(W1030-02-11) Machine Learning Approaches to Analysis of Microscopy Images of Polysorbate Degradation Product Particles

Purpose: Nonionic surfactants such as polysorbate 20 (PS20) are common excipients used in therapeutic protein formulations to protect them against interfacial stresses.^1–4 However, PS20 may degrade through host-cell protein catalyzed hydrolysis of PS20.^1,3,5,6 Some of the products of this hydrolysis are free fatty acids (FFAs) that are poorly soluble in aqueous environments.⁷ Consequently, subvisible particles may form, compromising drug product quality.^3,8 FFA-derived subvisible particles are difficult to characterize and monitor due to their heterogenous nature.^1,3,9 In several works, subvisible particles relevant to biopharmaceutical formulations were analyzed by a combination of flow imaging microscopy (FIM) and machine learning techniques.^10–21 Herein, we demonstrate how machine learning techniques can distinguish between FFA-derived particles obtained by incubating a lipase (representative of a host-cell protein) with different grades of PS20. Follow up experiments showed that particles generated during hydrolysis reactions in the presence of protein and the lipase concentration can further affect the morphologies of FFA-derived particles. The results from this work demonstrate the need for careful consideration for methods to detect polysorbate degradation events due to their various degradation products.

Methods: Subvisible particles were generated by incubating lipase from Pseudomonas Cepacia (PCL) with high purity (HP) and super refined (SR) PS20 grades in the presence and absence of IVIg. HP and SR PS20 were chosen since they contain different distributions of fatty acids¹.  PCL was dissolved in phosphate buffered saline at concentrations of 0.4 and 0.05 mg/mL. Different concentrations were used to evaluate the effect that PCL concentration and consequent kinetics had on resulting particles. 0.1% w/v HP and SR PS20 were spiked into separate 1.5 mL polypropylene tubes containing stock solutions of PCL and were refrigerated at 4°C for 3 hours (0.4 mg/mL lipase) and 96 hours (0.05 mg/mL lipase), respectively. Some samples were incubated with 1 mg/mL intravenous immunoglobulin (IVIg) at t =0 hours. Digital images of the resulting particles were recorded with FIM after the specified incubation times. Particle standards composed of aqueous suspensions of ethylene tetrafluoroethylene (ETFE) were also imaged using FIM.²² Following the basic procedure described in Greenblott et al.¹², CNNs were trained on sets of FIM images and their associated labels to classify single images into one of the classes on which the CNNs were trained. The trained classification CNN models (with fixed parameters) were subsequently applied to test images (not seen by the networks during training) allowing prediction of the most likely hydrolysis formation conditions that had generated each individual particle. Confusion matrices were used to visualize how well CNNs classified single images of particles.

Results: Figure 1 shows representative FIM images of ETFE particles and of particles obtained from incubating PCL at 0.4 mg/mL with HP and SR PS20 for three hours. ETFE particles generally appeared distinct from PS20-derived particles. There were subtle morphological and optical feature differences between the two collections of FFA-derived particles, e.g., some HP-derived particles tended to be brighter or more branched than SR-derived particles. Figure 2a shows the classification results from training a CNN on images of ETFE, HP PS20, and SR PS20 derived particles and applying it to test images obtained form the same (vial 1) and different vials (2 and 3) from which training images were obtained. The classification accuracies were similar for all three replicates indicating that the hydrolysis reaction was repeatable and that the particles obtained from solutions containing each PS20 grade were distinct. When IVIg was incubated with PCL and PS20, the resulting particles were different as evidenced by the different classification accuracies shown in Figure 2b. This suggested that IVIg interacted with FFAs during their precipitation and particle formation to form particles that were distinct from those made in the absence of IVIg. The CNN used to generate Figure 2a was also applied to images of particles obtained by incubating lipase at a lower concentration with both PS20 grades as described in Methods, and the classifications results are shown in Figure 2c. Single images tended to be classified correctly with an accuracy of slightly greater than 50% suggesting that the reaction rate decrease caused by a lower enzyme concentration affected the generated particles. However, they were generally still distinguishable as either HP or SR-derived. When IVIg was added to the same samples obtained from the same starting conditions as Figure 2c, classification accuracies were different (Figure 2d) from samples in the absence of IVIg, further suggesting that IVIg interacted with FFAs as they formed particles.

