Poor generalizability is a major barrier to clinical implementation of artificial intelligence in digital pathology. The aim of this study was to test the generalizability of a pretrained deep learning model to a new diagnostic setting and to a small change in surgical indication. A deep learning model for breast cancer metastases detection in sentinel lymph nodes, trained on CAMELYON multicenter data, was used as a base model, and achieved an AUC of 0.969 (95% CI 0.926–0.998) and FROC of 0.838 (95% CI 0.757–0.913) on CAMELYON16 test data. On local sentinel node data, the base model performance dropped to AUC 0.929 (95% CI 0.800–0.998) and FROC 0.744 (95% CI 0.566–0.912). On data with a change in surgical indication (axillary dissections) the base model performance indicated an even larger drop with a FROC of 0.503 (95%CI 0.201–0.911). The model was retrained with addition of local data, resulting in about a 4% increase for both AUC and FROC for sentinel nodes, and an increase of 11% in AUC and 49% in FROC for axillary nodes. Pathologist qualitative evaluation of the retrained model´s output showed no missed positive slides. False positives, false negatives and one previously undetected micro-metastasis were observed. The study highlights the generalization challenge even when using a multicenter trained model, and that a small change in indication can considerably impact the model´s performance.

Monoclonal antibodies (mAbs) constitute a rapidly growing biopharmaceutical sector. However, their growth is impeded by developability issues such as polyspecificity and lack of solubility, which leads to attrition as well as manufacturing failures. In this study a multitool hybrid quantitative structure–activity relationship (QSAR) model development framework is described. This framework uses four novel datasets derived from the primary sequences of IgG1‐κ‐humanized mAbs with varying degrees of resolutions. Unsupervised pattern recognition is first performed on the descriptor sets to visualize any intrinsic property‐based clustering, followed by regression of descriptors against cross‐interaction chromatography (CIC) retention times. Model optimization is performed via unsupervised variable reduction followed by supervised variable selection. Finally, the models and datasets are benchmarked based on the regression model performance metrics such as R2, Q2, and RMSE. The results show that datasets containing localized descriptors rather than averaged value over the entire protein have better predictive performance of CIC retention behavior with R2 > 0.8 and RMSE < 0.3. Furthermore, the results indicate the physicochemical, electronic, and topological properties of hypervariable regions of antibodies that contribute most to the CIC retention times. The results of these studies could contribute to early‐stage screening and better design of mAbs.

Monoclonal antibodies (mAbs) constitute a rapidly growing biopharmaceutical sector. However, their growth is impeded by high failure rates originating from failed clinical trials and developability issues in process development. There is, therefore, a growing need for better in silico tools to aid in risk assessment of mAb candidates to promote early-stage screening of potentially problematic mAb candidates. In this study, a quantitative structure–activity relationship (QSAR) modelling workflow was designed for the prediction of hydrophobic interaction chromatography (HIC) retention times of mAbs. Three novel descriptor sets derived from primary sequence, homology modelling, and atomistic molecular dynamics (MD) simulations were developed and assessed to determine the necessary level of structural resolution needed to accurately capture the relationship between mAb structures and HIC retention times. The results showed that descriptors derived from 3D structures obtained after MD simulations were the most suitable for HIC retention time prediction with a R2 = 0.63 in an external test set. It was found that when using homology modelling, the resulting 3D structures became biased towards the used structural template. Performing an MD simulation therefore proved to be a necessary post-processing step for the mAb structures in order to relax the structures and allow them to attain a more natural conformation. Based on the results, the proposed workflow in this paper could therefore potentially contribute to aid in risk assessment of mAb candidates in early development.

In today's biopharmaceutical industries, the lead time to develop and produce a new monoclonal antibody takes years before it can be launched commercially. The reasons lie in the complexity of the monoclonal antibodies and the need for high product quality to ensure clinical safety which has a significant impact on the process development time. Frameworks such as quality by design are becoming widely used by the pharmaceutical industries as they introduce a systematic approach for building quality into the product. However, full implementation of quality by design has still not been achieved due to attrition mainly from limited risk assessment of product properties as well as the large number of process factors affecting product quality that needs to be investigated during the process development. This has introduced a need for better methods and tools that can be used for early risk assessment and predictions of critical product properties and process factors to enhance process development and reduce costs. In this review, we investigate how the quantitative structure-activity relationships framework can be applied to an existing process development framework such as quality by design in order to increase product understanding based on the protein structure of monoclonal antibodies. Compared to quality by design, where the effect of process parameters on the drug product are explored, quantitative structure-activity relationships gives a reversed perspective which investigates how the protein structure can affect the performance in different unit operations. This provides valuable information that can be used during the early process development of new drug products where limited process understanding is available. Thus, quantitative structure-activity relationships methodology is explored and explained in detail and we investigate the means of directly linking the structural properties of monoclonal antibodies to process data. The resulting information as a decision tool can help to enhance the risk assessment to better aid process development and thereby overcome some of the limitations and challenges present in QbD implementation today.