A data-driven robotic platform for autonomous formulation of antibody therapeutics

Helena Ros, Christian Wells, Geoff Platt, Paul Dalby, David Shorthouse, Michael T. Cook. UCL School of Pharmacy, London, WC1N 1AX. UCL Department of Biochemical Engineering, London, WC1E 6BT. Croda, UK

Switching monoclonal antibody therapies from intravenous to subcutaneous administration improves patient autonomy and eases healthcare costs, but remains a challenge due to high viscosity of mAbs in the high-concentration, low-volume formulations needed for this switch. To address this a high-throughput robotic workflow for screening protein formulations at dilute concentrations as predictors of viscosity at high concentrations was developed. 16 formulations with 6 varying excipient concentrations were dispensed using an Opentron’s OT-2 liquid handling robot and A33 Fab fragment added as a representative protein. Dilute formulations were analysed by dynamic light scattering, intrinsic protein fluorescence, and thermal shift assays to assess protein aggregation and stability. A33 protein was then concentrated to 100 mg/mL and analysed using rheology, allowing correlation between dilute regime endpoints and rheology in the high concentration regime. This study presents steps towards an automated workflow for formulation of high-concentrated antibody therapeutics. Future work will integrate Bayesian optimisation to allow iterative, data-driven optimisation of formulations.

 

Decoding Tactile Perception: Bridging Physical Measurements into Sensory Experience for Accelerating Formulation Development.

Ferdaous EL BENNI1 ; Magali Bonnier2 ; Uyai Ikpatt1. 1R&T, Croda, Liverpool, United Kingdom; 2R&T, Croda, Paris, France

Tactile perception of skincare formulations, including spreadability, thickness and stickiness, arises from dynamic interactions between formulation composition, skin surface and finger motion. These interactions involve physical parameters such as friction, pressure, vibration and shear, which stimulate mechanoreceptors and contribute the sensory response. Relating these measurable signals to perception provides a complementary route for characterising formulations beyond in-vivo sensory testing. 

Sixteen formulations were selected as reference materials to represent various colloidal structures, including aqueous gels, emulsions, oils, and petrolatum to cover the sensory scale. Instrumental characterisation combined complementary techniques to capture the mechanical and tactile behaviour of formulations. The ForcePlate and tribometer were jointly employed to quantify vibration, friction, and lubrication under controlled load conditions. Rheological measurements included both flow protocols and LAOS tests, providing information on flow behaviour and viscoelasticity. Texture analysis was conducted through compression and stickiness tests, offering additional parameters related to tactile behaviour.

Several machine learning algorithms were tested, including linear and non-linear models. Gaussian Process Regression provided the most consistent performance across validation and test data, with a mean absolute error below 10. To evaluate robustness, models were trained using 63 input configurations combining instruments and protocols. This reflects laboratory diversity and practical constraints, confirming acceptable predictions can be achieved even with a single instrument. Validation on seven formulations with closely related sensory profiles confirmed the accuracy of the approach. 

This framework enables predictive modelling of tactile attributes and has been implemented into a digital interface, offering formulators an innovative practical tool for guiding formulation design. 

 

 

 

The role of data analysis in new process analysis technologies 

Patricia Blanco-García, Pilar Gómez Jiménez, Chandresh Malde, George Platt.  Johnson Matthey Technology Centre, UK  

This talk will discuss some of characterisation methods used in Johnson Matthey’s product portfolio at the research and manufacturing scales, and how the use of statistical tools in data analysis combined with in-process measurements can advance pass/fail criteria and decision making towards future formulations. 

It is often the case that in industry formulation and product characterisation occur at different scales. In R&D, there is greater emphasis on extracting numerous fundamental properties of the formulation or finished product from a wide range of characterisation techniques. In manufacturing process, it is common practice to reduce the number of measurement metrics to qualify and validate a given formulation/product. 

There is a global emphasis for industry to move towards sustainable manufacturing. For most industries this will require considerable momentum shift in capital expenditure to modernise existing manufacturing processes in terms of process efficiency (e.g., powder transport, slurry mixing, milling, drying etc), in-line analytics, and recycling. Automation, statistical data analysis, and data management will play a major role for the improvements necessary in manufacturing processes. 

Characterisation metrics that define pass/fail criteria can be based on historical agreements, single value data from complex distributions and/or non-linear profiles and dated instrumentation. This presentation will discuss the use of statistical tools in data analysis combined with in-process measurements to advance pass/fail criteria and decision making towards future formulations. 

 

 

Integrating Mechanistic and Data-Driven Modelling for Optimising Probiotic Tablet Formulation

Bide Wang¹, Xilu Wang², Oleksiy V. Klymenko¹, Rachael Gibson³, Andrew Middleton³ & Chuan-Yu Wu^1. 1 School of Chemistry and Chemical Engineering, University of Surrey, Guildford, GU2 7XH, UK. 2 School of Computer Science and Electronic Engineering, University of Surrey, Guildford, GU2 7XH, UK.  3 P&G Innovation Centre, Reading, Berkshire, RG2 0QE, UK.

Tablet is a popular dosage form for probiotic delivery, offering accurate dosing, ease of administration, and scalability for industrial production. However, probiotics are highly sensitive to mechanical and thermal stresses during powder compaction, which often leads to significant loss of viability and diminished clinical efficacy. To address this challenge, a finite element (FE) model was developed to predict probiotic survival during tabletting, integrating a Drucker–Prager Cap (DPC) compaction model with the thermal tolerance characteristics of probiotics. The model was rigorously validated against experimental data. Building on this foundation, Gaussian Process Regression (GPR) coupled with active learning (AL) was then employed to construct a surrogate model capable of accurately predicting probiotic viability using a limited dataset, achieving 95% predictive accuracy. Beyond viability, tablet porosity represents another critical quality attribute. To reconcile these competing objectives, multi-objective optimisation using NSGA-II was applied to identify processing conditions that simultaneously maximise probiotic survival and optimise tablet porosity. This integrated modelling and optimisation framework provides both a scientific basis and practical guidance for the rational design of robust probiotic solid dosage forms.

