AI isn't ready to tackle our convoluted genomes, and a recent Wired report, reprinted from Quanta Magazine, explains why. The human genome still resists the kind of input-output modeling that machine learning handles well elsewhere. Genomic foundation models such as Evo 2, Genos and Google DeepMind's AlphaGenome can predict how a DNA change might affect a trait. They cannot yet explain why, because our genome's regulatory logic is far messier than a script or a blueprint.
This matters beyond academic biology. Health tech startups, pharma data teams and software companies exploring AI and ML development for life sciences clients need to understand this gap. It changes how confidently AI should be trusted compared to image recognition or text generation. The distance between prediction and understanding has direct consequences for product design, regulatory risk and client trust.
Why AI isn't ready to tackle our convoluted genomes today
Only about 2 percent of the human genome consists of actual protein-coding genes. Researchers once dismissed the rest as junk. They now know much of it governs gene regulation, the process of switching genes on and off depending on cell type, timing and external signals.
The article quotes Harvard Medical School biologist Karen Adelman, who notes that genes are "probably not the most interesting part of the genome." The harder problem is regulation. Transcription factors and enhancers interact in combinatorial, context-dependent ways. Unlike bacteria, where a single signal often triggers an on/off response, human regulation behaves more like an "AND" logic, where multiple signals must align before a gene fires. Wendy Bickmore of the University of Edinburgh points out that 23 years after the Human Genome Project was completed in April 2003, scientists still cannot map all enhancers to the genes they control.
Where genomic foundation models still struggle
Genomic foundation models are trained to find statistical correlations in sequence data. That approach works when enough labeled examples exist and the underlying pattern stays relatively stable. Gene regulation in humans does not behave that way. It involves several layers at once:
- Enhancers that can sit far from the genes they influence, sometimes looping through 3D chromosome folding to make contact
- Transcription factors that behave differently depending on which other factors share the same cell
- Regulatory combinations specific to a cell type and moment in time, not a fixed rule
A model trained mostly on well-studied cell lines and populations is likely to perform worse on rare cell types, underrepresented ancestries, or developmental stages not well covered in training data. That is a classic generalization problem. A model can look highly accurate in benchmark papers while still making confident, wrong predictions on real-world edge cases.
Why this matters for business leaders, not just scientists
Companies building products on genomic AI, or advising clients who do, should treat this as a data governance and risk management issue, not a purely scientific curiosity.
- Prediction is not explanation. A model can flag a sequence variant as likely pathogenic without indicating the regulatory mechanism, which makes clinical or legal review harder.
- Training data bias compounds quietly. Genomic datasets skew toward certain populations and well-studied cell types, so model confidence does not equal model accuracy for everyone.
- Regulatory scrutiny is rising. Health authorities in the USA, the EU and increasingly in Asia are asking AI vendors to document how predictions were validated, not just report accuracy scores.
Any company integrating genomic or biological AI into a product, whether a diagnostics startup or a data platform, needs a validation layer before model outputs reach a decision-maker.
Concrete steps a company should consider
- Keep a human expert in the loop for any genomic prediction feeding into a clinical, insurance, or hiring decision
- Document which populations and cell types your training data actually covers, and flag predictions outside that coverage as lower confidence
- Budget for wet-lab or clinical validation cycles rather than treating model output as a final answer
- Ask AI vendors directly how they handle enhancer-gene mapping uncertainty, since this is an unsolved problem across the field, not a vendor-specific gap
- Separate exploratory research use cases from product feature use cases; the bar for the latter should be much higher
Where this intersects with everyday software and data work
Most companies will never build a genomic model themselves. Many will still integrate AI predictions from partners or vendors into dashboards, CRMs or internal tools. The same discipline applies there too: log the model version, log the confidence score, and never let a downstream system silently treat a probabilistic prediction as confirmed fact. We see similar patterns when we build AI and machine learning features into business systems, where the biggest technical risk is rarely the model itself. It is how confidently the surrounding software presents its output to a non-technical user.
Teams working with sensitive health or genetic data also need to think about data protection and access control long before the AI layer. A leaked dataset is arguably a bigger business risk than an imperfect prediction. If your team is evaluating API integration work to connect a genomic AI vendor into an internal system, build in an audit trail from day one. Track who queried what, which model version answered, and what confidence score came back. Our cybersecurity services team can help scope that layer before go-live.
A note on timelines
Biologists in the Quanta piece are candid that a full mechanistic understanding of gene regulation may be decades away, if it arrives at all. That is a useful calibration for business planning. Genomic AI will keep improving incrementally, but it is not on a trajectory to produce fully explainable, audit-ready answers in the next product cycle. Plan integrations, compliance documentation and client commitments around that reality rather than around vendor marketing claims. If your roadmap depends on explainable genomic predictions arriving soon, build a fallback plan now, because the underlying science has not caught up with the demand. Teams weighing where AI fits into a broader business intelligence strategy should apply the same caution to any vendor claiming near-term explainability.
