A computational result can look finished when it reaches the bench: a ranked list of candidates, a predicted structure, a set of mutations, perhaps accompanied by a confidence score that makes one option appear more promising than another.
From the computational side, this can feel like an answer. From the experimental side, however, it is often the point where a different kind of work begins.
Before anything is ordered, cloned, expressed or measured, the prediction has to be translated into an experiment. That means understanding which candidates are actually worth testing, what the ranking represents, which aspects of the system the model captures, and which ones remain outside its scope. All of this has to be brought back to the biological question we are trying to answer.
I have often found that this translation requires much more effort than expected. What initially looks like a straightforward list of promising candidates can quickly become a longer discussion about how they were selected, what a particular score really represents, and whether the assumptions behind the prediction still hold under the conditions of the experiment.
The difficulty is not simply one of communication. A computational model and an experiment can be looking at the same biological system while describing different aspects of it. A model estimates what is likely within the assumptions, representations and training data that shape it, whereas an experiment observes what happens under a particular set of physical and biological conditions. Protein structure prediction offers a useful example of how important this distinction can become.
A predicted structure may reproduce the overall architecture of a protein remarkably well while still missing local conformations, interactions or alternative states that are important for function. AlphaFold has made this particularly visible: its predictions can be extremely valuable for generating hypotheses, but even high-confidence regions do not necessarily reproduce every structural detail observed experimentally, especially when ligands, post-translational modifications or environmental conditions are involved [1].
At the bench, details like these can change the way an experiment is designed and interpreted. A residue that appears perfectly positioned in a predicted structure may behave differently in the presence of a ligand; a protein that appears stable in silico still needs to be expressed and folded correctly; and a sequence ranked highly by a model still has to perform the function we care about under the conditions in which that function is measured.
The distance between a computational prediction and an experimental observation can therefore be substantial, and it becomes even more important when models move from describing existing biology to designing new sequences.
In one recent study, researchers expressed and purified more than 500 natural and AI-generated enzyme sequences in order to investigate which computational metrics were actually informative about experimental activity [2]. What is particularly interesting here is not simply that some AI-generated proteins worked, but that the criteria used to evaluate them computationally also had to be confronted with experimental reality.
The first predictions generated experimental data, and those data could then be used to improve how subsequent candidates were selected. Rather than acting as the final validation step at the end of a computational pipeline, the experiment became part of an iterative process in which each round could inform the next.
This way of working is already central to machine-learning-guided protein engineering. In active-learning workflows, a model proposes a relatively small set of variants, those variants are tested experimentally, and the resulting measurements are fed back into the next computational round. In a recent example, researchers repeated this process over several rounds to navigate a difficult enzyme sequence space, using what they learned experimentally at each stage to decide where to explore next [3].
Seen from this perspective, the boundary between computational and experimental work becomes much less rigid. The model helps narrow a very large search space and identify regions that may be worth exploring, while the experiment reveals where those predictions are useful, where they fail, and sometimes whether the original question itself needs to be reconsidered.
For an experimentalist, this also changes what makes a computational result genuinely useful. The prediction matters, of course, but so does the context that travels with it: what the model was optimized for, what information it received, what its scores actually represent, which biological conditions are missing from its representation, and where uncertainty remains.
Without that context, even a very good prediction can be difficult to translate into a good experiment. With it, the computational result becomes something that can be tested, challenged and refined, rather than simply accepted or rejected.
As biomolecular AI becomes increasingly capable of generating predictions and proposing new biological designs, this exchange between computation and experiment will become even more important. The challenge is not only to move predictions successfully from the model to the bench, but also to make sure that what we learn at the bench can travel back and improve the next computational step.
The real opportunity is to make that loop between computation and experiment work better.
References
- Terwilliger, T. C. et al. AlphaFold predictions are valuable hypotheses and accelerate but do not replace experimental structure determination. Nature Methods 21, 110–116 (2024). Link
- Johnson, S. R. et al. Computational scoring and experimental evaluation of enzymes generated by neural networks. Nature Biotechnology 43, 396–405 (2025). Link
- Yang, J. et al. Active learning-assisted directed evolution. Nature Communications 16, 714 (2025). Link