The study tackled a persistent problem in epilepsy drug discovery. Human genetics is one of the most effective routes to identifying new drug targets — targets with genetic support show roughly a two-fold higher probability of progressing to approval — but previous epilepsy target-discovery efforts have relied on expression measured in blood plasma or in bulk brain tissue, where signals from millions of different cells are averaged together. These approaches have yielded inconsistent candidate lists, with different studies nominating entirely different genes. The new work set out to test whether this instability has a structural explanation: that a gene detectable in one cell type is simply missed when expression is measured in another.

To do this, Gul took the genes prioritised at the genome-wide significant loci of the largest epilepsy meta-analysis, conducted by the ILAE Consortium on Complex Epilepsies, and screened 23 of them using a two-step genetic design. First, drug-target Mendelian randomisation (MR) was used to ask whether genetically predicted expression of each gene was associated with epilepsy risk. MR exploits the random allocation of alleles at conception to estimate causal effects largely free from the confounding and reverse causation that plague observational studies. When the gene encodes a druggable protein, the MR estimate further predicts the effect of modulating that target with a drug. But MR alone can be misled: two nearby variants inherited together can produce a spurious association even when one controls the gene and a different one affects disease. So the second step applied Bayesian colocalisation, testing whether the expression signal and the disease signal at each locus were driven by a single shared causal variant. A gene was retained only when both steps agreed.

The crucial innovation was the resolution at which expression was measured. Rather than relying on bulk tissue, the study drew on SingleBrain, a single-nucleus RNA-sequencing brain eQTL meta-analysis of four European-ancestry cohorts, providing expression quantitative trait loci (eQTL) for inhibitory neurons, excitatory neurons, and four glial cell types. Bulk cortex eQTL came from an independent resource, MetaBrain. Each gene was assessed across all seven panels, allowing a direct comparison of colocalisation posteriors across resolutions while holding the epilepsy outcome and colocalisation window fixed.

Of the 23 genes screened, MR nominated ten as having a significant effect on epilepsy risk. Colocalisation retained only two. The first, KCNIP2, produced the study’s most striking result. Its expression in inhibitory neurons colocalised with genetic generalised epilepsy at a posterior probability of 0.95 — well above the 0.8 threshold — and higher expression was associated with lower risk, at an odds ratio of 0.64 (p = 2.5 × 10−11). In bulk cortex, however, the same gene carried a strong expression signal that mapped to a completely different causal variant, with the posterior for a shared variant falling to zero. The bulk-cortex eQTL lead variant was in negligible linkage disequilibrium with the inhibitory-neuron lead (r² = 0.016), and the inhibitory-neuron lead carried no detectable effect on KCNIP2 expression in bulk tissue. The two resolutions were measuring different regulatory events, not one diluting the other.

This pattern offers a mechanism for the weak and inconsistent results of earlier analyses. Bulk tissue does not merely average away a cell-type-restricted signal — at this locus, it replaces the neuronal signal with a separate variant that does not colocalise with epilepsy at all. A plasma or bulk-tissue screen would discard KCNIP2 entirely. The finding was formally tested rather than inferred: the full glial summary statistics showed that KCNIP2 carries a genuine cis-eQTL in oligodendrocytes, with a directionally concordant but sub-threshold colocalisation posterior of 0.53, while three other glial panels carried no eQTL reaching their discovery thresholds. Cochran’s Q across the seven panels confirmed significant heterogeneity carried by inhibitory neurons against the rest.

The second retained gene, AP3D1, behaved differently. It colocalised with genetic generalised epilepsy across inhibitory neurons, excitatory neurons, and bulk cortex alike (posteriors 0.94, 0.94, and 0.92), with higher expression associated with increased risk (OR 1.46). Its pattern was set by the phenotype rather than by resolution — for all epilepsy, it did not colocalise in any panel. Both genes were then tested in FinnGen, an independent Finnish biobank. The directions of effect reproduced at the phenotype-matched generalised-epilepsy endpoint, with KCNIP2 protective (OR 0.51) and AP3D1 risk-increasing (OR 1.67). Colocalisation itself did not reproduce in FinnGen, which the study attributes to the smaller Finnish case count and different registry ascertainment rather than a contradiction of the discovery result.

