Artificial Intelligence, particularly Generative AI, is becoming common vernacular and starting to play a role in R&D plans, however adoption of generative AI has been slow in life sciences so far. When deployed thoughtfully, AI can accelerate scientific discovery in several areas of life sciences, including drug discovery and development. But getting the most out of AI requires robust infrastructure, seamless integration into research workflows and a cultural shift among researchers and clinicians to view AI as an essential component of their toolkit rather than some futuristic concept.
It is important to distinguish between predictive and generative AI. Bioinformaticists have been training predictive AI for more than 20 years, and it could be argued that informaticists in the biopharma world created much of modern AI. Generative AI, however, especially in life sciences, is not the same as the predictive models that have become a standard in the industry.
Obstacles to wider adoption of generative AI in this field include researchers’ mistrust of technology vendors, IT mandates from institutional leadership and a real concern for the potential regulatory impact of hallucinations. Plus, AI is opening up a new world of collaboration across multiple teams, that likely feels a little foreign to researchers
“The way it used to work is you would discover a target and then you give multiple research organizations within the same company that target and tell them to go figure it out,” said Alexander Long, Head of Life Sciences Sales Strategy at Dell Technologies, which is providing critical high-performance servers, storage and a roster of software partners to support generative AI in life sciences.
According to Long, there are three truths about biopharma researchers: They generally distrust IT, they don't know the cost of anything and they love sharing their work—after it has been published. IT-enabled collaboration to moderate costs is antithetical this mentality. AI has to be user-friendly and it also has to yield good results in terms of helping researchers meet endpoints but also in avoiding “hallucinations” that can lead to false positives and false negatives.
IT is not the enemy here; IT should be the friend of researchers. “All of the pharma industry is going through a big transformation of the old way of competitive analytics to collaborative engagement,” Long said.
Another barrier is the fact that biopharma is a heavily regulated industry that requires compliance and scientific rigor, plus pharma companies closely guard their intellectual property.
Large language models (LLMs) are the foundation of generative AI, and well-known LLMs like OpenAI’s GPT-4, Google’s Bard and the open-source Llama rely on public datasets. That is a nonstarter for drug development, so pharma companies and their academic partners have been deploying private LLMs to encourage collaboration in a secure environment.
Researchers are sitting on huge quantities of data that are completely unexplored right now. By licensing AI applications and training them on private LLMs, then extracting results from their own data, biopharma investigators are able to develop new insights and, hopefully, stop rediscovering some ideas they have seen before. Reconfirming previous discoveries is a costly, inefficient part of current drug discovery processes.
An LLM also makes it easy to consider clinically adjacent information not captured in electronic medical records, such as telemetry data, environmental information and genomics. Long-read whole genomes, spatial omics, cryogenic electronic microscopy, cryogenic electron tomography and other advanced technologies are adding exponentially to research datasets “It’s a deluge,” Long said.
As whole-genome and exome sequencing becomes more commonplace, AI can speed up the addition of genomic data to the LLM to cross-reference gene expression and protein data in search of new targets, Long explained. AI might also be useful in the firmware of imaging and sequencing instruments to produce better raw reads.
Dell Technologies has been teaming with Northwestern Medicine to develop a generative multimodal LLM to help interpret chest X-rays in an emergency department where radiologists are not always readily available. In a paper published last year in the journal Emergency Medicine, the Northwestern Medicine team described how the model was just as accurate as human interpretation, and the results fit into existing workflows.
Northwestern Medicine built the model in-house with Dell Technologies infrastructure. Rather than revolt, radiologists embraced the model because they shared the benefits of increased efficiency, allowing them to spend more time on the interesting cases; in short, they gained from the deployment.
As for the ongoing fear many face, AI has the potential to automate many tasks that humans currently perform, but AI is not a creative thinker. It is outstanding at finding correlations but poor at understanding causations. In other words, it will not replace researchers, but researchers who use AI will replace those who don't. By leveraging AI, scientists and researchers will get to spend more time on the interesting things they find, not the what of the problem, but the why of the solution.
We live in a fundamentally better world than we did 10 years ago because of new therapeutics. Thanks to GLP-1 drugs, we now have an ability to respond to diabetes in a completely different way.
Drug discovery should only accelerate in the near future thanks to the transformational potential of generative AI.
“The discovery process for a long time has been a solo journey. It's a lone researcher doing their job,” Long said. “I hope we can help make it a collaborative one.”
To see a specific example of Dell Technologies working with partners like Intel to make AI more accessible and impactful in life sciences, read this solution brief on using AI to drive innovation in your organization with the right infrastructure and support: https://www.delltechnologies.com/asset/en-us/solutions/industry-solutions/briefs-summaries/democratizing-ai-with-dell-and-intel.pdf