Target Identification and Prioritization
Target identification is a decision under uncertainty. AI systems help organise evidence, but the scientific risk is confusing association with therapeutic tractability. Open Targets is the de facto open reference for evidence integration; AI adds ranking, representation learning, and literature mining at scale. A prioritized target is rarely just the top-ranked node. It is the target with a testable mechanism, feasible modality, safety rationale, and disease-relevant assay path.
Use this chapter to:
- Prioritize therapeutic targets by combining genetics, biology, tractability, safety, modality fit, and translational rationale.
- A target score is not a target hypothesis unless it states disease context, direction of effect, mechanism, safety rationale, and feasible modality.
Prerequisites: Evaluation Principles for Life Sciences AI for calibration and split discipline; Variant Effect Prediction for genetics-evidence quality.
Summary: Prioritize therapeutic targets by combining genetics, biology, tractability, safety, modality fit, and translational rationale. Evidence aggregation and genetics-informed ranking are useful; reliable target selection without experimental or human-genetic support is not established.
Key point: A target score is not a target hypothesis unless it states disease context, direction of effect, mechanism, safety rationale, and feasible modality. Open question: whether ranked targets become modality-ready hypotheses with safety, directionality, and disease-context support.
Bottom line: Target discovery connects genomics, systems biology, disease biology, modality selection, chemical biology, trials, and real-world evidence.
What is this field trying to solve? Prioritize therapeutic targets by combining genetics, biology, tractability, safety, modality fit, and translational rationale.
What is the core idea? A target score is not a target hypothesis unless it states disease context, direction of effect, mechanism, safety rationale, and feasible modality.
What is the current state of the field? Evidence aggregation and genetics-informed ranking are useful; reliable target selection without experimental or human-genetic support is not established.
What do we know, and what remains open? Known reference points include Open Targets, GWAS Catalog, ClinVar, gnomAD, UK Biobank, DepMap, CRISPR screens, knowledge graphs, Mendelian randomization, and pathway resources. What remains open is whether ranked targets become modality-ready hypotheses with safety, directionality, and disease-context support.
Why does this matter? Target discovery connects genomics, systems biology, disease biology, modality selection, chemical biology, trials, and real-world evidence.
Introduction
The Open Targets Platform integrates genetics, genomics, chemistry, literature, and drug evidence to support systematic target-disease prioritisation (Ochoa et al., 2021). AI adds ranking and representation learning, but the underlying problem remains biological evidence integration with appropriate weighting of causal versus associational evidence.
Human genetics is the strongest general prior because it links perturbation of a gene to human disease biology before a program begins. Approved drug indications are enriched for human genetic support (Nelson et al., 2015), but the genetics signal still has to be paired with tractability, modality fit, tissue expression, safety rationale, and disease-stage logic. Finan and colleagues framed the druggable genome as a target-identification substrate, not as an automatic route to medicines (Finan et al., 2017).
How a target hypothesis is built
A target hypothesis has more parts than a target score. The minimum professional argument names the disease biology, patient subgroup, intervention direction, expected biomarker movement, tractable modality, safety concern, and first experiment that would falsify the hypothesis. A knowledge-graph score can make that argument easier to assemble, but it cannot decide which evidence layer is causal and which is only correlated.
The Open Targets Platform is useful because it decomposes target-disease evidence into evidence types rather than presenting one undifferentiated recommendation. The 2025 Nucleic Acids Research update frames the platform as a therapeutic-hypothesis building system, not a replacement for disease-area review (Buniello et al., 2024). The graph helps organise genetics, expression, known drugs, literature, and tractability. The program team still owns mechanism, modality, and stopping rules.
Direction of effect is the common missing sentence
Genetic association does not by itself say whether to inhibit, activate, degrade, silence, replace, or modulate a target. Loss-of-function protection points in a different direction from gain-of-function risk. Expression elevation in diseased tissue may be causal, compensatory, or a downstream marker. Mendelian randomisation strengthens some causal claims, but it still requires attention to instrument validity, pleiotropy, effect size, and disease timing.
Recent genetics studies have tightened the target-to-program argument. Trajanoska and colleagues reviewed how human genetics can support the path from target discovery into clinical development when the evidence is read with mechanism and effect direction in mind (Trajanoska et al., 2023). Minikel and colleagues refined how genetic evidence relates to clinical success and emphasized that genetic support is more informative when the therapeutic perturbation mirrors the disease-protective or disease-risk mechanism (Minikel et al., 2024).
