By Dr. Neil Taylor
When AlphaFold2 was publicly released, it marked a genuine turning point for structural biology. Suddenly, researchers had access to high-confidence predicted protein structures for almost the entire human proteome. It was a major leap forward for AI in drug discovery, and the excitement around it was well deserved.
But now that the initial wave of enthusiasm has settled, a more important question deserves attention: has protein structure prediction meaningfully changed how drugs are discovered?
The honest answer is: not yet to the extent many had hoped. That is not a criticism of the technology. It is a reflection of where the real challenge lies. Predicting structures is one thing. Turning those structures into usable insight for structure-based drug design is another entirely.
The field has become very good, very quickly, at generating model structures. What it is still learning to do is use them well.
In many respects, this is a familiar problem in computational chemistry. For decades, researchers have had the ability to generate vast hit lists of candidate molecules, docking poses and ranked predictions. The bottleneck has rarely been producing options. It has been knowing what to do with them next. AI in drug discovery has accelerated that challenge. Today, many teams can generate more modelling results, more quickly, from more tools than ever before. But more output does not automatically mean better decisions.
This is where the conversation around drug discovery machine learning needs to mature. The value is not in producing another ranked list. The value is in improving predictive accuracy, refining hypotheses and accelerating the discovery of viable candidates with data that scientists can interrogate and trust.
That is also why analysis of protein structure remains such a critical discipline. A predicted structure, even a highly accurate one, is still only a model. It can be enormously useful, but it does not remove the need for scientific judgement, experimental validation and robust data infrastructure.
One of the core limitations of structure prediction is that predicted proteins are generally represented as static snapshots. Drug targets are not static.
Proteins move. They breathe, flex, shift between conformational states and reveal cryptic binding pockets only transiently. A single structure, no matter how strong the prediction, cannot fully capture that behaviour. This matters enormously for structure-based drug design, ligand-based drug design and computational workflows that depend on understanding how a target behaves in realistic biological conditions.
To move beyond static models, teams need to integrate predicted structures with richer dynamic information. That includes molecular dynamics, steered molecular dynamics, MD simulation workflows, cryo-EM ensembles and other experimental or simulation-derived data. Tools such as GROMACS and related molecular dynamics platforms help generate this kind of insight, but without the right data pipeline and the right environment to manage and interrogate the output, the value can quickly become fragmented.
This is the real next frontier: not just structure generation, but structural utilisation at scale.
What is especially exciting now is not only predicting what a protein looks like, but beginning to design new proteins with intended functionality.
Recent advances in deep learning are pushing structural biology and computational chemistry into a new phase. Researchers are increasingly exploring how to engineer proteins that switch states, respond to stimuli or adopt specific geometries on demand. That changes the conversation dramatically.
It suggests a future in which scientists are not only screening for small molecular binders, but also designing entirely new protein-based therapeutic tools.
Dr. Neill Taylor, Founder, DesertSci
For drug discovery, the implications are significant. It suggests a future in which scientists are not only screening for small molecule binders, but also designing entirely new protein-based therapeutic tools with tailored structural and functional properties. That opens new possibilities across biologics, engineered binders and broader therapeutic discovery.
But it also raises the bar for infrastructure. Designing proteins with predicted function requires more than isolated models. It requires connected systems for analysing, comparing, validating and sharing structural data across multidisciplinary teams.
For pharma and biotech organisations, the strategic implication is clear. Structural biology can no longer sit in a silo.
The organisations most likely to lead the next decade of innovation will be those that treat structural data as a living scientific asset – something continuously refined, computationally interrogated and closely integrated with medicinal chemistry, biology and discovery teams. That shift matters not just for structural biologists or modelling specialists, but for all the data consumers and data producers involved in modern discovery programmes.
This is where infrastructure becomes decisive. Better structural data management platforms, AI-ready repositories, stronger data pipelines and more connected scientific environments are no longer optional. They are becoming essential to making artificial intelligence in drug discovery and development genuinely useful in day-to-day research.
Platforms such as DesertSci’s Proasis play an important role here. By helping teams manage, interrogate and act on structural data at scale, Proasis supports the move from isolated prediction to practical scientific use. It helps bridge the gap between predicted structures and the iterative demands of real-world SBDD workflows, where reliable access to data, context and history improves both speed and scientific confidence.
AlphaFold helped democratise access to protein structure in a way that was genuinely revolutionary. But democratised access is only the starting point.
The scientists and organisations that will shape the future are those asking bigger questions. Not just what a protein looks like, but how it behaves, how its behaviour can be modified and how that understanding can be translated into better medicines.
That is why this moment feels larger than the arrival of a single prediction tool. The field is moving into a design era, where the challenge is no longer simply to generate structures, but to use structural insight more intelligently across the discovery process.
For that to happen, the industry needs to be ready. That readiness starts with infrastructure: the ability to manage structural data, integrate dynamic insight, support computational chemistry workflows and turn information into action.
In that environment, AI in drug discovery becomes more than a source of predictions. It becomes part of a practical, connected and scientifically rigorous system for discovery.
For more insights into AI in drug discovery, computational chemistry and research-grade scientific software, follow follow Dr Neil Taylor on LinkedIn. Or, if you’d like to arrange a demonstration of DesertSci’s Proasis, please get in touch with our team.
