Case-Based Reasoning for Tablet Formulation

Case-Based Reasoning is a recent Artificial Intelligence problem-solving methodology where new problems are solved by re-using solutions to similar problems in the past. We use this approach to address the formulation problem for tablets: given the dose of a particular new drug decide which excipients to use as filler, binder, disintegrant, lubricant and surfactant, and the quantity of each, so that a viable tablet can be manufactured [1]. The physical and chemical properties of the drug and excipients are provided: e.g., solubility, and stability with each excipient. The idea is to balance the physical properties of the drug and excipient, whilst ensuring that sufficient stability is attained. Although rules can be applied to suggest formulations for new drugs, it is difficult to acquire effective problem solving rules initially, and equally difficult to maintain these rules as different formulation practices evolve and more difficult drugs require to be formulated. This makes the corporate memory approach of Case-Based Reasoning (CBR) attractive [2]. As tablets for new drugs are formulated, the properties of the drug and excipients, together with the formulation that was successfully applied, are stored in a database. When a new drug is presented, the formulations for drugs with similar properties are retrieved from the database. Although a new drug with identical properties to an existing formulation is unlikely, drugs with similar properties are likely to have similar formulations – this is a fundamental assumption of CBR. The most similar retrieved formulations are re-used to suggest a new formulation; e.g. by choosing the most frequently occurring filler from the retrievals and taking its average quantity. When the new drug is sufficiently different from the retrieved formulations, an adaptation step amends the proposed formulation by exploiting these differences; e.g., a harder drug may need a softer filler. The knowledge acquisition overhead of rules is largely overcome in CBR because successful new formulations can be inserted into the database. Recently it has been noticed that knowledge sources other than the database are required, notably retrieval knowledge to determine similarity, and adaptation knowledge. In an EPSRC-funded project we are exploring ways to ease the knowledge acquisition for these other knowledge containers – we learn retrieval and adaptation knowledge from the formulation knowledge already captured in the database [3, 4]. This knowledge must also be refined when the database grows significantly or when a new formulation policy is introduced [5]. CBR is used to solve a wide range of problems: e.g., failure prediction from ultrasonic NDT of rails for Dutch railways, maintenance scheduling for Boeing 737 engines, mission planning for the US navy, customisation in e-commerce sales, personalised TV listings for satellite channels. However, the most similar application to tablet formulation comes from a very different domain – formulating rubber for specialised tyres in racing cars [6]! Further information about CBR for tablet information can be found at http://www.scms.rgu.ac.uk/research/kacbd/