Safety genetics and target tractability
Target confidence also depends on why trials stop. A target with strong disease genetics may still fail because the intervention has an unacceptable safety margin, the tissue window is wrong, the modality cannot reach the target, or the downstream biology is redundant. Genetic factors associated with clinical trial stoppage provide a reminder that target evidence and safety evidence need to be read together, not sequentially after a program has already launched (Razuvayevskaya et al., 2024).
Tractability is therefore not a late chemistry question. For small molecules it includes pocket quality, selectivity, ADMET liabilities, and feasible exposure. For antibodies it includes accessible extracellular target biology, epitope specificity, developability, and Fc or format choice. For RNA or gene therapies it includes delivery, tissue specificity, durability, and reversibility. Target evidence should be paired with modality evidence because a target without a modality hypothesis is not ready for a program.
What is demonstrated?
Demonstrated capability includes evidence aggregation, target-disease scoring, literature mining, and genetics-informed prioritisation. Open Targets provides a transparent public example of systematic target prioritisation infrastructure (Ochoa et al., 2021). FDA also recognises AI and machine learning use across drug development submissions and discussion papers (FDA, 2026). Industrial knowledge-graph platforms (Iktos, BenevolentAI, others) and academic systems extend the evidence-integration approach.
Functional genomics then tests whether perturbing a target changes a relevant cell state. The Cancer Dependency Map used large-scale loss-of-function screens to identify selective cancer dependencies across cell lines (Tsherniak et al., 2017). That evidence is most useful when the disease model, lineage, genotype, and assay endpoint match the intended therapeutic context. A dependency in a transformed cell line is not automatically a safe target in a patient.
| Evidence Anchor | What It Supports | Practical Constraint |
|---|---|---|
| Open Targets | Integrated target-disease evidence | Scores require biological interpretation |
| Human genetics | Causal-leaning target-disease support | Pleiotropy, linkage, and effect direction still require review |
| Druggable genome mapping | Modality and tractability framing | Druggable does not mean disease-relevant |
| Cancer Dependency Map | Context-specific functional dependency | Cell-line dependency requires tissue and safety translation |
| Open Targets Genetics | Variant-to-gene-to-disease links | GWAS and rare-variant evidence quality varies |
| FDA drug AI materials | Regulatory attention to AI in drug development | Regulatory acceptance depends on context of use |
| Knowledge-graph platforms | Heterogeneous evidence integration | Source quality bounds graph quality |
What is theoretical?
Theoretical capability includes target selection models that forecast efficacy, toxicity, patient subgroup, and modality fit before a program begins. The causal and translational links remain difficult, especially when disease biology is heterogeneous. Multi-omics-informed target prioritisation that integrates genetics, transcriptomics, proteomics, and clinical data at production quality is an active area; published work covers selected indications and is not yet a general recipe.
What is beyond current capability?
Beyond current capabilities includes reliable target discovery without wet-lab or human genetics validation. Computational ranking alone does not establish disease causality or therapeutic window. Fully autonomous target discovery without expert biology and chemistry review remains beyond current capabilities.
What would make this more promising?
Target identification becomes more promising when a ranked association becomes a modality-ready therapeutic hypothesis with a falsification path.
| Claim | Evidence that raises or lowers confidence |
|---|---|
| “This target is associated with disease” | Evidence type, source, effect direction, tissue context, and uncertainty are separated rather than collapsed into one score |
| “This target is causal” | Human genetics, perturbation data, rescue logic, or disease-relevant functional genomics support the direction of intervention |
| “This target is tractable” | Modality fit, structural or accessibility evidence, delivery path, selectivity rationale, and safety biology are specified |
| “This target is program-ready” | Patient subgroup, assay path, biomarker, first no-go experiment, and rejected-target record are documented |
The strongest change is not a higher graph score. It is orthogonal evidence that connects disease mechanism, intervention direction, modality, and early stopping rule.
What should researchers, biotech teams, funders, and program leaders do with this?
- Separate evidence for disease association from evidence for intervention.
- Require a modality hypothesis before program launch.
- Use orthogonal evidence layers rather than one graph score.
- Write the expected perturbation direction. Loss-of-function genetics, gain-of-function biology, agonism, antagonism, degradation, and RNA knockdown do not imply the same therapeutic move.
- Record why lower-ranked targets were rejected; this is the most-overlooked discipline in target selection.
- Cite Open Targets evidence with the specific data type (genetics, expression, literature) rather than a single composite score.
- For genetics-informed targets, distinguish Mendelian randomisation, rare-variant burden, and common-variant GWAS evidence quality.