The biology of KCNIP2 lends the genetic finding particular weight. The gene encodes KChIP2, an EF-hand calcium-sensing accessory subunit of the Kv4 A-type potassium channel. KChIP2 raises the A-type potassium current, and its deletion reduces that current, increases neuronal excitability, and raises kindling-induced seizure susceptibility in mice. The protective direction estimated by MR — higher expression, lower risk — is exactly what this channel biology predicts. Human genetics implicates the same channel from the pore-forming side as well: a de novo variant in KCND2, which encodes the Kv4.2 subunit that KChIP2 modulates, alters A-type-current inactivation and causes intractable seizures. The Kv4-KChIP2 complex is a structurally resolved target, making positive modulation of the A-type current the candidate on-target therapeutic direction. Yet the study is careful to note that KChIP2 also regulates the equivalent potassium current in cardiac muscle, so a potentiating strategy carries an on-target cardiac liability, and the closely related subunit family complicates selective targeting.

The study is equally candid about its limits. KCNIP2 cannot be separated from its neighbour CUEDC2 on genetic evidence alone: the two share a single inhibitory-neuron regulatory variant, with their strongest eQTL variants in near-complete linkage disequilibrium (r² = 0.975), and conditioning on the shared variant removed the epilepsy association for both. KCNIP2 is nominated on the strength of its genome-wide significant eQTL, its larger expression effect, and the established channel biology — not because CUEDC2 has been excluded. Every statement about KCNIP2, the study notes, should be read as a statement about the KCNIP2/CUEDC2 locus. AP3D1, meanwhile, encodes a subunit of the intracellular AP-3 vesicle-trafficking adaptor, a protein class with no catalytic ligand pocket and therefore low small-molecule tractability, leaving it a supported epilepsy gene but a weak drug-target candidate.

The demonstration that cell-type resolution determines detectability rests mainly on this single gene, and colocalisation has power only at genome-wide significant loci, so the discovery space was necessarily small. No functional work was performed, and the causal role of the two genes in seizure liability remains to be established. Still, the proof of principle is clear: at this locus, single-cell resolution is what made the signal detectable. Cell-type-restricted candidates have now been reported in ischaemic stroke and migraine as well, suggesting the approach may generalise. For a field where a third of patients lack effective treatment, the message is that the search for new epilepsy drug targets may need to look not just at the right genes, but at the right cells.

Subject of Research: Genetics

Article Title: Cell-type-resolved colocalisation prioritises KCNIP2 as a candidate epilepsy drug-target detectable at inhibitory-neuron but not bulk-tissue resolution: a Mendelian randomisation and colocalisation study

Article References: Gul, A. (2026). Cell-type-resolved colocalisation prioritises KCNIP2 as a candidate epilepsy drug-target detectable at inhibitory-neuron but not bulk-tissue resolution: a Mendelian randomisation and colocalisation study. Frontiers in Genetics, 17, Article 1920945. https://doi.org/10.3389/fgene.2026.1920945

Image Credits: AI Generated

DOI: 10.3389/fgene.2026.1920945

Keywords: epilepsy, KCNIP2, Mendelian randomisation, colocalisation, inhibitory neurons, drug targets, genetics, eQTL, Kv4 potassium channel, CUEDC2, AP3D1, FinnGen

The research, led by Xin Gao and Xiaofeng Guo of Shanxi Provincial People’s Hospital together with Shaopeng Yang of Zhejiang Cancer Hospital and corresponding author Qi Zhang of Shandong Cancer Hospital and Institute, describes a protein-prioritised, colocalisation-informed framework for generating target-outcome safety hypotheses. Rather than waiting for post-marketing surveillance to reveal unexpected adverse events, the team asked whether human genetics could flag plausible safety concerns for eighteen genes that represent established or emerging inflammatory bowel disease therapeutic targets. The logic is elegant: if genetically lowered levels or expression of a drug’s target protein are associated with increased risk of a particular disease outcome in population data, then pharmacologically inhibiting that same target deserves closer scrutiny for a similar effect.

