Oncology drug development has a well-documented attrition problem. The majority of agents that enter clinical trials do not proceed to approval, and a substantial fraction of those failures occur early in development when target-related hypotheses are still being tested. The pattern is persistent across tumor types, mechanistic classes, and development eras. Preclinical data that looks compelling by conventional standards is frequently not predictive of clinical outcomes.
Understanding why this happens at the target level, rather than at the drug level, is important for anyone working on target identification. Some failures are pharmacological: the right target, the wrong drug. Those are not the failures this post is about. This post is about failures where the target itself turns out not to be the right biological intervention point in human tumors, despite preclinical evidence that said it was.
Model-to-Human Translation Failures
The most discussed category of translational failure is the discordance between preclinical models and human tumor biology. The reasons for this discordance are multiple and worth examining specifically.
Cell-line models are the most common preclinical starting point for target validation. They are fast, tractable, and well-characterized. They are also, by definition, adapted for growth in culture conditions that share relatively little with the in-vivo tumor microenvironment. They lack stromal interactions, immune infiltrate, vascular architecture, and the hypoxic gradients that characterize solid tumors in vivo. Targets that are critical for cell survival in standard two-dimensional culture can be redundant or even irrelevant in the in-vivo context where stromal survival signals are available.
Standard xenograft models in immunocompromised mice address the three-dimensional tumor architecture problem partially. They do not address the immune component at all, which is a significant limitation for targets whose relevance is tied to the tumor-immune interaction. They also frequently use cell lines propagated extensively in culture, which drift substantially from the genetic and phenotypic profile of primary tumors over time. Patient-derived xenograft models preserve more of the original tumor's heterogeneity and are more predictive than cell-line xenografts, but they are slower and more resource-intensive, which limits their use early in target screening.
The Publication Landscape Problem
There is a systematic bias in the preclinical literature that compounds the model-to-human translation issue: positive findings are published at much higher rates than null or negative findings. This means that the published evidence base for a given target is substantially more favorable than the true empirical record of experiments performed. A target that has fifteen published positive papers and twelve unpublished negative results looks like a well-validated target if you only see the literature. It is something quite different in the full experimental record.
This publication bias has been recognized for decades in medicine and is particularly acute in preclinical oncology research. It means that the evidence synthesis problem is not just a quantity problem (not enough synthesis capacity) but also a quality problem: the literature itself provides a systematically distorted signal. A team doing evidence synthesis is, to a degree, working with a biased dataset regardless of how thorough they are.
This is not an argument that literature synthesis is useless. The positive signal is real, even if the null signal is hidden. A target with strong, replicated positive evidence from diverse experimental systems is genuinely better supported than a target with only weak or non-replicated positive evidence. But it does mean that the evidence quality bar should be set higher than the literature count alone suggests. Targets need corroborating evidence from independent sources, not just accumulation of positive findings from what may be a single research group or methodology.
Clinical Context Misalignment
A different failure mode involves the relationship between the target's preclinical evidence and the clinical context in which it is being pursued. Targets can have legitimate preclinical validation in a laboratory setting but be advanced into the wrong clinical context: the wrong tumor type subpopulation, the wrong patient selection criteria, or a biological context where compensatory mechanisms are active that were not present in the preclinical model.
Preclinical models often use tumor cell lines or animal models that are genetically uniform relative to the heterogeneous human tumor populations that enter clinical trials. A target that is highly relevant in a cell line carrying a specific sensitizing alteration may produce a much weaker clinical signal when studied in an unselected patient population where the sensitizing alteration is present in only a subset of enrolled patients. If the clinical signal in that subset would have been strong, the study was designed to dilute it.
This kind of clinical-preclinical misalignment is partially addressable through better patient stratification design, but it starts at the target identification stage. Evidence that a target's relevance is subtype-specific or alteration-dependent should be visible in the evidence base before a development program is designed. If that signal is present but not synthesized, programs advance with the wrong assumption about how broadly their target evidence applies to the intended patient population.
What Better Evidence Synthesis Can and Cannot Fix
Improved evidence synthesis at the target identification stage is one lever on the translational problem. It is not the whole solution. Some translational failures are genuinely unpredictable from the available evidence because the biology of the specific human tumor context was not present in any accessible preclinical system. Immune-mediated resistance mechanisms in immunocompetent patients, for example, cannot be predicted from immunocompromised mouse models regardless of how thorough the synthesis of those models' results is.
What better evidence synthesis can do is reduce the frequency of failures that were preventable with the evidence that existed. When targets fail because their preclinical evidence was concentrated in a single non-representative model, or because contradictory in-vivo evidence existed in the literature but was not retrieved and weighted appropriately, or because the clinical context alignment was absent and that absence was not made visible during target selection, those are failures that a structured evidence assessment might have caught.
We built the Avenzo evidence framework around this specific distinction: separating the preventable translational failures (those traceable to inadequate evidence synthesis) from the genuinely unpredictable ones. The framework cannot eliminate the latter. It is designed to reduce the former by surfacing evidence quality and breadth information before commitment, rather than after the assay campaign is underway.
The Evidence That Should Raise Confidence
Based on what translational failures teach us about the limitations of preclinical evidence, the evidence characteristics that should carry the most weight in a target selection decision are: functional validation in multiple independent experimental systems spanning at least one in-vivo model; consistent directional signal across both preclinical and clinical context evidence; evidence from models that reflect the target population's alteration profile; and the absence of contradictory in-vivo findings.
None of these criteria guarantee clinical success. The translational gap is partly a biological reality that no target selection process can fully eliminate. But targets that meet these criteria have a different failure probability distribution than targets that do not. The failures that remain are more likely to be genuinely unforeseeable biology than preventable information gaps.
That shift in the distribution of reasons for failure is the value proposition for rigorous evidence synthesis at the front of the discovery pipeline. If you are working on oncology target selection and want to understand how your current candidates compare against these evidence criteria, the Avenzo platform is designed to surface that analysis. Reach out to discuss your program context.