The repository is public. Now what?
Once a promising model appears in a paper and its repository is public, the architecture is described, and the reported results look relevant to a problem we care about, the practical work begins.
The software environment has to be reconstructed. Dependencies need the right versions. Internal data have to be mapped onto the expected inputs. Preprocessing steps that occupy a few lines in the Methods section may correspond to several functions in the repository. Eventually, a more scientific question appears: does the model operate on information that our experimental workflow actually produces?
These practices help establish whether the computational procedure can be rerun under specified conditions. In biomolecular AI, reproducing a model also requires reconstructing how biological data were transformed into the representations the model actually receives. Preprocessing, feature construction, graph definition and other domain-specific choices can affect the instantiated model even when the software itself runs correctly. Reproducibility therefore depends on recovering both the computational environment and the scientific choices encoded in the pipeline. [2,3]
Getting the code to run establishes computational reproducibility at the software level. In biomolecular AI, the next question is whether we can also reconstruct how the biological data were processed and represented to produce the model inputs.
From biological phenomenon to model input
Looking at the general path from biological phenomenon to prediction, we have to consider that biology does not enter a model in the form in which a scientist encounters it.
biological phenomenon → measurement → data → representation → computation → prediction
Each transition contains choices.
Which properties are measured? Which are ignored? How are missing observations handled? Which entities become nodes? What defines an edge? Which symmetries should the representation preserve? How is a continuous quantity discretized or normalized?
In molecular machine learning, representation is already recognized as a central part of model design. The way molecular information is encoded determines which properties and relationships are available to the learning algorithm. Established best practices therefore emphasize representations that retain the information relevant to the scientific problem and respect appropriate physical symmetries. [4]
For a reader trying to understand or reuse a model, this also means reconstructing what those representational choices actually encode. This matters for understanding because a representation is selective by design. It preserves some aspects of the scientific object and transforms or omits others. Reading the architecture alone rarely tells us the full story.
When performance feels like understanding
Messeri and Crockett recently described a broader risk associated with the growing use of AI in science. AI systems can increase researchers' ability to produce results while also encouraging what they call illusions of understanding: situations in which our capacity to generate answers grows faster than our understanding of the world those answers describe. [1]
Their argument concerns scientific practice broadly. Biomolecular AI provides a useful setting in which to examine a related tension.
A model can execute successfully, perform well on a benchmark and produce a plausible prediction while leaving important questions difficult to answer:
- What information was available to the model?
- How was that information represented?
- Which scientific assumptions shaped the representation?
- Which parts of the implementation determine the prediction we are looking at?
- How does the output relate back to an experimental decision?
Performance metrics require context. Biological datasets contain dependencies that can create subtle forms of data leakage, leading to optimistic estimates of how a model will behave in its intended setting [7]
In drug discovery, this connects to a broader evaluation problem: strong retrospective benchmark performance does not by itself show that a computational method will improve downstream scientific decisions. Recent work has therefore argued for evaluation strategies that are more closely tied to prospective use and decision-making, especially given the limited evidence for clinically relevant impact compared with the volume of model development and benchmarking. [8]
Prospective studies provide a clearer view of how computational models can be integrated into experimental workflows. In a 2025 large-scale screening study, ML-guided iterative screening tested 5.9% of a two-million-compound library while recovering 43.3% of the primary actives identified by the full high-throughput screen. In this setting, the model was used directly within an experimental decision loop. [9]
Different disciplines ask different questions
A biomolecular AI model is often shared by people whose expertise sits at different points along this chain.
An AI engineer may ask whether the model generalizes and whether the representation supports effective learning.
A computational biologist may ask what biological information the representation preserves and which assumptions entered during preprocessing.
An experimental biologist may ask why a prediction deserves the time and resources required to test it.
These questions refer to the same model but their answers live in different places: papers, datasets, equations, code, metadata, experimental protocols and sometimes the tacit knowledge of the people who produced them.
One person can learn to move across several of these layers. Expecting a single researcher to hold all of them together becomes increasingly unrealistic as models and scientific workflows grow more complex. This makes the connections between disciplines part of the scientific infrastructure itself: the information needed to interpret a model has to remain accessible across the different contexts in which the model is developed, evaluated and used.
Continuity of understanding
We find it useful to think of this as a problem of continuity of understanding. Much of the information required to interpret a biomolecular AI model is already available, but it is distributed across different sources and levels of description. The scientific claim may be stated in the paper, the corresponding operation formalized in an equation, the implementation distributed across several functions in the repository, and the provenance of the data recorded in supplementary material or metadata. Experimental context may sit elsewhere again.
Understanding the model therefore requires more than access to these individual pieces. It requires being able to follow how they relate to one another as a scientific idea is translated into data, representation, computation and eventually a prediction that may inform an experiment. As biomolecular AI workflows become more complex, preserving these relationships becomes increasingly important for researchers who need to inspect, evaluate or reuse a model.
How much of this chain should a scientist be able to reconstruct before a model can meaningfully inform the next experimental decision?
References
- Messeri L, Crockett MJ. Artificial intelligence and illusions of understanding in scientific research. Nature. 2024. Link
- Heil BJ et al. Reproducibility standards for machine learning in the life sciences. Nature Methods. 2021. Link
- Walsh I et al. DOME: recommendations for supervised machine learning validation in biology. Nature Methods. 2021. Link
- Artrith N et al. Best practices in machine learning for chemistry. Nature Chemistry. 2021. Link
- Wilkinson MD et al. The FAIR Guiding Principles for scientific data management and stewardship. Scientific Data. 2016. Link
- Barker M et al. Introducing the FAIR Principles for research software. Scientific Data. 2022. Link
- Bernett J et al. Guiding questions to avoid data leakage in biological machine learning applications. Nature Methods. 2024. Link
- Bender A et al. Artificial intelligence in drug discovery: what it is, where we stand and the path forward. Nature Reviews Drug Discovery. 2026. Link
- Müller AT et al. Machine Learning-Assisted Iterative Screening for Efficient Detection of Drug Discovery Starting Points. Journal of Medicinal Chemistry. 2025. Link