In today's rapidly evolving technological landscape, the integration of advanced technologies such as large language models (LLMs) and generative AI has the potential to revolutionize laboratory analysis. We present a novel approach to enhancing laboratory information management systems (LIMS) through the integration of a knowledge-driven chatbot powered by LLMs. By leveraging internal databases, the chatbot is designed to provide specialized and tailored assistance to scientists, streamlining their workflow and improving overall efficiency. The chatbot utilizes the power of LLMs to understand user queries, retrieve relevant information from the database, and generate informative responses in a conversational manner.

Substantial data analysis is routinely performed to support critical decisions and filings. The process from analysis to reporting results requires time-consuming, repetitive manual effort. We will share our insights on scripted add-ins developed that take this process from hours to seconds. These tools can be distributed to empower scientists, increase development speeds, reduce human error, and streamline workflows, allowing us to focus on more strategic and innovative activities.

Machine Learning (ML) has shown promise in drug discovery and manufacturing, but its effective utilization faces hurdles across ML lifecycle including traceability, meeting regulatory requirements, siloing of related tools and their interoperability, model understanding, cross-organization collaboration, and dataset and model reuse. To address these hurdles, this presentation introduces the Machine Learning Lifecycle Ontology (MLLO), a standardized framework to capture ML metadata throughout the lifecycle. A MLLO-based prototype tool, ML Lifecycle Explorer, will be demonstrated. The talk concludes with a future work which will explore an additional MLLO potential to enhance model development and reuse by connecting it to domain knowledge.

Protein characterization with mass spectrometry involves a series of tools. Most labs have Orbitraps and QTOFs. Methods include reduced and non-reduced peptide maps, intact and subunit mass, as well as native MS methods such as native SEC and native CEX. It is desirable to have a cross-platform data processing suite that can open data files from all instrument manufacturers as well as specialized in-house software tools.  A set of those tools will be presented.

The occurrence of subvisible particles (SVPs) in monoclonal antibody (mAb) development presents challenges in assessing product stability. This study utilizes quartz crystal microbalance with dissipation (QCM-D) in silico and experimental physicochemical properties to investigate SVP formation risks associated with various containers and stress types. MAb adsorption kinetics were found to strongly correlate with SVP propensity in the stirring study, and in silico predictors significantly improved all model performance.

High-quality datasets are key to successful biologics design and pre-developability prediction. Automated collection and curation of pre-developability data from electronic lab notebooks enables rapid retraining and testing of predictors, and it has also been used to direct model-focused data-collection efforts. Furthermore, we use internal data to augment and validate models trained on publicly available datasets, increasing the impact of these models by improving their predictivity on internal assays.

The development of biologic therapeutics is an exacting process, from discovery of a protein with desired function to optimizing for developability. To accelerate this process, we develop a modular multi-objective optimization method that synergistically harnesses deep learning and statistical models combined with structure-based and data-driven rational design. We apply this method to engineer immunoglobulin degrading proteases as potential therapeutics with desired target specificity and potency, low immunogenicity, and favorable manufacturability.

Explore the benefits of being a hybrid scientist in my journey from bench scientist to digital practitioner. Discover how digital skills empower the creation of quick, effective solutions, reducing reliance on external experts and manual processes. Learn about the future impact these tools have on portfolio progression through regulatory filings. Understand the importance of fostering digital development opportunities for future scientists.

Bioprocess Development groups face challenges with complex modalities, faster development cycles, and more experimental data from HT and PAT technologies. Historically, lab experimental data is dispersed in many different instrument software and unstructured files formats requiring substantial manual data manipulation efforts for experimental insights, decision making, process modeling, and tech transfer. Here we will share our journey to build and deploy a fit-for-science digital ecosystem within our bioprocess development labs.

We measured the viscosity of a large panel of 229 mAbs to develop predictive models for high-concentration screening. We developed DeepViscosity, consisting of 102 ensemble models to classify low-viscosity and high-viscosity mAbs at 150 mg/mL, using a sequence-based DeepSP model. Two independent test sets, comprising 16 and 38 mAbs, were used to assess DeepViscosity’s generalizability. The model exhibited an accuracy of 87.5% and 89.5% on both test sets, respectively.

This study investigated the metabolism of Chinese Hamster Ovary (CHO) cells through a multi-omics approach, integrating proteomic and metabolomic datasets. By developing genome-scale models, metabolic patterns were successfully identified and analyzed, leading to new hypotheses about the regulatory mechanisms driving cellular behavior. These findings provided insights that enhance the understanding of CHO cell metabolism and may improve strategies for optimizing protein bioproduction.