Conclusion: We demonstrated the ability of CNN-based analyses to distinguish between FIM images of particles derived from the hydrolysis of different PS20 grades. Particles were also affected by the presence of protein or when the concentration of enzyme was changed. Ultimately, the results of this work will inform formulation scientists about the challenges associated with polysorbates and will contribute machine learning approaches to aid in characterizing polysorbate degradation products.

References: 1. Doshi N, Fish R, Padilla K, Yadav S. Evaluation of Super RefinedTM Polysorbate 20 With Respect to Polysorbate Degradation, Particle Formation and Protein Stability. J Pharm Sci. 2020;109(10):2986-2995. doi:10.1016/J.XPHS.2020.06.030
2. Glücklich N, Carle S, Diederichs T, Buske J, Mäder K, Garidel P. How enzymatic hydrolysis of polysorbate 20 influences colloidal protein stability. European Journal of Pharmaceutical Sciences. 2023;191. doi:10.1016/j.ejps.2023.106597
3. Zhang J, He J, Smith KJ. Fatty Acids Can Induce the Formation of Proteinaceous Particles in Monoclonal Antibody Formulations. J Pharm Sci. 2022;111(3):655-662. doi:10.1016/j.xphs.2021.10.008
4. Sreedhara A, Buecheler J, Brosig S, et al. End-to-End Approach to Surfactant Selection, Risk Mitigation, and Control Strategies for Protein-Based Therapeutics. AAPS Journal. 2023;25(1). doi:10.1208/s12248-022-00773-3
5. Yuk IH, Koulis T, Doshi N, et al. Formulation mitigations for particle formation induced by enzymatic hydrolysis of polysorbate 20 in protein-based drug products: insights from a full-factorial longitudinal study. Published online 2022. doi:10.1186/s41120-022-00064-3
6. Hall T, Sandefur SL, Frye CC, Tuley TL, Huang L. Polysorbates 20 and 80 Degradation by Group XV Lysosomal Phospholipase A2 Isomer X1 in Monoclonal Antibody Formulations. J Pharm Sci. 2016;105(5):1633-1642. doi:10.1016/J.XPHS.2016.02.022
7. Doshi N, Demeule B, Yadav S. Understanding Particle Formation: Solubility of Free Fatty Acids as Polysorbate 20 Degradation Byproducts in Therapeutic Monoclonal Antibody Formulations. Mol Pharm. 2015;12(11):3792-3804. doi:10.1021/acs.molpharmaceut.5b00310
8. Glücklich N, Dwivedi M, Carle S, Buske J, Mäder K, Garidel P. An in-depth examination of fatty acid solubility limits in biotherapeutic protein formulations containing polysorbate 20 and polysorbate 80. Int J Pharm. 2020;591. doi:10.1016/j.ijpharm.2020.119934
9. Saggu M, Demeule B, Jiang L, et al. Extended Characterization and Impact of Visible Fatty Acid Particles - A Case Study With a mAb Product. J Pharm Sci. 2021;110(3):1093-1102. doi:10.1016/j.xphs.2020.11.027
10. Daniels AL, Calderon CP, Randolph TW. Machine learning and statistical analyses for extracting and characterizing “fingerprints” of antibody aggregation at container interfaces from flow microscopy images. Biotechnol Bioeng. 2020;117(11):3322-3335. doi:10.1002/bit.27501
11. Calderon CP, Daniels AL, Randolph TW. Deep Convolutional Neural Network Analysis of Flow Imaging Microscopy Data to Classify Subvisible Particles in Protein Formulations. J Pharm Sci. 2018;107(4):999-1008. doi:10.1016/j.xphs.2017.12.008
12. Greenblott DN, Zhang J, Calderon CP, Randolph TW. Machine Learning Approaches to Root Cause Analysis, Characterization, and Monitoring of Subvisible Particles in Monoclonal Antibody Formulations. Biotechnol Bioeng. Published online 2022. doi:10.1002/bit.28239
