Target prioritization frameworks have become more sophisticated over the past decade. Genetic association data, transcriptomic target validation across tissue types, and proteomics-based selectivity profiling all get factored in. The computational infrastructure for target scoring has matured considerably. But the single most predictive variable for Phase II success, the tractability of the target in the subpopulation that will actually be enrolled, is almost never part of how targets get ranked.
The reason is not ignorance of the principle. Drug hunters understand that different patient populations respond differently to the same mechanism. The problem is operational: the tools for characterizing population heterogeneity relative to a specific target have been difficult enough to use that they rarely get integrated into the target selection workflow. By the time a development team is thinking carefully about which patients a target works in, they are designing Phase II, not choosing among candidates.
The cost of that sequencing is substantial. A target that shows strong aggregate genetic and molecular evidence but only generates meaningful clinical signal in a subpopulation comprising 30% of the enrolled patient group is structurally harder to develop than a target with more modest aggregate evidence but strong signal uniformity across the enrolled population. Aggregate prioritization frameworks cannot distinguish between these cases.
What aggregate signal strength actually measures
Genetic association studies measure the association between a variant and an outcome across the entire studied population. A high effect size in a well-powered GWAS or proteomics study means the association is strong and replicable. It does not mean the mechanism is active in all patients with the phenotype. Many robust genetic associations derive their statistical power from a subgroup with a specific genetic background or environmental exposure history, while being essentially null in the complementary subgroup.
The same problem appears in transcriptomic target validation. Differential expression analysis of disease versus healthy tissue aggregates over a heterogeneous patient sample. If 40% of the disease population has elevated target expression due to a specific upstream driver, and 60% does not, the aggregate differential expression signal will be diluted. The target may still score adequately in a conventional framework. But the implication for trial design is that you are building a development program around a mechanism that is active in 40% of the population, without knowing that you need to find and enrich for that 40%.
Aggregate validation evidence therefore contains a systematic bias: it penalizes targets that are highly active in one subgroup relative to targets that have modest activity across the full population, even when the subgroup-specific activity is pharmacologically preferable for a drug development program that can identify and enrich that subgroup.
The population heterogeneity model as a target selection input
At Valinor Discovery, we think of population heterogeneity modeling as an additional layer of the target scoring framework, not a replacement for conventional evidence. The question it answers is: given everything known about the target's mechanism and the disease population, what is the likely distribution of patients across response-relevant molecular and clinical subgroups?
This requires a model of the patient population for the indication, built from the kind of multimodal reference data that Bayesian cohort reconstruction can assemble. You map the target's known mechanisms of action onto the molecular and cellular phenotype dimensions captured in that population model. Subgroups where the upstream pathway driving target expression is active become visible as clusters in the virtual patient distribution. Subgroups where the mechanism is absent or redundantly compensated by alternative pathways become separately identifiable.
The output is not a binary "target works or doesn't." It is a characterization of the patient population into response-relevant clusters, with estimates of their relative sizes and their identifiability from clinical variables that could serve as enrichment criteria. A target that works well in 35% of patients but where that 35% can be identified by genomic stratification is categorically different from a target that works in 35% of patients distributed randomly across the clinical population without any identifiable enrichment handle.
Where heterogeneity analysis changes a prioritization decision
The cases where population heterogeneity modeling is most likely to change a prioritization call are not the obvious cases. If a target's mechanism is restricted to a patient subgroup that can be identified by a well-established companion diagnostic, that is already factored into the development plan. The cases that get miscategorized by aggregate scoring are subtler.
Consider a program where two candidate targets have comparable aggregate efficacy signal and safety profiles. Target A has strong, uniform signal across the indication population. Target B has a stronger signal in a definable subgroup comprising roughly half the population, and weak signal in the other half, but both halves are spread across the current standard-of-care treatment categories and molecular subtype classifications that are typically used as enrichment criteria.
A conventional prioritization framework will likely score them comparably, possibly slightly favoring Target A on signal consistency. But if the team can build a population heterogeneity model showing that Target B's responsive subgroup is identifiable by a combination of genomic variant frequency and a disease-severity-related biomarker, the development economics change. The trial can be designed with a stratification criterion that concentrates the responsive patients, changing the effective signal density in the enrolled cohort. That is the kind of developability input that shifts a prioritization decision and is invisible to aggregate-scoring approaches.
The integration challenge and what it realistically requires
We want to be direct about the operational challenge here. Integrating population heterogeneity modeling into target selection requires having a reference population model for the indication at the time targets are being evaluated, not months later when protocol design begins. For most drug discovery organizations, that is a significant process change. Population modeling has typically happened in the clinical development function, after the target has been committed to and a development candidate identified.
The organizational friction is real. We are not suggesting it is trivial to move population modeling upstream. What we are arguing is that the information value of doing so is high enough to justify the effort, particularly for programs in heterogeneous diseases where the probability of enrolling the wrong population in Phase II is already elevated. The question is not whether population heterogeneity modeling is theoretically useful at target selection. It clearly is. The question is whether the organizational cost of front-loading that analysis is less than the cost of running a Phase II trial that fails because the enrolled population diluted a real signal in a non-enriched subgroup.
For development programs in oncology, rare metabolic disease, and complex inflammatory conditions, we believe the calculus strongly favors early modeling. The trial costs in those indications are high enough that even a meaningful probability shift in Phase II success probability justifies the investment in characterizing population heterogeneity before committing to a target.