Target-Ancestry Fine-Tuning and Transfer Learning

Why Fine-Tune a Score?

A polygenic score developed in one ancestry can lose predictive performance in a target ancestry because allele frequencies, linkage disequilibrium (LD), and genetic effects may differ. Earlier material on ancestry portability distinguishes these contributors and motivates using target-ancestry data when available. Target-ancestry fine-tuning uses those data to update discovery-derived weights, aiming to improve prediction in the target population.

The central decision is how much to move away from the discovery weights. A small target GWAS may provide noisy evidence, so replacing the original weights with target estimates can make prediction worse. A larger, well-designed target GWAS can support a more substantial update. But sample size alone does not guarantee better performance: LD can make some effects poorly identified, overlapping samples can make evidence dependent, and tuning on the same individuals used to assess performance can overstate gains.

Definition: Let \(w_D\) be a vector of discovery-derived score weights, and let \(\beta_T\) denote the vector of genetic effects relevant to the target ancestry under a specified outcome model. Target-ancestry fine-tuning is an estimation procedure that combines \(w_D\) with target-ancestry GWAS evidence to produce updated weights \(w_F\). The resulting score for an individual with genotype vector \(G\) is \(S_F=G^T w_F\). Fine-tuning updates weights; it does not, by itself, guarantee improved target prediction or calibrated risk.

A Shrinkage Update and Its Assumptions

A useful starting point is a target GWAS that provides estimates \(\hat{\beta}_T\) of the target effects, with sampling-error covariance matrix \(V_T\). For the moment, suppose these are estimates of joint effects for the same variants and outcome model represented by \(\beta_T\), so that \(\hat{\beta}_T\mid\beta_T\sim N(\beta_T,V_T)\). Suppose also that the difference between the target effects and discovery weights has a normal prior: \(\beta_T\mid w_D\sim N(w_D,\tau^2 I)\). Here \(I\) is the identity matrix, and \(\tau^2\) describes how much the target effects are expected to differ from the discovery weights. This is a modeling assumption, not a fact established merely by having two GWAS.

Under these assumptions, the posterior mean is a precision-weighted combination of the target estimates and discovery weights:

w_F = \left(V_T^{-1}+\tau^{-2}I\right)^{-1} \left(V_T^{-1}\hat{\beta}_T+\tau^{-2}w_D\right).

The equation shows why target sample size matters. With more target participants, sampling variances often decline, so the target estimates receive more weight. If target estimates remain imprecise, the update stays closer to \(w_D\). In practice, the covariance structure, trait scale, variant set, and model used for both estimates must be compatible; otherwise this formula does not describe a coherent combination.

Model qualification: The preceding update treats \(w_D\) as fixed and assumes target estimation errors have covariance \(V_T\) under the stated likelihood. If discovery weights are uncertain, or the discovery and target GWAS overlap, their joint uncertainty and covariance must be represented. Treating dependent estimates as independent can exaggerate the amount of information.

Worked Example: A Noisy Target Estimate Gets a Conservative Update

Suppose one discovery weight is \(w_D=0.20\), in units of outcome change per genotype unit. A target GWAS estimates the corresponding effect as \(\hat{\beta}_T=0.10\), with standard error \(0.20\). For this hypothetical example, assume independent discovery and target data, a normal target likelihood, and a prior \(\beta_T\mid w_D\sim N(w_D,0.10^2)\). Thus, the prior variance is \(0.01\), and the target estimate variance is \(0.20^2=0.04\).

In the one-weight case, the posterior mean is \[ w_F= \frac{\hat{\beta}_T/SE^2+w_D/\tau^2}{1/SE^2+1/\tau^2}. \] Substituting the values gives \[ w_F= \frac{0.10/0.04+0.20/0.01}{1/0.04+1/0.01} = \frac{2.5+20}{25+100} = 0.18. \]

The updated weight is \(0.18\), closer to the discovery weight than to the noisy target estimate. This is a consequence of the assumed uncertainty, not a finding that the discovery estimate is necessarily correct. With a larger target sample and a smaller standard error, the target evidence would generally exert more influence under the same model.

