1209 Feasibility of AI based framework for new product assessment in reproductive cancers

<h3>Background</h3> Pharmaceutical new product development and commercial strategy teams aim to estimate, if not predict, the relative efficacy of various competing drugs used in treatment for specific indications or patient segments affected by a disease. AI tools can be harnessed for a rapid and more comprehensive analysis and prediction of commercial markets. <h3>Methods</h3> This paper leverages AI tools to extract data from clinical trials, regulatory repositories, and scientific publications for Reproductive cancers for women. The paper describes an encoder-decoder model for a knowledge graph based contextual model for data extraction and validates the specificity and selectivity of the extracted data. Furthermore, this paper develops a set of non-generative AI based models to assess the key predictive parameters (if any) for performance across these indications by evaluating the past studies and then employing a model for projecting the relative probability of clinical success for future studies. For evaluation, the model considers several factors such as performance by mechanism of action, route of administration, pharmacokinetic profile, the cell biology of the targets, and the sequence of indications for which these drugs were developed. <h3>Results</h3> For reproductive cancers’ data we show that the accuracy of extraction from clinical and regulatory related data by our model is higher than 95% with 95% confidence while for toxicity, safety and pharmacokinetic and outcome related data, accuracy is higher than 90% with 90% confidence. Our early results suggest that success of completion of trials is, at best, a weak indicator, with a few drugs performing well (from an approval as well as a revenue perspective) despite having lower completion success in early studies than others. This is especially true with drugs sponsored by larger companies, leveraging their extensive experience and institutional knowledge base to fine-tune their models and drive success. This is also reflected in the overall success in approval by larger companies. This weighs the success of solutions employed by larger companies. Initial results indicate the combination of ADME parameters, protein binding and the success of the given Mechanism of Action when observed in a similar cancer, may be an early indicator of success as defined above. <h3>Conclusions</h3> AI models are effective in extracting a large amount of data and building new product analysis framework¹- suggesting that such frameworks may change by therapeutic area but are worthy tools in strategy planning for new product development. <h3>Reference</h3> Gartlehner G, Kahwati L. Data extraction for evidence synthesis using a large language model: a proof-of-concept study. <i>Res Synth Meth</i>. 2024:1–24.