13. Thite NG, Ghazvini S, Wallace N, Feldman N, Calderon CP, Randolph TW. Machine learning analysis provides insight into mechanisms of protein particle formation inside containers during mechanical agitation. J Pharm Sci. Published online July 2022. doi:10.1016/j.xphs.2022.06.017
14. Calderon CP, Ripple DC, Srinivasan C, et al. Testing Precision Limits of Neural Network-Based Quality Control Metrics in High-Throughput Digital Microscopy. 2022;1:3. doi:10.1007/s11095-021-03130-9
15. Witeof AE, Daniels AL, Rea LT, et al. Machine Learning and Accelerated Stress Approaches to Differentiate Potential Causes of Aggregation in Polyclonal Antibody Formulations During Shipping. J Pharm Sci. 2021;110(7):2743-2752. doi:10.1016/J.XPHS.2021.02.029
16. Gambe-Gilbuena A, Shibano Y, Krayukhina E, Torisu T, Uchiyama S. Automatic Identification of the Stress Sources of Protein Aggregates Using Flow Imaging Microscopy Images. J Pharm Sci. 2020;109(1):614-623. doi:10.1016/J.XPHS.2019.10.034
17. Shibata H, Terabe M, Shibano Y, et al. A Collaborative Study on the Classification of Silicone Oil Droplets and Protein Particles Using Flow Imaging Method. J Pharm Sci. 2022;111(10):2745-2757. doi:10.1016/j.xphs.2022.07.006
18. Thite NG, Ghazvini S, Wallace N, Feldman N, Calderon CP, Randolph TW. Interfacial Adsorption Controls Particle Formation in Antibody Formulations Subjected to Extensional Flows and Hydrodynamic Shear. J Pharm Sci. Published online July 2023. doi:10.1016/j.xphs.2023.07.010
19. Salami H, Wood C, Ouyang H, Zhao X, Skomski D. Deep learning image analysis models pretrained on daily objects are useful for the preliminary characterization of particulate pharmaceutical samples. Biotechnol Bioeng. Published online July 12, 2023. doi:10.1002/bit.28488
20. Salami H, Wang S, Skomski D. Evaluation of a Self-Supervised Machine Learning Method for Screening of Particulate Samples: A Case Study in Liquid Formulations. J Pharm Sci. Published online October 12, 2022. doi:10.1016/J.XPHS.2022.10.010
21. Fedorowicz FM, Chalus P, Kirschenbühler K, Drewes S, Koulov A. Image classification of degraded polysorbate, protein and silicone oil sub-visible particles detected by flow-imaging microscopy in biopharmaceuticals using a convolutional neural network model. J Pharm Sci. Published online 2023. doi:10.1016/j.xphs.2023.07.003
22. Telikepalli S, Gonzalez K, Dragulin-Otto S, Ripple D, Carrier M, Khan M. Development of Protein-Like Reference Material for Semiquantitatively Monitoring Visible Proteinaceous Particles in Biopharmaceuticals. PDA J Pharm Sci Technol. 2019;73(5):418-432. doi:10.5731/pdajpst.2018.008953

Acknowledgements: Funding for this work was provided by the University of Colorado, Boulder.

Figure 1. Representative FIM images of ETFE and of particles obtained from incubating a lipase with HP and SR PS20 for three hours.

Figure 2. A CNN classifier was trained on images of ETFE, HP PS20, and SR PS20 particles obtained from incubation with PCL at 0.4 mg/mL for three hours. The CNN was applied to test images of PCL at 0.4 mg/mL incubated with HP or SR PS20 without (a) or with 1 mg/mL IVIg spiked into the solution at t = 0 hours. FIM images were recorded after incubation at 4 °C for three hours. The CNN was also applied to test images obtained from a solution of PCL at 0.05 mg/mL incubated with HP or SR PS20 without (c) or with (d) 1 mg/mL IVIg spiked into the solution at t = 0 hours. FIM images were recorded after incubation at 4 °C for 48 hours. Rows of each confusion matrix represent separate vials.