Transfer Learning for Genomic Prediction in Underrepresented Populations

Transfer Learning for Genomic Prediction in Underrepresented Populations

In September 2026, Google Research published a blog post titled Transfer Learning for Genomic Prediction in Underrepresented Populations. The title captures something that matters far beyond the machine learning laboratory: if genomic prediction is to be clinically useful, it must work across populations, not only in the people who historically dominate research datasets. A model that performs beautifully in one genetic ancestry group can be dangerously misleading in another. Transfer learning—the practice of reusing knowledge from one prediction task to improve another—has become one of the most promising ways to address that gap.

This article examines the core problem, the logic behind transfer learning, and the practical and ethical conditions under which it can help. The referenced blog post is a useful indicator of the research direction. More importantly, the problem it names deserves a careful unpacking.

The Data Asymmetry Problem at the Heart of Genomic Prediction

Genomic prediction uses genetic information to estimate risk or predict a trait. It is the engine behind polygenic risk scores, agricultural breeding programs, and a growing number of precision medicine experiments. The usual workflow begins with a large dataset of genotyped individuals whose traits or health outcomes are known. A statistical model learns relationships between genetic variants and the outcome. Later, the model is applied to new individuals who may only have genetic data.

The challenge is that those large training datasets are not evenly distributed across world populations. Many widely used genomic resources are heavily drawn from people of European genetic ancestry. Research participation, healthcare access, national investment in biobanks, and historical patterns of scientific recruitment have all contributed to the imbalance. Consequently, a model may have enormous relevant training data for one group and very little for another.

Underrepresented populations face a double disadvantage. Their genetic variation is less likely to be tagged by existing variants in the model’s measurement scheme. Their clinical and environmental contexts may differ from the original training population. And even when they are included in a study, they may be represented by far too few individuals to train a robust prediction model. Small sample size means noisy effect estimates and unstable performance. Transfer learning offers a way to avoid starting from nothing in such settings.

What Transfer Learning Really Does

Transfer learning is not a single method. It is a family of approaches that share a simple insight: learning one task can help learn another related task. In classical machine learning, a model is trained from scratch for each dataset. In transfer learning, a model first learns general patterns from a data-rich environment. Then it is adapted to a data-sparse environment. The underlying expectation is that not all knowledge must be relearned in every new context.

The terminology often distinguishes between a source domain and a target domain. In the context of genomic prediction, the source domain might be a large, well-characterized cohort of one genetic ancestry. The target domain might be a much smaller cohort from another ancestry, or a clinical population whose combination of ancestry and environmental exposure is underrepresented in current databases. The goal is not merely to copy the source model. It is to move the useful parts of its "understanding"—for example, general patterns about how genes influence a trait—and then adapt the model to the target population’s specific genetic architecture.

This is essential because genomic prediction is not like object recognition, where a picture of a cat remains a picture of a cat across settings. Genetic effects are context-dependent. A variant may have a larger effect in one population because it is more common, because it is in a different linkage disequilibrium structure, or because it interacts with local environmental factors. Transfer learning can potentially handle this by allowing some model layers or parameters to remain general while other parameters are fine-tuned to the target data.

Why Genomic Prediction Does Not Transfer Automatically

It is tempting to assume that biology is universal and that a genetic risk score learned anywhere should apply everywhere. Human DNA is, after all, nearly identical across populations. But the differences, though small in percentage terms, are not random. They are shaped by migration, drift, selection, and admixture. When those differences intersect with the statistical shortcuts used by prediction models, transfer becomes hard.

One complication is linkage disequilibrium, the non-random association between genetic variants. A model may learn that a certain variant is predictive. That variant might not itself cause the trait. It might be a marker located near a causal variant. Because patterns of linkage disequilibrium differ across populations, the same marker might be a good proxy in one population and a poor proxy in another. A model that has learned "marker A predicts trait B" may fail once the marker-trait relationship is reassembled in a different population.

Another complication is allele frequency. Some genetic variants that are common in a source population are rare in a target population, and vice versa. A model trained on common variants in one population may not have seen the variants that matter in another population. Copying the model may therefore be equivalent to using a map written in invisible ink on a different terrain.

Environmental and clinical context matters too. Genomics does not unfold in a vacuum. A trait such as blood pressure is influenced by the same DNA across populations, but the environments in which that DNA is expressed—diet, stress, healthcare quality, medication availability, pollution—are not identical. If the source model has implicitly captured the environmental average of the source population, it may be severely miscalibrated when applied elsewhere. Transfer learning cannot magically infer a new environment from DNA alone.

An Illustrative Workflow: How Transfer Learning Might Be Used

A concrete example can clarify what researchers mean by transfer learning in this context.

Imagine an institution in a country whose population has been largely absent from global genomic databases. The institution can collect genetic data from thousands of patients, but not the hundreds of thousands often needed for a high-quality risk model. Funding and clinical outcomes are available, yet the sample is too small for a conventional genome-wide approach. Should the institution build a model from scratch? That would probably produce results that are noisy and unreliable.

Instead, the researchers could begin with a model that was pretrained on a massive, ancestrally European dataset. That model might not be directly appropriate for the local population, but it contains valuable information: a list of relevant variants, an understanding of regulatory regions, and approximate effect directions for many loci. The researchers could then fine-tune this model using their local data. During fine-tuning, the model adjusts its earlier assumptions while still retaining knowledge from the source domain. It might also incorporate an ancestry-aware component that accounts for differences in allele frequency and linkage disequilibrium.

