AI Genomics Limitations: Why DNA Still Defies Prediction

Genomes are not simple blueprints

It is tempting to think of DNA as a fixed blueprint: one sequence, one set of instructions, one outcome. In practice, a genome is more like a library of possible instructions. Cells use different parts of it depending on their role and circumstances.

Even a single genetic change can have different effects depending on where it occurs and what else is happening in the cell. A variant might alter a protein, change when a gene turns on, or have little measurable effect in one context and a substantial effect in another. The sequence alone does not always reveal which outcome will occur.

Evolution has also built genomic systems through many small changes over billions of years. The result is intricate organization, not a clean design with one obvious rule for every gene. Regulatory signals can overlap, interact, and sometimes work at a distance from the genes they influence. Machine learning in biology has to find useful patterns within this complexity without assuming that the patterns are simple or universal.

Gene activity depends on layers of interacting regulation

A gene’s activity is shaped by multiple processes. Models trained on genomic data may be able to predict some of these processes, but the processes themselves do not operate independently.

After a gene is transcribed, splicing determines how sections of its RNA are joined. The same gene can produce different RNA forms, which may lead to different proteins or affect how much usable RNA is made. Epigenetic marks add another layer: chemical modifications to DNA or its associated proteins can influence which regions are accessible and active, without changing the underlying DNA sequence.

These processes can affect one another. A DNA change may alter a binding site, which could influence transcription, while the resulting RNA is also subject to splicing choices. A model that predicts one measurement may not capture the full chain of events that produced it.

Chromatin loops and packing therefore add spatial context that a model cannot infer reliably from a short sequence alone. Even when a model uses long DNA segments, its predictions depend on the training data and on how well those data represent the relevant cell state and molecular interactions.

The same DNA can behave differently across cells and over time

Nearly all cells in a person carry broadly the same DNA, yet a liver cell and a neuron use different sets of genes. They differ in the regulatory proteins they produce, the accessibility of parts of their genome, and the signals they receive. Context-dependent gene expression is central to biology, not an exception to it.

Activity can also change over time. Development, aging, infection, medication, and environmental conditions can all shift gene regulation. A model trained on cells measured under one set of conditions may not accurately predict what happens in another. Even data labeled as coming from the same cell type can vary because of differences in sampling, experimental methods, or the cells’ state at the time of measurement.

This makes the task more demanding than mapping a DNA sequence to one fixed answer. To make a useful prediction, a system may need information about the cell type, biological condition, and process being measured. Without that context, an accurate average prediction can still miss important effects in a particular setting.

Genomic foundation models can predict patterns, but prediction is not understanding

Genomic foundation models learn statistical relationships from large datasets, then use those patterns to make predictions about sequences or their effects. They offer a way to combine evidence across many regions and regulatory measurements. But a model can perform well on a prediction task without representing the biological mechanism that caused the outcome.

However, a prediction remains conditional on the model’s data and task. If its training examples underrepresent a cell type, biological condition, or kind of genetic variation, performance may not transfer well to that case. Predictions should therefore be treated as evidence to investigate, not as final answers about what a variant does in every living system.

Mechanistic understanding usually needs additional evidence. Researchers can compare predictions with independent datasets, test whether they hold in relevant cell types, and use experiments to alter a suspected regulatory element. If changing that element produces the predicted effect, it offers stronger support for a causal explanation than a model score alone.

What it will take to make AI predictions useful in biology

Progress depends on evaluating models against the complexity of living systems, not just their performance on a convenient test set. Useful checks include:

• Test in relevant contexts: Evaluate predictions across cell types and conditions, especially those not represented well in training data.
• Check generalization: Use independent datasets and experimental methods to see whether a model’s results hold beyond its original examples.
• Separate prediction from explanation: Report what a model predicts, how uncertain it is, and whether experiments support the proposed mechanism.
• Use AI to guide experiments: Treat model outputs as ways to prioritize questions, then test important claims in the lab.

AI in genomics can help researchers find patterns that would be difficult to identify manually. Its limits become clearer when a prediction is mistaken for a universal rule or a complete explanation. The strongest results will come from combining computational predictions with biological context and experimental validation. Understanding gene regulation is a practical starting point for interpreting what these models can - and cannot - tell us.

Step-by-Step Guide

1. 1

   Define the biological context

   Specify the cell type, condition, time point, and molecular outcome you want to predict so the model's task matches the biology being studied.

2. 2

   Check training-data coverage

   Review whether the model's training data represent the relevant cell types, conditions, and kinds of genetic variation; note important gaps.

3. 3

   Test predictions independently

   Evaluate model outputs on datasets and experimental methods that were not used for training, and compare performance across relevant contexts.

4. 4

   Validate mechanisms experimentally

   Treat predictions as hypotheses, then test important claims by altering candidate regulatory elements or variants in an appropriate biological system.

Key Statistics

• The human genome contains approximately 3 billion DNA base pairs.The National Human Genome Research Institute describes the human genome as roughly 3 billion bases; this scale illustrates the breadth of sequence information genomic models must analyze.
• The human genome contains approximately 20,000 protein-coding genes.The National Human Genome Research Institute gives an estimate of about 20,000 protein-coding genes. This is a gene count, not a measure of how many regulatory interactions or context-specific effects a model can explain.

Frequently Asked Questions