Machine learning approaches to root cause analysis, characterization, and monitoring of subvisible particles in monoclonal antibody formulations

Abstract

Processing stresses on therapeutic proteins may cause formation of subvisible particles. Different stress mechanisms generate particle populations with characteristic morphological "fingerprints," and machine learning techniques like convolutional neural networks (CNNs) allow classification of microscopy images of these particles according to known stresses at their root cause. Using CNNs to classify novel particle types not included during network training may lead to inaccurate classification, however, using CNNs to monitor the presence of particulate matter not explicitly used in training could serve as a useful process analytical technology.We used CNNs to classify and identify the root cause of particles generated by subjecting three monoclonal antibodies (mAbs) to various common manufacturing stresses. We probed the generality of particles generated by stressing different mAbs in different formulations and showed that CNN analyses were sensitive not only to the applied stress, but also the buffer conditions and the particular mAb that generated particle populations. Thus, models trained on images of particles created with one mAb and buffer system may not provide accurate root cause analysis when applied to particles generated by other mAb and buffer systems. A lever-rule analysis of CNN-derived fingerprints was used to characterize the composition of mixtures of particle types. Finally, we monitored the temporal evolution of CNN-derived fingerprints when novel populations of particles, which were not included during training, were generated by pumping mAb solutions through a peristaltic pump.

Deep learning image analysis models pretrained on daily objects are useful for the preliminary characterization of particulate pharmaceutical samples

“…In such cases, other possible methods for dimensionality reduction like UMAP can be explored. Other alternatives may include the addition of a self‐supervised learning step to produce the 2D feature vector (Salami, Wang, et al, 2022), or applying a weak supervised fine‐tuning step using, for example, a triplet loss function (Greenblott et al, 2022).…”

“…One example is building supervised or self-supervised classifiers to partition particles in prefilled syringe-administered biotherapeutics into high-risk and lowrisk classes based on images captured by microflow imaging (MFI) (Wang et al, 2022). Other examples include building machine learning models that assign a protein aggregate to a certain forced degradation process (e.g., heat, pH, freeze-thawing, or mechanically induced), or produce particle fingerprints that highlight features of particles belonging to specific classes (Calderon et al, 2018(Calderon et al, , 2022Gambe-Gilbuena et al, 2020;Greenblott et al, 2022;Saggu et al, 2017).…”

Supervised and unsupervised machine learning approaches for monitoring subvisible particles within an aluminum‐salt adjuvanted vaccine formulation

“…The pumping protocol used in this work was adopted from Greenblott et al (2022). Briefly, samples of either drug substance and placebo were recirculated in a closed loop between the peristaltic pump and a 50 mL polypropylene Falcon tube (Sigma-Aldrich) collection vessel.…”

“…Analysis based only on the size of particles within a vaccine suspension discards other information about the system and introduces human bias from this selected feature (Maddux et al, 2017). Data‐driven, supervised machine learning techniques like convolutional neural networks (CNNs) and multivariate data analysis tools can be used to augment analyses of FIM images of particulate matter commonly found in protein formulations (Calderon et al, 2018; Calderon, Krhač Levačić, et al, 2022; Calderon, Ripple, et al, 2022; Daniels et al, 2020; Gambe‐Gilbuena et al, 2020; Greenblott et al, 2022; Shibata et al, 2022; Thite et al, 2022; Witeof et al, 2021). Recently, these methods were explored as a potential process analytical technology (PAT) tool for detecting and monitoring populations of novel particles that were not seen during the training of CNNs.…”

“…The CNN analyses used in these works demonstrated that different particle populations can possess unique distributions of morphological features or "fingerprints." Populations of particles were compared with Rosenblatt-transform-based goodness-of-fit hypothesis testing (Calderon, Ripple, et al, 2022;Daniels et al, 2020) or by differences between the fingerprint embedding medians (FEMs) between two fingerprints (Greenblott et al, 2022).…”