LD Changes What Target GWAS Estimates Mean

A target GWAS usually reports marginal associations: each variant is tested separately, rather than estimated jointly with every other variant in the score. Under a standardized-genotype linear model \(Y=G\beta_T+\varepsilon\), with genotype correlation matrix \(R_T\), the expected vector of marginal effects is \(R_T\beta_T\), not generally \(\beta_T\). Thus, directly treating marginal effects as joint effects can produce an inappropriate update, especially when variants are correlated.

If \(\hat b_T\) is the vector of marginal estimates and its sampling-error covariance is \(V_T\), a compatible normal likelihood is \(\hat b_T\mid\beta_T\sim N(R_T\beta_T,V_T)\). With the same prior centered on \(w_D\), the posterior mean becomes

w_F = \left(R_T^T V_T^{-1}R_T+\tau^{-2}I\right)^{-1} \left(R_T^T V_T^{-1}\hat b_T+\tau^{-2}w_D\right).

This expression makes the role of target LD explicit. Directions in the effect vector that produce similar marginal associations can be difficult to distinguish. In particular, when variants are highly correlated, the data may be informative about their combined signal but weakly informative about how that signal should be divided between their individual weights. Regularization or a suitable prior stabilizes such weakly identified directions; it cannot create information that the target data do not contain.

Worked Example: LD Makes One Effect Direction Hard to Estimate

Consider two standardized variants with target-ancestry correlation \(r=0.80\), so \(R_T=\begin{pmatrix}1&0.80\\0.80&1\end{pmatrix}\). Suppose the hypothetical marginal GWAS estimates are \(\hat b_T=(0.18,0.16)^T\), their sampling-error covariance is \(V_T=0.01I\), and the discovery weights are \(w_D=(0.10,0.10)^T\). Use the prior covariance \(\tau^2 I=0.01I\).

The two orthogonal directions of this LD matrix are the common direction \((1,1)^T\), with eigenvalue \(1.80\), and the contrast direction \((1,-1)^T\), with eigenvalue \(0.20\). In these directions, the marginal estimate must be divided by the corresponding eigenvalue to estimate the effect direction. The observed common-direction estimate is \((0.18+0.16)/\sqrt{2}=0.2404\), giving an effect estimate \(0.2404/1.80=0.1336\). The contrast-direction estimate is \((0.18-0.16)/\sqrt{2}=0.01414\), giving \(0.01414/0.20=0.07071\).

The target-estimate variances for these two effect directions are \(0.01/1.80^2=0.003086\) and \(0.01/0.20^2=0.25\), respectively. For each direction, the posterior mean is the discovery value plus a fraction of the difference between the target estimate and discovery value; that fraction is \(0.01/(0.01+\text{target-estimate variance})\). In the common direction, the discovery value is \(0.20/\sqrt{2}=0.1414\), and the fraction is \(0.7642\), giving an updated value of approximately \(0.1354\). In the contrast direction, the discovery value is zero and the fraction is \(0.03846\), giving approximately \(0.00272\).

Transforming back to the two variant weights gives \(w_F\approx((0.1354+0.00272)/\sqrt{2},(0.1354-0.00272)/\sqrt{2})^T=(0.0977,0.0938)^T\), with small differences from rounding. The weakly identified contrast direction is strongly shrunk toward the discovery weights, while the common direction receives more target-data influence. This example is hypothetical and illustrates why LD-aware updating matters; marginal estimates should not simply be substituted for joint effects.

Target Sample Size, Overlap, and Overfitting

Larger target samples usually reduce sampling error and can make fine-tuning more useful. Yet the relevant quantity is not sample size alone. The effective information also depends on phenotype noise, allele frequencies, LD, variant coverage, and the number of weights being updated. If many weights are estimated from a small target sample, a flexible update can fit noise even when each individual GWAS estimate appears plausible. Shrinkage, dimension reduction, or a restricted set of update parameters can reduce this risk, but their tuning choices also need evaluation.

