 10.1002/pro.4497 · DOI · Paper page
  40. 10.1073/pnas.2016239118 · DOI · Paper page
  41. 10.1109/TPAMI.2021.3095381 · DOI · Paper page
  42. 10.1038/s41598-021-81132-8 · DOI
  43. 10.1093/bioinformatics/btp163 · DOI · Paper page
  44. 10.1007/s10994-006-6226-1 · DOI · Paper page
  45. 10.1016/j.cpc.2010.04.018 · DOI
  46. 10.1080/19420862.2023.2248671 · DOI
  47. 10.1613/jair.953 · DOI · Paper page
  48. 10.1093/bioinformatics/bts565 · DOI · Paper page
  49. 10.1093/bioinformatics/btl158 · DOI
  50. Lundberg SM , Lee SI . A unified approach to interpreting model predictions. In: Guyon I , Luxburg U V , Bengio S , et al. , eds. Advances in Neural Information Processing Systems 30. Curran Associates, Inc.; 2017:4765–4774. http://papers.nips.cc/paper/7062-a-unified-approach-to-interpreting-model-predictions.pdf
  51. 10.1023/A:1010933404324 · DOI · Paper page

Metadata

Authors
Vidal-Henriquez, E., Holder, T., Lee, N. F., Pompe, C., & Teese, M. G.
Affiliations
Leukocare AG, Klopferspitz 19a, 82152 Martinsried/Munich, Germany
Corresponding author
Teese, Mark George
ORCID
Venue
Computational and Structural Biotechnology Journal
Publisher
openRxiv
Published
2025-02-16
Volume
27
Pages
4517-4525
ISSN
2001-0370 (ISSN-L)
DOI
10.1101/2025.02.12.637685
PMC
PMC12597001
OpenAlex
W4415197028
Cited by
9
References
1
Citations by year
2026: 6 · 2025: 3
FWCI
2.71
Type
journal article
Language
en
Open access
Licence
Locations
  • Computational and Structural Biotechnology Journal (publishedVersion, cc-by-nc-nd)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Technische Universität München
  • University of Colorado Boulder
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein purification and stability
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Drug Discovery › Computational Drug Discovery Methods
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
Keywords
excipient selection, protein language model embeddings, protein structural properties

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review