Why better AI needs better biology

Artificial intelligence has fundamentally reshaped drug discovery, accelerating target identification, decoding high-dimensional biological patterns, and prioritising actionable therapeutic concepts. Yet as the conversation shifts toward next-generation frontier models, the industry remains disproportionately fixated on capability. 

This focus overlooks the primary bottleneck: data fidelity. Unlike deterministic domains where data is structured and predictable, legacy biological data is inherently noisy, fragmented, and heavily reliant on non-translatable animal models. 

As AI transitions from retrospective data mining to predictive models capable of simulating human physiological responses, compute power is no longer the rate-limiting step. The bottleneck is the availability of scalable, reproducible, and physiologically faithful human biological data. 

AI is only as good as the biology it learns from 

The garbage in, garbage out principle is well known in technology, yet in life sciences the quality of the biological information used to train AI is hard to control. 

Over decades, researchers had been relying on a combination of animal models, immortalised cell lines and other surrogate systems to understand the processes of the disease and test potential therapeutics. Such approaches have driven substantial scientific progress, however, they  remain approximations of human physiology.. Data obtained in one model system is not interchangeable with data generated in another, and species-specific biology introduces systematic divergence.. 

This divergence  starts to matter when these datasets are used for training AI models. An AI model can be exceptionally effective  at identifying patterns in the training data it was fed but it cannot correct for the inherent biological biases in this information. If the experiments themselves do not properly represent the reality of human biology, no amount of architectural sophistication compensates for it, and increasing complexity has diminishing returns against a flawed input. 

Compounding the problem is variability, when  generating data at the scale needed for machine learning. Cells behave differently in different batches, in different laboratories and in different experiments; variance that is manageable within a single study but becomes a structural liability at high data volumes. 

Reproducibility is becoming a data problem 

One of AI’s greatest strengths is its ability to work at scale, surfacing patterns in data far beyond human analytical capability. However, scale is only an asset if the underlying data points are comparable. 

Consider training a  model on thousands of experiments in which the biological material is different every time. Patterns will still emerge, but it will be hard to distinguish whether these patterns represent genuine biological signal or just variation between experiments. 

This is why the problem of reproducibility is becoming increasingly weighty in the AI era. It’s no longer simply a marker of experimental rigour; it now determines whether results can be translated into useful training data too.. A great example here is CRISPR screening. By systematically perturbing gene activity across large cell populations, CRISPR screens generate dense mechanistic datasets that map which genes drive a given process, how these processes interact, and where tractable drug targets sit.. 

When run at scale in  consistent human cell models, the collected data becomes highly valuable to computational research, shifting AI from correlation-mining in historical datasets to learning from perturbation datasets explicitly designed to probe how human cells respond to defined changes This is where virtual cell models become particularly powerful, trained on  high-quality experimental data to predict cellular behaviour under different conditions. Those predictions are then interrogated in wet lab, generating new experimental data that refines the model in turn; a closed loop between computational prediction and experimental validation. 

The next AI advantage will come from human data 

The convergence of AI and human biology may become one of the most influential technological transformations of the next decade, but for that to happen, both of these spheres need to develop together. 

For example, the development of induced pluripotent stem cells (iPSCs) has given researchers direct access to robust human cell types in vitro.. Paired with technologies that enable manufacturing of these cells with batch-to-batch consistency at scale, iPSC-derived cells make it possible to generate large, standardised datasets grounded in human biology rather than surrogate models.  

The trajectory of AI development in this space will result in a future that looks markedly different to today. . Instead of relying mainly on fragmented, historical datasets, researchers will increasingly design experiments purpose-built to generate structured biological data on a large scale, test the computational predictions and use the results to improve their models. 

None of this, though, implies that computation should replace experimental biology; the value lies precisely in the feedback loop between the two.. Computational models can identify which questions are worth asking and where to look for answers; wet-lab biology provides empirical validation that answers are correct.  

For drug discovery, this results into better decision-making, earlier in the pipeline.. Rather than advancing hypotheses based on  signals that fail to translate humans, researchers can embed human biology in computational and experimental discovery from the outset. The AI conversation to date has centred mostly on models, algorithms and computing power. As the field matures, the lasting competitive advantage will shift to something harder to replicate: access to high-quality, human-relevant biological data that competitors cannot easily produce.