AI Plus The Scientific Method Might Reinvent Drug Discovery
By Sergey Jakimov and Artem Trotsyuk, LongeVC
Recent breakthroughs have fueled a wave of enthusiasm around the potential of AI in drug discovery. AlphaFold’s prediction of 200 million protein structures was a massive moment in biology. Insilico Medicine’s rapid design of a novel drug candidate for pulmonary fibrosis in under 18 months — and at a fraction of traditional cost — underscores AI’s power. Yet for all the promise, AI alone is not enough. Biology is messy, context-dependent, and governed by mechanistic nuance. The true potential lies in coupling AI with the rigor of the scientific method, turning AI into a force multiplier rather than a replacement for experimentation.
In its classic form, the scientific method is a disciplined loop: ask a question, form a hypothesis, design and run experiments, analyze the results, and refine your understanding. It is this structured, iterative approach — rooted in reproducibility and evidence — that turns observations into reliable knowledge. Without it, even the most sophisticated algorithms risk producing results that look compelling on paper but fail when tested.
AI’s Limitations
AI is a world-class pattern recognition tool. But in the absence of good data, clear hypotheses, and experimental feedback, its outputs can quickly become unreliable. Many models learn from limited or biased public data sets, which often overrepresent positive results and underreport failed experiments. This prejudice creates blind spots that stall generalization and discovery. Meanwhile, models like AlphaFold capture static snapshots, while authentic biology plays out in dynamic, cellular contexts. Trust is also an issue. Many systems are black boxes, offering little transparency into why AI made that prediction or how to test it.
Another limitation is that many AI-generated predictions aren’t actionable in real-world experimental settings. A model might suggest a molecule that’s chemically intriguing but impossible to synthesize. Or it might ignore the physical limitations of an assay setup. Predictions that look good in a simulation can collapse under the weight of experimental constraints: time, cost, feasibility, or simply the lack of a compatible testing system. Then there’s the matter of speed. Biology is slow. It can take weeks to run an assay or generate a data point, which breaks the fast feedback loops that machine learning thrives on. Without the guiding structure of the scientific method — where every prediction is framed as a testable question and verified experimentally — AI risks producing outputs that look compelling on paper but fail in practice. Treating AI as a static tool that spits out answers rather than a dynamic system embedded in experimental workflows limits its utility.
The Solution
The solution lies in integration. Embedding AI into an iterative, closed-loop system of experimentation — what engineers might call a design-build-test-learn (DBTL) cycle — makes it dramatically more powerful. In this model, AI not only makes predictions but also generates hypotheses, informs the design of experiments, learns from real outcomes, and refines its understanding with each cycle. A drug discovery loop might start with a generative model designing a library of potential compounds. These compounds are synthesized, often via robotic or automated platforms, and tested in high-throughput assays. Results are fed back into the system, updating the models. The next round of suggestions is better informed, more targeted, and more likely to succeed. Over time, this loop accelerates both discovery and learning.
We’re seeing early but promising results from this paradigm. Insilico Medicine, as previously mentioned, uses AI to design novel compounds and identify new targets. They’ve advanced the first AI-discovered small molecule into clinical trials. Deep Longevity is a model that uses AI to generate predictive biomarkers and personalized longevity strategies by continuously learning from real-world data, which provides personalized health insights designed to drive preventive care and early interventions. Importantly, every round of failure is a learning opportunity. The system gets smarter with each iteration. Rather than starting from scratch each time, the model remembers what didn’t work and why.
Room For Improvement
To scale this model, labs need robotic systems capable of running experiments autonomously, cloud platforms for data sharing, and standards for interoperability. Findable, accessible, interoperable, and reusable (FAIR)-compliant data management systems, semantic knowledge graphs, and standardized lab protocols are critical. Without them, AI and experimental biology remain separated. Companies like Emerald Cloud Lab are pioneering cloud-based, remote-controlled laboratories that can run assays and capture data 24/7. Startups are also developing domain-specific languages (DSLs) and application programming interfaces (APIs) that allow AI systems to write, dispatch, and monitor lab protocols in real time. The endgame is a self-driving lab: AI proposes, robotics execute, data returns, and models update. The entire cycle happens with minimal human intervention, but always under scientific supervision.
The implications go far beyond speed. Drug discovery becomes not just faster but also smarter through the tight coupling of AI and experimentation. Instead of screening 10,000 molecules to find one hit, you might screen 100 because the model has learned what to avoid. Instead of waiting years for a lead compound, you might get there in months. By tracking, validating, and explaining each step, this approach addresses growing concerns about reproducibility and regulatory transparency. Black-box models will never pass muster with regulators. But a system that generates a hypothesis, tests it, and shows its work? That’s science.
Moving Forward
Of course, there are challenges. Cross-disciplinary talent remains rare. Most teams still separate AI, chemistry, biology, and lab operations into silos. Integrating them requires not just tools but culture change as well. It also requires investment in training, infrastructure, and collaboration across sectors. Academia can help by focusing on the messy, underexplored edge cases by publishing failed experiments and developing interpretable models. Policymakers can help by funding open data sets and automated lab infrastructure. Industry needs to move beyond hype toward truly integrated systems, where AI is not a feature but a foundational part of the discovery process.
AI alone won’t reinvent drug discovery, but AI embedded in the scientific method might. The real breakthrough will come when algorithms and experiments talk to each other, learning is continuous and rigorous, and by testing hypotheses in real time. The labs that master this integration will define the next era of biotechnology.
About The Authors:
Sergey Jakimov is a founding partner of LongeVC, a venture capital fund supporting early-stage biotech and longevity-focused founders that are changing the world. He is a serial entrepreneur, having co-founded three deep-tech ventures and raised more than $50 million in venture funding for his own ventures and as an entrepreneur in residence. He has worked with several other early-stage companies in the therapeutics space on fundraising, IP protection, and clinical trial strategies. He is also a visiting lecturer at several universities on venture capital and intellectual property rights. He co-authored a master's program in technology law at the Riga Graduate School of Law. Since 2018, he has co-founded a medical tech startup, Longenesis. In 2020, he co-founded LongeVC, and in 2021, the Longevity Science Foundation, a non-profit organization advancing the field of human longevity by funding research and development of medical technologies to extend the healthy human lifespan. He holds a BSc in international affairs from Rīga Stradiņš University and two MScs in political science and government and law and finance from Central European University and Riga Graduate School of Law, respectively. He was named Forbes Latvia 30 Under 30 in technology and healthcare in 2020.
Artem Trotsyuk is a partner at LongeVC, a venture capital fund supporting early-stage biotech and longevity startups. He is a bioengineer and computer scientist by training and an AI Fellow at Stanford University. Trotsyuk's experience focuses on early-stage investments (pre-seed, seed, up to Series A), supporting entrepreneurs to turn their ideas and visions into successful companies. Previously, he was an OpenAI Forum Member and an AI Trainer and has worked as an entrepreneur in residence for the R42 venture capital fund. He completed his Ph.D. in Bioengineering and master’s in computer science with an AI specialization at Stanford University under the supervision of Dr. Geoffrey Gurtner in the Department of Surgery. Trotsyuk's research interests lie in bioengineering, gene editing, wearables, CRISPR therapy, regenerative medicine, and ethical use of data in drug development.