AI co-scientists are emerging as a new category of tool in drug discovery, sitting between general-purpose AI and specialized computational platforms to work directly within the environments scientists already use, from ELNs to LIMS. In this Ask the Expert article, we spoke with Robert D. Brown, Ph.D., GVP and Head of the Scientific Office at Sapio Sciences, about what AI co-scientists can and cannot do today, how they fit into the design-make-test-analyze cycle, and what labs need to invest in now to get ahead of the shift.
Biocompare:What exactly is an AI co-scientist, and how does it differ from traditional lab automation tools or general-purpose AI like ChatGPT?
Robert: “General-purpose AI like ChatGPT knows a lot about science in the abstract but has no idea what experiment you are running or where your data lives. Traditional computational tools, such as molecular docking platforms or QSAR modelers, sit at the other extreme. They are built for specialists and require years of training, which is why bench scientists have always depended on a separate computational team to get those answers.
“A co-scientist works at two levels. The first is understanding the informatics software environment well enough to handle it on the scientist’s behalf: templates, searches, and experiment setup. The second, and more important, is understanding the scientific content of an experiment well enough to analyze it and advise on what it means or what to do next. That second level is the equivalent of having a highly skilled medicinal chemist, bioinformatician, or data scientist looking over your shoulder, acting as a genuine force multiplier.”
Biocompare:What are the main things an AI co-scientist can do to support scientists?
Robert: “There are really two capabilities. The first is platform intelligence. The co-scientist understands the informatics environment the scientist works in, so it can handle templates, reports, sample queries, and the data entry and search mechanics that pull a scientist’s time away from science.
“The second is the more consequential one. Discovery science runs in a cycle of Design, Make, Test, Analyze, with the design and analyze phases generally requiring a computational team running highly validated tools and models. The problem has always been the orchestration layer connecting those tools to the scientist. A co-scientist fills that gap, coordinating computational and design tools from within the scientist’s environment and pulling every output back into the scientific record.
Biocompare:Are AI co-scientists being used in labs today, and can you share examples of real-world implementations?
Robert: “We all know the stats: 10 to 12 years to move from target to patient, an approximately $2 billion investment per approved drug, and nine in ten clinical candidates still fail. What AI is starting to shift is the front end of that equation.
“The discovery phase from target to candidate usually runs three to five years, but we’re starting to hear from biopharma companies such as Recursion that they are cutting that to around 17 months and using far fewer physical compounds by moving from a model that was 90% wet lab to one that is largely dry lab, or in silico, work.
“In their case, what used to be almost all physical experimentation is now mostly virtual design. A better candidate gets to the clinical trial sooner, and that matters because two years earlier in the clinic means longer patent life and patients getting access faster.”
Biocompare:What are the biggest limitations or hurdles in deploying AI co-scientists right now?
Robert: “The hurdles are coming from multiple directions. The first is scientific credibility. Scientists are trained skeptics, and if the AI does not give answers as credible as what their computational team would have produced, they will wait and get the right answer the old way. Trust has to be earned through demonstrated scientific quality.
“The second is the gap between organizational ambition and scientific readiness. Mandates to become an AI-driven organization often arrive from boards and C-suites, complete with dedicated budgets, before the infrastructure needed to support them is in place.
“The third, often underappreciated, is agent overload. As every software vendor builds AI agents, the risk is that scientists go from managing 10 pieces of computational software to managing 200 agents instead.
“The answer to all three is a central orchestration layer that is scientifically credible, well-governed, reduces complexity rather than compounding it, and deploys AI in the environment scientists already live in, namely the ELN and/or LIMS.”
Biocompare:Looking ahead, how do you see AI co-scientists evolving over the next three to five years?
Robert: “The pace of development has been faster than most people predicted, and that makes the next three to five years genuinely difficult to forecast. But I can say with certainty that the research process will move from being predominantly wet-lab centric to having a much greater ratio of dry-lab work. AI co-scientists and AI-powered virtual research studies will drive that paradigm shift.
“The labs that benefit most from AI co-scientists will be the ones that invest in how their scientists work with them. Scientific training has historically been about using the software, navigating menus, building queries, and generating reports. That model is becoming obsolete.
“Prompt engineering is going to become an integral part of scientific practice. AI is designed to be helpful, so if you ask it to validate your hypothesis, it will often tell you that you are right. Learning to challenge the AI rather than simply consult it is a skill that has to be developed. The labs that invest in that capability now will be the ones that get ahead.”