I was deeply focused on studying the structure and dynamics of protein-ligand interactions, and explored working with a variety of molecular dynamics sampling methods in my dissertation, "Computational simulation studies of conformational changes of proteins and ligand binding mechanisms". But as the research progressed, I found myself facing a fundamental challenge: how to effectively select slow degrees of freedom and improve sampling efficiency in complex systems.

The turning point came when I met Prof. Weinan E and Dr. Linfeng Zhang, two pioneers in AI for Science  whose work on the Deep Potential Molecular Dynamics algorithm would later win the 2020 ACM Gordon Bell Prize, one of the highest honors in high-performance computing. Their way of thinking fundamentally changed how I approach scientific problems.

From a mathematical perspective, I came to a realization that has stayed with me ever since: AI, at its core, is a powerful tool for fitting high-dimensional functions. I began to see that the challenges I was facing in protein dynamics research were exactly the kind of problems this capability could help address. Building on this understanding, I developed what became the Reinforced Dynamics (RiD) enhanced sampling method, and eventually the RiDYMO® platform we built, Atombeat's AI-driven platform for "undruggable" targets.

My commitment didn't come from predicting future trends, but rather from recognizing that AI could solve the pressing scientific problems I was facing. In a sense, I was naturally drawn toward AI for Science as it revealed solutions that had previously been inaccessible or even unimaginable.

I've always taken a pragmatic view. Every technology, whether experimental techniques or AI, has its strengths and limitations. Even between animal models and clinical trials, there is an inherent gap; that is simply part of how science progresses.

What I've learned is that AI is essentially a toolkit with significant potential, particularly for what we call the "impossible targets". It has shown strong performance in hit discovery and rapid lead optimization, especially when combined with molecular dynamics. However, AI still faces challenges in areas such as ADMET prediction and clinical translation.

As data quality improves and agent-based reasoning gets smarter, this gap will likely narrow. But I don't think it will—or should—disappear entirely. That's why at Atombeat, we have always taken an integrated approach that combines AI, experimental validation, and human expertise. Ultimately, what makes AI meaningful in science is not the algorithm itself, but the human ability to understand what lies beneath the problem—and what is truly worth solving.

Coming into drug discovery when AI is a reality means we can approach problems through a fundamentally different lens. AI is not merely an add-on, but a new toolkit that offers multi-level, multi-depth impact.

For instance, it's hard to fully characterize the dynamic movements of proteins through experiment alone. With computational simulation and AI, we can model and visualize these processes more effectively. Similarly, AI enables virtual screening across a chemical space that is far larger than what was previously possible, significantly expanding the scope of discovery. In addition, AI agents can autonomously deploy different research tools and methods, enabling more quantitative and structured decision-making.

At the same time, AI does not replace the need for experimental validation. It's more about a symbiotic relationship: AI provides powerful new tools that must be integrated with and validated by experimental evidence. For me, this blend of advanced computation and rigorous experimentation is the defining characteristic of modern drug discovery.

Today, only about 20% of drug targets are currently tractable with conventional approaches. The remaining 80% are often labeled "undruggable", yet many represent some of the most promising opportunities in oncology, neurology, and immunology. To me, "undruggable" doesn't mean impossible. It simply means we need a different way of approaching these targets.

What has always fascinated me is the dynamic nature of proteins. They are not static objects but are constantly moving and sampling different conformations. Conventional drug design often assumes a fixed binding pocket, overlooking this dynamic structure. The AI platform we built was around a different perspective: modeling these motions and identifying transient binding sites that conventional approaches often miss.

This way of thinking has also enabled us to expand from small molecules into cyclic peptides, opening new possibilities for intracellular protein-protein interactions that were previously beyond reach. It has also strengthened our collaboration with partners such as BioDuro, as we share the belief that tackling the most challenging biological problems requires a combination of AI-driven design, experimental validation, and human expertise, and can lead to the greatest clinical impact.

Working on cyclic peptide rational design right now feels like standing at the crest of a transformative wave.

For decades, we've understood the immense potential of cyclic peptides—their stability, their ability to target difficult proteins, and their promise for oral bioavailability. Their complex conformational space, together with sheer computational burden of optimization, has made large-scale rational design was largely out of reach. However, oral cyclic peptides targeting PCSK9 and IL23R have demonstrated clinical viability, providing validation for the field. This convergence of clinical evidence  and technological progress  makes the present moment particularly meaningful.

Recent breakthroughs in AI and physics-based modeling means have given us tools that are now sophisticated enough to navigate high-dimensional problems. AI's ability to learn from and extract patterns in vast datasets, combined with advanced molecular simulations, enables more precise and efficient rational design than ever before. For me, this means being able to directly address critical unmet clinical needs, not only with oral cyclic peptides, but also in emerging modalities such as RDCs and POCs.

This is an area where AI, physics, and high-throughput experimental validation come together to open new therapeutic possibilities, and it is incredibly rewarding to be part of it.

Cyclic peptides are built from amino acids through well-understood amide bond formation, which makes them inherently "deliverable" as real therapeutic molecules that can be synthesized and developed—an important starting point.

Their natural fit with AI and physics-based modeling comes from their structural complexity, which is both significant and challenging. While cyclic peptides are ultimately designed to adopt stable 3D structures for their biological function, identifying those stable structure within a vast conformational space remains highly challenging.

This is exactly where AI and physics-based methods shine. Using advanced tools or high-dimensional sampling and representations, we can explore this vast chemical space, characterize complex 3D conformations, and predict how they may interact. This creates a reinforcing data-model flywheel, in which AI and physics models guide molecular design and screening, generate new data, which in turn helps refine and strengthen the models and accelerates discovery for promising molecules.

In our experience, the most productive collaborations happen when AI and experimental teams are tightly integrated and aligned at the level of molecular design and experimental validation.

An AI platform can rapidly generate hypotheses and prioritize the most promising compounds, but these predictions are only truly valuable when they are synthesized and tested. As our wet lab partner, BioDuro plays a critical role in closing this loop by validating results quickly and at high quality, generating data that continuously improves our models.

In this sense, we amplify each other: AI helps focus experimental efforts on the highest-value opportunities, while experimental results feed back into the models to further improve their accuracy and relevance over time.

Ten years from now, success will be measured by the outcomes it enabled in patients. Patients benefiting from medicines for previously untreatable diseases, and researchers routinely pursuing once undruggable targets, will be the strongest validation of the work being done today.