AI for drug Discovery: A powerful tool, but not a single solution

Artificial intelligence has become a routine tool in drug discovery, but its success has fueled a widespread misconception: that breakthroughs such as AlphaFold mean drug discovery is now largely solved. While AI-predicted protein structures can dramatically speed up the first steps of drug research, they do not automatically translate into effective medicines. In reality, the gap between prediction and cure is still a formidable obstacle, one that requires lab experimentation to overcome.

Advances in machine learning have lowered the barrier to obtaining plausible structural hypotheses, reshaping how early-stage discovery begins and compressing timelines that once stretched for months or years in a field where each delay compounds cost, risk and failure. Nonetheless, the central role of AI today is not to deliver answers, but to accelerate the search for them. Only lab experiments can determine which ideas hold up in the complex, dynamic context of real biology.

Where AI excels in drug discovery

In many drug discovery programs today, projects no longer begin in a lab, but rather on a screen. AI aids the earliest, most uncertain stages of research by making them faster, more focused and more selective. By rapidly generating structural models and testable hypotheses about how proteins might fold, interact or bind small molecules, AI helps researchers decide where to look first and what is worth testing at all.

Systems such as AlphaFold have made it routine to begin a project with a plausible three-dimensional model of a target protein, even in the absence of experimental structures. These models are best understood as starting points. They provide an initial spatial framework: where binding pockets might exist, which regions are likely rigid or flexible and how a protein could plausibly interact with small molecules. That context allows researchers to ask sharper questions much earlier than was previously possible.

AI is also particularly effective at helping interpret experimental data that is inherently noisy or incomplete. In techniques such as cryo-electron microscopy, machine learning systems assist with identifying meaningful signals in noisy images, automating repetitive processing steps and producing more consistent structural models than manual workflows alone. The advantage here is not just speed, but focus. AI removes much of the low-level work that once dominated early analysis, allowing scientists to concentrate on interpretation rather than data cleanup.

Perhaps most importantly, AI excels as a triage and decision-support tool in early discovery. In structure-based drug discovery, it helps prioritize which targets, binding pockets or compound designs are most worth experimental follow-up. Instead of testing thousands of possibilities blindly, teams can concentrate limited experimental resources on a smaller, more promising set of hypotheses.

Why good predictions can still be wrong

Modern AI models produce detailed structures, clean visualizations and neat numerical confidence scores. Seen on screen, their outputs can seem decisive, especially when confidence is read as a reliable measure of correctness. But in practice, it often is not. Many models are poorly calibrated. A prediction labeled “high confidence” does not consistently mean it is likely to be correct, particularly when it ventures beyond the kinds of biology or chemistry the model has seen before.

These weaknesses become most obvious under what researchers call domain shift. AI systems tend to perform best on familiar ground, such as well-studied protein families, common binding pockets and known chemotypes. As projects move toward unusual targets, highly flexible regions or novel compounds, performance can degrade quickly. 

A major drawback of AI models is that they sometimes create predictions that appear reliable, even when they’re not accurate. Generative models are designed to produce outputs that look plausible given patterns in existing data. Over time, they learn strong priors about what a “good” structure or interaction should resemble, including tidy hydrogen-bond networks, familiar geometries and chemically reasonable fits. That coherence is compelling. However, it is not the same as physical or biological truth, which depends on dynamics, energetics and context that the model may not fully capture.

In the lab, this gap can be deceptive. Polished images, authoritative-sounding confidence scores and a coherent molecular story are easy to mistake for proof. The core danger is not that AI is wrong. It is that AI can be wrong in ways that are systematic, persuasive and difficult to detect, unless experiments are deliberately designed to prove predictions wrong, not just to confirm them.

When experiments disagree with AI

The abstract limitations of AI become tangible when a prediction that looks impeccable on screen meets experimental reality in the lab. One illustrative case comes from fragment-based drug design, a common approach in which small chemical building blocks are gradually expanded into drug candidates. Here, AI models are often used to suggest how those fragments can be chemically elaborated into stronger binders.

Starting from experimentally determined fragment poses, a generative model proposed an expanded compound that looked ideal, at least on screen. The predicted binding mode (the orientation the molecule was expected to adopt inside the protein) placed it neatly in the pocket, forming an elegant network of hydrogen bonds and hydrophobic contacts. By every visual and chemical criterion the model had learned, the pose looked exactly right.

When the compound was tested experimentally, the initial result seemed encouraging: it did bind to the target. But structural validation revealed a crucial discrepancy. The molecule was not binding in the way the model had predicted. Its orientation in the pocket was different, and the interactions stabilizing the complex were not the ones that had guided the computational design. The prediction had captured plausibility, but not the actual physical mechanism. This was not necessarily a failure of the AI model; it’s simply that the complexity of real biological systems can reshape how a molecule behaves once it encounters its target.

This is why experiments play a deeper role than simple confirmation. In many cases, experimental data forces a revision of the story the model tells, revealing which elements are meaningful and which are artifacts of learned expectation. In drug discovery, progress often comes not from being proven right, but from seeing precisely how a reasonable prediction was incomplete, and what reality looks like instead.

Where experiments are needed (no matter how advanced AI becomes)

Cases such as this, where experiments contradict elegant predictions, illustrate an important truth: some of the most important questions in drug discovery cannot yet be answered by models alone. No matter how sophisticated AI becomes, experiments remain essential for understanding how molecules behave in real biological systems, where physical chemistry, cellular context and dynamics intersect.