The final model could then be evaluated in the local population using metrics that matter clinically—not average accuracy, but calibration and performance in specific subgroups. This is the basic pattern of transfer learning: pretrain on something large, fine-tune on something small, validate on something independent.

The illustration is deliberately idealized. In practice, data harmonization, variant representation, and informed consent create enormous difficulties. But it captures the structural promise of the approach: a small, underrepresented dataset does not have to remain isolated if models can learn general principles from larger datasets without committing to them permanently.

What Needs to Be Preserved During Transfer

The phrase "transfer learning" can give the misleading impression that a model is moved as a complete package. In clinically meaningful genomic prediction, there are several things that should never be moved carelessly.

One is calibration. Even if a model ranks people correctly by relative risk, the absolute risk estimates can be wrong if the population baseline is different. Transfer learning should therefore be paired with recalibration in the target population. A risk score that says "ten percent" in one population might need to become "six percent" when applied to another population, even if the ranking remains the same. Without calibration, patients and clinicians cannot interpret the result.

Another element is local ancestry. Many populations are admixed, with segments of DNA originating from different ancestral groups. A person may carry one set of variants that came from a European ancestor and another that came from an African ancestor. A simple transfer model that treats each individual as belonging to a single group will obscure this structure. Model architectures that explicitly incorporate local ancestry may be able to use information from multiple source populations simultaneously.

Finally, the model’s uncertainty needs to be reported. If fine-tuning is performed on a dataset with only a few thousand cases, the model will have larger uncertainty around rare variants and around variants that were not present in the pretraining source. A responsible genomic prediction system should communicate when the prediction is solid and when it rests on very little direct evidence.

Validation Is the Hardest Part

Transfer learning can reduce the amount of target data needed for training, but it does not remove the need for target data in evaluation. In fact, an independent validation dataset from the underrepresented population is more essential than ever.

Without external validation, a transfer learning model can look excellent during development and then fail at the point of clinical contact. This happens because fine-tuning can overfit to the small target dataset, especially when there are multiple ancestry groups or when the clinical outcome has many missing records. The model might memorize a handful of local patterns instead of learning the actual genetic relationship with the trait. Evaluation should therefore be performed on individuals who were not used for fine-tuning, and ideally across demographic subgroups within the underrepresented population as well.

Metrics also need to be chosen carefully. A model with high discriminative accuracy could still misallocate clinical resources if it is poorly calibrated. A model that performs well on average could perform poorly for a particular linguistic, geographic, or ethnic group hidden inside a broad category. For this reason, validation plans should not be left until the end of the project. They should be designed before transfer learning begins, when researchers still have the chance to collect appropriate outcome data.

The Ethical Limits of Transfer Learning

The biggest risk of transfer learning is not technical failure. It is the temptation to treat it as a substitute for global investment in diverse genomic research.

Transfer learning is an adaptation tool. It uses existing data more efficiently, but it cannot recover knowledge that was never collected. If a disease has never been studied in a particular population, no amount of transfer learning can invent the relevant clinical labels. If a community has not been meaningfully involved in genomic research, a model built through transfer learning may still lack the cultural and clinical sensitivity needed for real use. The model may be statistically valid and ethically invalid.

There are also concerns about data sovereignty. Transfer learning often relies on a pretrained model that was built on a separate population’s data. That population may not have consented to its use in deriving models for entirely different groups. The legal and ethical pathways for such transfers are still evolving. Researchers should be transparent about the origins of pretrained models and about any benefit-sharing agreements with source communities.

Moreover, the existence of a transfer learning method must not become an excuse for continuing to build massive resources in only a handful of populations and then exporting the results elsewhere. While the technical tool is valuable, the long-term goal should be equitable data generation, not merely equitable data borrowing.

Practical Takeaways for Researchers and Clinicians

For groups working on genomic prediction in underrepresented populations, a few practical guidelines are worth keeping in mind.

First, understand the source model. Researchers should know who contributed to the training data, what genetic variants were measured, and how the outcome was defined. A source model that is poorly documented is difficult to transfer responsibly.

Second, do not overweight pretrained effect sizes. Pretrained information is a prior, not a verdict. Fine-tuning should be strong enough to override harmful biases while weak enough to preserve useful general patterns.

Third, use ancestry-aware evaluation. Even underrepresented populations contain substructure. Risk models should be evaluated separately in groups that differ in ancestry, age, sex, and clinical context where possible.

Fourth, report what cannot be transferred. The absence of certain variants in a target dataset should be stated explicitly. If a model is being used under conditions unlike its validation environment, that should be visible on the output itself.

Fifth, involve the target population in the design process. A prediction model that works in a research dataset may be unusable in a community clinic. Questions about data collection, consent, result delivery, and clinical decision support are inseparable from model design.

Looking Forward

The Google Research contribution underscores that transfer learning has moved from an abstract machine learning concept into a central problem of genomic equity. The practical success of such methods will depend on more than algorithmic cleverness. It will depend on data sharing, on careful evaluation, and on recognizing that "underrepresented populations" are not a single monolithic category. Every transfer carries assumptions, and those assumptions must be named.

Transfer learning cannot solve every problem in genomic prediction. It cannot replace large, representative, community-engaged cohorts. It cannot fix missing clinical data. But it can change the trajectory of model development. Instead of asking every population to build its own infrastructure from nothing, the field can ask how existing resources can be repurposed without erasing local context.

In the end, the most important transfer is not the transfer of numeric weights from one population to another. It is the transfer of trust. Genomic prediction will only be accepted in underrepresented populations when the people who contribute their data can see themselves in the model, in the validation plan, and in the scientific decisions being made. Transfer learning is a technical bridge, but it must be built with that human relationship in mind.