Biomolecular AI has given us a strange new problem: having too many ideas.
A model can generate molecules, propose protein sequences, rank mutations or suggest possible interactions long before a scientist has finished examining the first few outputs. Producing another computational candidate is becoming progressively cheaper. Deciding whether that candidate deserves anyone's attention is not.
For years, much of the excitement around AI in biology has been about expanding what we can predict. We are now beginning to see the consequence of that success. Once predictions become abundant, the difficult part starts to move elsewhere.
The scarce resource is no longer only computation or even experimentation. It is scientific attention.
From scarcity of hypotheses to scarcity of attention
Drug discovery has never lacked possible hypotheses. What is changing is how quickly we can generate and explore them. Generative models can propose molecules and protein sequences at a scale that would be impossible to explore manually, while predictive models can rapidly evaluate and prioritize large numbers of candidates.
That expansion is useful, but the number of outputs is a poor proxy for scientific progress.
Recent work on AI-generated molecules makes the distinction explicit. Metrics such as novelty, validity and diversity can characterize what a model generates, but they do not establish that a proposed molecule is chemically credible, biologically relevant or ready to become an experiment. The authors argue that generated molecules are better treated as hypotheses whose translational readiness has to be established through multiple layers of evidence. [1]
Another recent review describes the same transition as moving from chemical space to “synthetic reality”: generated molecules eventually encounter route planning, available precursors, laboratory execution and medicinal-chemistry judgment. In other words, expanding the space of possible answers does not remove the work required to decide which answers matter. [2]
This creates an asymmetry. The bandwidth of hypothesis generation is increasing quickly. The bandwidth available to inspect those hypotheses, connect them to existing knowledge and decide what deserves further work is much harder to scale.
More experiments are only part of the answer
One obvious response is to increase experimental throughput.
Automation, high-throughput experimentation and self-driving laboratories are already pushing in this direction. These systems combine machine learning, robotics and automated analysis so that larger numbers of experimental decisions can be executed and updated with less manual intervention. The field is now developing benchmarks not only for automation, but for whether these systems actually accelerate learning. [4], [5]
That matters enormously. But making experiments faster does not make every computational hypothesis equally worth testing.
This is where the idea that “the bottleneck is testing” becomes more interesting. Testing capacity is certainly finite, but the problem begins before a sample enters an instrument. Someone, or some system, has to distinguish between candidates that are merely high-scoring and candidates that are scientifically informative.
In reports of real-world AI/ML deployment in drug discovery, this distinction is already visible. Computational prioritization can narrow candidate sets, but medicinal chemistry and experimental knowledge continue to shape what is practical, interpretable and useful to pursue. Model recommendations live inside a much larger decision process [3].
And these decisions depend on the scientific question. A protein scientist may need to assess whether a proposed interaction remains plausible in its molecular environment. A medicinal chemist may ask whether a molecule opens a useful chemical series rather than simply optimizing a score. A structural biologist choosing a designed binder may care about epitope accessibility, sample stability or whether the binder will actually contribute useful information to the intended experiment.
Each decision brings together experimental knowledge, biological context and prior evidence. As the number of AI-generated candidates grows, the challenge is not simply to process more outputs, but to determine which ones are consistent with what we already know, which ones are worth challenging, and which ones can lead to an informative experiment.
AI can expand the space of possibilities much faster than it can reduce the scientific work required to navigate it.
The organizations being built around AI already reflect this
There is an interesting signal in the organizations that describe themselves as AI-native or AI-first in biology.
Current job openings are too transient to support a meaningful industry-wide ratio between experimental scientists and AI engineers. But they do show something more useful. Companies such as Isomorphic Labs, Xaira Therapeutics, insitro and Generate:Biomedicines are not being assembled as machine-learning teams with a laboratory attached at the end. Their organizations combine machine learning and computational design with medicinal chemistry, pharmacology, biology, data generation, automation and experimental development. [6], [7], [8], [9]
This makes sense. As computational systems become capable of producing more candidate hypotheses, experimental expertise becomes part of the mechanism by which those hypotheses are filtered and turned into knowledge.
The interesting scaling problem is therefore not simply how to employ fewer experimental scientists. It is how to increase what each scientist can meaningfully evaluate without reducing scientific judgment to a model score.
Scientific attention needs infrastructure too
Software has already transformed our ability to scale computation. Automation is beginning to transform our ability to scale experiments. We think a similar question should be asked about scientific reasoning itself.
What would it take to help a scientist inspect a larger number of computational hypotheses without simply presenting a longer ranked list?
Part of the answer is better models. But part of it is also how information around the model is organized.
A computational output sits inside a network of context: the model architecture and representation that produced it, the data on which the model learned, the assumptions embedded in the task, the evidence already available in the literature, known biological constraints, alternative explanations and the experimental question that motivated the analysis.
Today, much of that context remains fragmented. Some of it is in a paper, some in supplementary information, some in a repository, some in the head of the computational scientist, some in the head of the experimentalist, and some is reconstructed only when a result becomes difficult to explain.
When there are ten computational hypotheses, scientists can compensate for this fragmentation manually. When there are thousands, that becomes a different kind of problem.
What we mean by building a bridge
This is where our thinking at TenesimBio is becoming more concrete.
For us, the bridge between biomolecular AI and experimental science is not simply a communication layer between two groups of people. It should help organize the scientific context required to interrogate a computational result.
That means being able to move from an output back toward the evidence and assumptions behind it; to compare candidates using criteria that make sense for the biological question rather than only the model objective; and to surface where information is missing, conflicting or uncertain before that uncertainty is hidden by a single score.
Imagine receiving one hundred predicted candidates. The useful interface is not necessarily the one that makes the list prettier. It is the one that lets a scientist ask why candidate #17 is interesting, what distinguishes it from candidate #18, which evidence supports that distinction, which assumptions it depends on, and what information would most change the decision.
This is not a replacement for scientific judgment. It is an attempt to give that judgment more leverage.
And it changes how we think about explainability. An explanation is useful not because every internal operation of a model needs to be translated into natural language, but because the scientist needs enough access to the relevant reasoning, evidence and uncertainty to decide what deserves attention.
Conclusion
As hypothesis generation continues to scale, scientific attention may become one of the resources we need to design for.
How do we help scientists navigate a growing space of computational possibilities without reducing scientific judgment to another score? What information needs to remain attached to a prediction for it to be questioned, compared and eventually acted upon?
And, perhaps most importantly, what should a biomolecular AI model actually explain to the scientist deciding what deserves attention next?
References
- Liu et al. (2026), Evaluating AI-Generated Molecules for Drug Discovery: From Generic Metrics to Translational Readiness, International Journal of Molecular Sciences. Link
- Simonetti et al. (2026), From chemical space to synthetic reality: A framework for makeable-by-design generative AI in drug discovery, Drug Discovery Today. Link
- Schneider et al. (2025), Real-World Applications and Experiences of AI/ML Deployment for Drug Discovery, Journal of Medicinal Chemistry. Link
- Tobias & Wahab (2025), Autonomous 'self-driving' laboratories: a review of technology and policy implications, Royal Society Open Science. Link
- Adesiji et al. (2026), Benchmarking self-driving labs, Digital Discovery. Link
- Isomorphic Labs, careers, accessed September 2026. Link
- Xaira Therapeutics, careers, accessed September 2026. Link
- insitro, Work With Us, accessed September 2026. Link
- Generate:Biomedicines, company platform, accessed September 2026. Link
- Géoui (2026), Experimental Science's AI Denial Is Becoming Arrogance, LinkedIn. Link