The first of these interactions appears at binding. AI may suggest that a compound should bind and propose a plausible pose, but only experiments can establish whether binding actually occurs under physiological conditions. More importantly, experiments are required to measure how strong that interaction is, how fast it forms and breaks, and how stable it remains over time. 

Kinetic and thermodynamic properties, including binding free energy, on and off rates, and residence time (how long a molecule actually stays bound) often determine whether a compound becomes a viable drug or fails, despite an attractive structure. In many cases, these failures arise not from a lack of model sophistication, but from limits in uncertainty calibration and generalization beyond the data regimes on which current models are trained.

Beyond binding, biological context quickly takes over. Inside a cell, molecules compete, collide and adapt in a crowded environment shaped by feedback loops and constant change. A compound that appears selective in silico may interact with multiple unintended targets once it enters a cell, leading to off-target effects or toxicity. Determining cellular impact, metabolism and safety still depends on experimental assays, because these outcomes emerge from system-level behavior that models can only approximate.

Protein behavior adds further complexity. Many drug targets are not rigid objects but dynamic systems that shift between conformations, form transient complexes and respond to their surroundings. A single predicted structure captures only a snapshot. Experiments are needed to observe which motions and states actually matter for function and drug response in native conditions.

The underlying reason is not a lack of ambition in AI, but a limitation of data. Biology still relies on comparatively sparse, uneven and highly biased datasets. Until richer, more systematic experimental data exists, experiments remain the final arbiter: not as a rival to AI, but as the foundation that gives its predictions meaning and guides what should be tested next.

The productive loop

The most effective drug discovery programs no longer treat AI and experimentation as separate phases. Instead, they operate as a single, iterative loop, one designed to generate insight, not just results.

That loop often begins with AI proposing hypotheses: concrete, testable ideas about how a molecule might bind to a protein or alter its behavior. These outputs are not answers. Their value lies in compression. AI narrows an enormous chemical and biological search space into a smaller set of plausible directions worth exploring.

Experiments then take over. Rather than screening everything that is computationally possible, researchers design focused tests that probe the model’s assumptions directly. Structural measurements, biophysical assays and cellular tests reveal whether the system behaves as predicted, quickly separating productive hypotheses from elegant but misleading ones.

The outcome is rarely a simple yes or no. Often, experiments partially agree with predictions while exposing unexpected interactions, alternative binding modes, or unanticipated dynamics. These discrepancies act as signals, pointing to missing interactions, overlooked motions or limits in the data on which the model was trained. Each result feeds back into the next round of modeling, refining predictions and reshaping priorities.

By making experimentation more selective and hypothesis-driven, the loop turns limited experimental capacity into a strategic advantage: fewer blind screens, more decisive tests. For scientists at the bench and teams deciding where to invest time and resources, this loop defines what responsible use of AI looks like in practice.

Going forward: AI as an ally

The productive loop between AI and experiments works only as long as both sides are allowed to challenge each other. In practice, that balance can be fragile. When a predicted structure appears clean, coherent and convincing on screen, it can begin to carry more weight than the experiments designed to test it. At that point, the same tools that accelerate discovery can quietly introduce risk.

As drug discovery processes move forward, the biggest gains from AI will not come from more powerful models, but from how their predictions are used at the bench. The first priority should be reliable uncertainty estimations: clear signals of how confident a model is in each specific prediction. When uncertainty is communicated honestly and calibrated well, researchers can see where a model is likely to be fragile and design experiments that probe those weak points directly, instead of unknowingly building on assumptions that only look solid on screen. A confident prediction, in this setting, becomes less of a verdict and more of a prompt for a carefully chosen test.

Equally important are larger, more standardized experimental datasets. Many of the most consequential questions in drug discovery remain data-poor, especially around binding dynamics, how tightly and how long a drug binds, its functional effects inside cells, and its off-target behavior. Without that grounding, even sophisticated models are forced to extrapolate. Expanding high-quality, shared datasets would reduce blind spots and give predictions a firmer footing in biological reality.

Beyond better data, there is a shift in how AI itself should be used. The most valuable systems will not be those that simply generate answers, but those designed to guide experimental strategy, for example, by suggesting which measurement would reduce uncertainty the most, or which result would most decisively challenge a working hypothesis. This approach treats prediction and validation as parts of the same process, not separate phases, and it requires tighter integration between modeling and experimental pipelines.

The long-term goal, then, is not AI as an unquestioned authority, but as a transparent scientific partner, one that helps researchers decide what to test next and that leaves the final judgment firmly in the hands of experiments.

Why the lab still matters

Artificial intelligence has permanently reshaped how drug discovery begins. It provides faster access to structural hypotheses, early and testable ideas about how a molecule might fit its target, narrows vast search spaces, and brings a level of efficiency that would have been difficult to imagine even a decade ago. In that sense, AI has moved the starting line of discovery forward, compressing early uncertainty into something researchers can engage with more deliberately.

But experiments still determine how discovery ends. Only measurements can show whether a molecule truly binds, how it behaves in cells, which unintended interactions emerge and which hypotheses survive contact with biological reality. These are questions no model can yet answer on its own, no matter how convincing the prediction looks on screen.

Modern drug discovery is therefore best understood not as a choice between computation and experimentation, but as a system built on their interaction. AI proposes possibilities; experiments decide what holds up. Data reshapes models; models guide what is tested next. Progress depends on teams that understand both sides of this exchange and know when to trust each.