 

 

The Cambridge Structural Database meets the Manufacturing Classification System

Andy Maloney, The Cambridge Crystallographic Data Centre

The physical and chemical properties of a new chemical entity, either as the active pharmaceutical ingredient or the drug product, are strongly influenced by the particle attributes and the underlying crystallisation process defined during development. Product performance can only be assured when the active ingredient is delivered to the patient in a chemically and physically stable particle. A structural understanding in this key area allows pharmaceutical organisations to support the selection of the right molecule from discovery, design of the right product during development, and optimisation of the manufacturing process in production.

This presentation will highlight the crystallographic data thread through development by linking the Manufacturing Classification System (MCS) back into the Cambridge Structural Database. Digital design tools are used to predict particle shape and particle properties (such as surface chemistry) before analysing this property data across the three MCS categories (wet granulation, dry granulation, direct compression) to identify trends. This demonstrates the potential for Particle Informatics modelling tools and data analytics to connect across the chemical, analytical and formulation disciplines providing a cornerstone for the accelerated development and manufacture of novel advanced pharmaceutical products.

 

 

 

 

 

 

 

 Mischa Zelzer. University of Nottingham.

 

There is significant potential for artificial intelligence approaches to contribute to the acceleration of drug formulation design and development. To implement this, the relative importance of physicochemical and methodological parameters of the processes have to be understood. In my group, we study the effect of physicochemical properties of drugs and formulation components on various aspects of the performance of drug delivery modalities. These include drug loading and drug release in lipid-based formulations as well as drug release from microneedles and solid dosage forms. Using our own experimental data as well as data from the literature, we use machine learning and model explainability tools to tackle challenges such as making formulation data AI ready and to determine relative parameter importance affecting drug loading and release.

 

A semi-automated method for accelerated prediction of new micellar nanomedicines

Antonia Gucic^a*, Chris Waudby^a, David Shorthouse^a, Michael T. Cook^a. aUCL School of Pharmacy, 29-39 Brunswick Square, London, WC1N 1AX, UK

Approximately 70% of currently available active pharmaceutical ingredients (APIs) are considered poorly water soluble, which often results in significant challenges in drug development and formulation (Moreton, 2021). 

Surfactant micellar systems have been extensively studied (Williams et al., 2013) for their ability to increase the amount of dissolved API (Lu et al., 2021), however less than 3% of marketed nanomedicines are micellar (Thapa et al., 2023). Additionally, the conventional trial-and-error methods of designing solubility-enhancing formulations are laborious and time-consuming (Noh et al., 2024).
Subsequently, a semi-automated saturation shake-flask solubility method (Baka et al., 2008) alternative has been developed with the view of exploiting machine learning (ML) to inform more efficient, automated formulation of micellar systems.

A method for semi-automated high-throughput small-volume equilibrium solubility studies in 96 well plates was developed using an Opentrons OT2 liquid handling robot (OT2) and plate readers commonly found in pharmaceutical laboratories (i.e. ClarioStar UV-Vis and DynaPro DLS plate readers). Curcumin (CUR) was employed as a model hydrophobic nutraceutical with polysorbate 80 (PS80) and polysorbate 20 (PS20) chosen as the model FDA-approved pharmaceutical surfactants. Following automated formulation of single-surfactant micellar formulations in well plates, the concentration of solubilised CUR was determined spectrophotometrically, whereas size and polydispersity of micellar particles were analysed using dynamic light scattering.

A semi-automated method for high-throughput preparation of micellar systems has been developed, and it will be further utilised to examine solubilisation power of a library of surfactants against a library of poorly soluble drugs, with the view of investigating multi-surfactant systems, and curating an extensive data collection for development of an ML model for prediction of novel micellar formulations.
This work has been funded by the EPSRC-SFI CDT in Transformative Pharmaceutical Technologies (grant code EP/S023054/1).

 

A data-driven robotic platform for autonomous formulation of antibody therapeutics

Helena Ros, Christian Wells, Geoff Platt, Paul Dalby, David Shorthouse, Michael T. Cook. UCL School of Pharmacy, London, WC1N 1AX.  UCL Department of Biochemical Engineering, London, WC1E 6BT.  Croda, UK

Switching monoclonal antibody therapies from intravenous to subcutaneous administration improves patient autonomy and eases healthcare costs, but remains a challenge due to high viscosity of mAbs in the high-concentration, low-volume formulations needed for this switch. To address this a high-throughput robotic workflow for screening protein formulations at dilute concentrations as predictors of viscosity at high concentrations was developed. 16 formulations with 6 varying excipient concentrations were dispensed using an Opentron’s OT-2 liquid handling robot and A33 Fab fragment added as a representative protein. Dilute formulations were analysed by dynamic light scattering, intrinsic protein fluorescence, and thermal shift assays to assess protein aggregation and stability. A33 protein was then concentrated to 100 mg/mL and analysed using rheology, allowing correlation between dilute regime endpoints and rheology in the high concentration regime. This study presents steps towards an automated workflow for formulation of high-concentrated antibody therapeutics. Future work will integrate Bayesian optimisation to allow iterative, data-driven optimisation of formulations.

 

 

 

 

 

 

 Organising Committee