Overlap matters because the target and discovery estimates may share participants. When their errors are correlated, a calculation that treats the target evidence as independent of the discovery weights can count the same information twice. The following scalar calculation shows the size and direction of the issue for a simple blend.

Worked Example: Overlap Raises the Variance of a Naive Blend

Suppose a discovery estimate and a target estimate each have variance \(0.04\), and their sampling-error correlation is \(0.50\). The covariance is therefore \(0.50\sqrt{0.04\times0.04}=0.02\). Consider the equal-weight blend \(\tilde\beta=0.5\hat\beta_D+0.5\hat\beta_T\). These values are hypothetical.

Its variance is \[ \operatorname{Var}(\tilde\beta) = 0.5^2(0.04)+0.5^2(0.04)+2(0.5)(0.5)(0.02) = 0.01+0.01+0.01 = 0.03. \]

If one incorrectly assumed independence, the covariance term would be omitted and the reported variance would be \(0.02\). The actual variance under the stated overlap is \(0.03\), 50% larger than that incorrect calculation. A principled analysis should account for the overlap covariance or use an estimation design that makes the evidence independent.

A separate concern is evaluating performance on the same target individuals who guided the weight update. The updated score was chosen partly to fit their outcomes, so its apparent performance can be optimistic. The strongest check is an independent target-ancestry validation sample that was not used to estimate or select the fine-tuning procedure. If data are limited, prespecified sample splitting or cross-validation can help estimate out-of-sample performance, but the validation design must reflect every step that used outcome information.

What Fine-Tuning Can and Cannot Establish

Fine-tuning changes genetic weights, and that change can alter score variance, ranking, and the relationship between the score and the outcome. It may improve prediction, have little effect, or reduce prediction in the target population. A weight update alone does not establish calibrated risk, guarantee a desired calibration slope, or demonstrate clinical usefulness. The updated score should be evaluated for its intended prediction target and outcome scale using suitable independent data.

In particular, it would be incorrect to say that fine-tuning cannot change a score's calibration slope. It can: changing the weights changes the score, and therefore can change its observed slope against the outcome. The correct distinction is that fine-tuning does not, by itself, calibrate the score or guarantee any particular slope. Calibration should be assessed separately rather than inferred from the fact that target GWAS data were used.

Key takeaway: Fine-tuning balances discovery information against target-ancestry evidence. The balance should reflect target-estimate precision, LD, uncertainty in the discovery weights, and any overlap. The resulting score requires independent evaluation; updating weights is not evidence by itself that target performance or calibration improved.

Common Mistakes

Replacing discovery weights with every target estimate. Noisy estimates, especially in a small target GWAS, can add substantial estimation error. Shrinkage or another regularized update may be preferable.
Treating marginal GWAS effects as joint effects. Under LD, marginal effects reflect correlated variants; the expected marginal-effect vector is \(R_T\beta_T\) in the stated linear model.
Assuming overlap is harmless. Shared participants can correlate estimation errors. Ignoring that covariance can make a blend appear more precise than it is.
Evaluating the update on its training sample. Reusing the target outcomes that informed the weights can overstate prediction gains. Use an independent validation design or account for all tuning in resampling.
Equating updated weights with calibrated risk. Weight changes can alter calibration, including the calibration slope, but do not guarantee a desired calibration result. Assess calibration separately.

Check Your Understanding

Use the models and examples above to reason through each question.

Under the normal shrinkage model, what happens to the influence of target estimates when their sampling variances decrease, holding the prior fixed?
Why are target GWAS marginal effects generally not equal to joint genetic effects when variants are in LD?
In the overlap example, what covariance term must be included in the variance of a weighted blend?
Give one reason a large target sample may still provide weak information about particular fine-tuned weights.
Why is evaluation on the same target participants used for fine-tuning potentially optimistic?
Can fine-tuning change a score's calibration slope, and does it guarantee a chosen slope?