Deanne Taylor, a computational biologist, has spent years chasing a question that sounds almost too basic to still be unanswered: what does typical gene expression look like in a five-year-old, compared with a fifteen-year-old? According to MIT Tech Review, there is still no comprehensive answer, because the reference datasets biomedicine relies on to define "normal" molecular biology were built almost entirely from adult donors.

That gap sounds like a niche curatorial problem until you notice what sits on top of it: every diagnostic tool, clinical trial, and machine-learning model that touches pediatric genomics. If a reference dataset never included a healthy eight-year-old's transcriptome, no algorithm trained on it can reliably tell a clinician whether a child's gene-expression pattern is unusual or simply age-appropriate.

Taylor's work, as profiled in the piece, is an attempt to close that gap — building a map of how gene activity shifts across childhood, the same way pediatricians already track height and weight against standardized curves.

A growth chart with no molecular equivalent

Every pediatrician's office has a percentile chart for height, weight, and head circumference, built from decades of population data. Nothing comparable exists at the molecular level. There is no widely used reference for how transcription, splicing, or protein abundance normally shift as a child ages from infancy through adolescence. The result is a strange asymmetry: clinicians can tell within a percentile point whether a child is growing normally, but have no equivalent tool for whether that child's underlying gene activity is progressing normally too. That's not for lack of interest from researchers — it's a data-collection problem that has been left unsolved for far longer than most people realize.

Why the big reference atlases skew adult

Most large-scale gene-expression atlases were assembled from post-mortem tissue-donation programs and adult clinical biobanks — populations where consented tissue collection is logistically and ethically far simpler than it is for children. Foundational resources that genomics researchers treat as a default "normal" baseline, such as the GTEx project, contain little data from young children and almost none from infants. That skew wasn't a deliberate choice against children; it's a side effect of where large-scale sequencing infrastructure was first funded and built, back when adult disease — cancer, cardiovascular conditions — dominated research priorities. Teams building pediatric-focused tools have historically worked around that hole using small single-institution cohorts, animal proxies, or adult baselines adjusted with rough age-scaling assumptions — none of which is a substitute for real developmental data.

What a missing baseline does to downstream AI

The absence of a pediatric reference doesn't just slow down basic research. It propagates directly into any AI system trained or evaluated on top of adult-skewed genomic data:

• Diagnostic and variant-classification models trained on adult reference distributions risk flagging normal childhood gene expression as anomalous, or missing true pediatric pathology because the model has no sense of what counts as unusual at a given age.
• Genomic foundation models trained on public expression atlases inherit the same skew silently — nothing in a typical training pipeline flags that the model has effectively never "seen" a toddler's biology.
• Rare-disease triage tools that use AI to prioritize which patients need expert genetic review depend on an accurate normal range at each developmental stage; get that range wrong and the tool either buries clinicians in false positives or waves through real cases.
• Pharmacogenomic and dosing models trained on adult trajectories carry the same blind spot into pediatric drug development, a field already under-studied because children are routinely excluded from clinical trials.

AiiN's takeaway

For teams building health-tech AI, this is a useful reminder that a model is only as trustworthy as the population its reference data represents. Work like Taylor's won't produce a headline-grabbing new model — it produces the unglamorous baseline data that any future pediatric genomics system needs in order to be trained or evaluated responsibly. In our estimation, as more health systems attempt to fine-tune genomic foundation models for pediatric care, projects that supply age-stratified reference data will likely become as commercially important as the models built on top of them, since without that baseline there is no way to confirm that a system trained mostly on adult biology behaves safely when applied to a child.