These reflections are from trends of knowledge share at AAPS.
There is no doubt that AI and LLMs are changing the way we show up in Pharma. An industry that has been the slowest adopter and highly gatekept with knowledge and information is now having to change its ways faster than ever with Drug Discovery to have the first movers advantage.
Small pockets of independent LLM and AI inspired models are being deployed and explored at Big Pharma while small biotech have also started to equip themselves with Software solutions and workforce that has computational skill set to be able to cope with the quick technological changes that are now creating ripples in experimentations to find the truth in pharma industry.
There is also a recognition in the industry on the quality of dataset that are available and the bias that these datasets have. So there have been discussions to move towards open datasets and to explore the loss of information with the bias of the dataset that are created by using LLM algorithms to solve for the same issues in drug discovery.
Many successful case studies have been highlighted in the area of Automation using GenAI:
There is no doubt that AI and LLMs are changing the way we show up in Pharma. An industry that has been the slowest adopter and highly gatekept with knowledge and information is now having to change its ways faster than ever with Drug Discovery to have the first movers advantage.
Small pockets of independent LLM and AI inspired models are being deployed and explored at Big Pharma while small biotech have also started to equip themselves with Software solutions and workforce that has computational skill set to be able to cope with the quick technological changes that are now creating ripples in experimentations to find the truth in pharma industry.
There is also a recognition in the industry on the quality of dataset that are available and the bias that these datasets have. So there have been discussions to move towards open datasets and to explore the loss of information with the bias of the dataset that are created by using LLM algorithms to solve for the same issues in drug discovery.
Many successful case studies have been highlighted in the area of Automation using GenAI:
- US-FDA leverage GPT-4 to do food effect summarization of reports using multi-turn interaction. The alogrithm is keyword-focused and used length-controlled prompts to refine the quality of generated summaries. They have conducted extensive evaluations, ranging from automated metrics to FDA professionals, on 100 NDA review documents (retrieved from Drugs@FDA) selected over the past five years. And they have shown to observe the quality of summarization improve. Their work shows how LLMs are now partners of FDA professionals for PSG assessment cycle and promote generic drug development.
- BMS and Regeneron talked about using LLMs and robotics for bio analysis specially to help bridge the non-clinical and clinical analysis gaps and time lines by thinking about continuous development and adaptation of analytical strategies. The ideas of creating decision trees for several challenges that arise at the early stages of a program, including limited information, restricted access to necessary reagents or platforms, unknown regulatory expectations, complex biology, and organizational structures where biomarker support activities are separate from bioanalysis groups and bridging the gap between the Silos to help create strong information and indicate use of automation for complex decision making was discussed and highlighted.
- Expert systems Inc shared their hypothesis and platform building for using LLMs in the context of drug reporposing. Specifically, they explored the potential of multitask models, which have recently demonstrated significant promise in predicting multiple drug-related properties. They showed the evaluation of combined multitask models with Combinatorial Fusion Analysis (CFA), a technique that aggregates and refines predictive outputs to improve accuracy and robustness.Their investigation into integrating these methodologies has yielded promising results, suggesting that the combined LLM / ML approach could offer a powerful new platform, suitable for optimizing drug candidates across multiple parameters and for addressing novel approved drug-disease combinations. They showcased their platform with case studies in oncology target identification and drug repurposing evaluation.
- Discussions on integrating empirical AI models with mechanistic models to make data work for the realms of preclinical, clinical and regulatory realm was discussed to highlight that gaps of LLM applications in biology while acknowledging that more experimentations with their applications as tools can carve frameworks for discovery.
- Many companies are realizing the potential of AI and LLMs for virtual twins as virtual populations alongside for virtual biomarkers or virtual organ on chips.
With all of this, the future of healthcare definitely looks like there will be many specialized language models in the market - and that the landscape and definition of proprietary solutions of software are rapid evolving. The democratization of knowledge in Pharma is also leading to opening of integration of proteomics, spacial genomics, transcriptomics to reduce the gap on the attrition of drugs coming to market.