The technical architecture of the study is where much of its novelty lies. The authors assembled twenty-five sets of genetic instruments, drawn either from cis-protein quantitative trait loci, which are variants that alter circulating protein levels, or from cis-expression quantitative trait loci, which are variants that alter gene expression in blood. These proxies covered the eighteen therapeutic genes and were tested against seventy-six prespecified safety outcomes drawn from the FinnGen R13 biobank release, a Finnish resource combining genetic data with national health registry records covering conditions from autoimmune disease to venous thromboembolism. In total, 1,900 tasks were planned, of which 1,824 proved analysable, generating a matrix of target-outcome hypotheses that could be evaluated in a statistically disciplined way.

A key methodological challenge was sparsity. Stringent selection of cis-acting variants, which are preferred because they are less prone to confounding by distant genetic effects, often leaves only one or a handful of usable instruments per gene. Conventional Mendelian randomisation methods that require many variants are therefore poorly suited to this setting. The workflow adapted by combining single-variant or few-variant Wald ratios with inverse-variance-weighted estimates where multiple variants were available, and then applying rigorous false-discovery-rate control. A global Benjamini-Hochberg correction across all 1,824 estimates served as the primary filter, with secondary corrections applied within forty narrower outcome domains. This two-tier approach balanced statistical conservatism against the risk of missing genuine signals buried in the multiple-testing burden.

The criteria for advancing a hypothesis were deliberately strict. An estimate had to pass the global false-discovery threshold, point in a risk-increasing direction under simulated target lowering, and lie outside the major histocompatibility complex, a densely variable genomic region notorious for producing misleading associations because of its extreme linkage disequilibrium. Signals meeting these conditions entered an objective follow-up phase involving sensitivity analyses, instrument-centred colocalisation and variant-level annotation. Colocalisation is the critical step: it asks whether the genetic variant associated with the molecular proxy, such as a protein level, and the variant associated with the disease outcome are in fact the same causal variant, rather than two nearby variants inherited together. The team used Bayesian colocalisation methods in windows of plus or minus 250 kilobases centred on each instrument, requiring a posterior probability of a shared causal variant above 0.5.

From the initial screen, one hundred estimates passed global false-discovery control and twenty more met only the narrow-domain threshold, producing fifteen tasks representing thirteen distinct target-outcome pairs for follow-up. The final classification stratified hypotheses by molecular layer, distinguishing protein-level P1 and P2 categories from expression-level E1 and E2 categories depending on whether colocalisation support exceeded the threshold. This stratification matters because a hypothesis supported at the protein level, where the drug will actually act, carries different translational weight than one supported only at the messenger RNA level, where the relationship between expression and protein abundance is often imperfect.

Two protein-level P1 hypotheses emerged as the strongest safety signals. The first linked lower ITGB7, the integrin targeted by vedolizumab, a widely used gut-selective biologic, with increased risk of multiple sclerosis and demyelinating disease, showing an odds ratio of 2.35 with a global q-value of 0.003 and colocalisation support of 0.935. The second connected TNFRSF1A, a tumour necrosis factor receptor, with myocardial infarction, an association replicated across two independent proteomic resources, the UK Biobank Pharma Proteomics Project and deCODE, with odds ratios of 1.56 and 1.37 respectively, both driven by the same variant, rs4149584, which encodes the R92Q amino-acid substitution. The convergence of two independent protein datasets on the same variant and the same outcome lends considerable credibility to the signal.

At the expression level, two E1 hypotheses stood out, both involving psoriasis, an outcome of particular interest because psoriasis shares inflammatory pathways with inflammatory bowel disease and is itself a comorbidity of the condition. Reduced IL23A expression was associated with dramatically increased psoriasis risk, with an odds ratio of 4.13 and an exceptionally strong q-value of 7.47 times ten to the minus nine, while reduced TYK2 expression showed an odds ratio of 1.53 with a q-value of 1.10 times ten to the minus eleven and near-perfect colocalisation support of 0.984. These findings are biologically coherent: IL-23 and TYK2 are the molecular foundations of some of the most successful recent therapies in both psoriasis and inflammatory bowel disease, and the genetic data suggest that their pathways are deeply intertwined with psoriasis susceptibility in ways that clinicians managing patients with both conditions should note.

Not every candidate signal survived scrutiny, and the framework’s treatment of these failures is as informative as its successes. IL12B-psoriasis and TNFSF15-lower-extremity deep-vein thrombosis were classified as P2 hypotheses because colocalisation posterior probabilities fell below the 0.5 threshold, meaning the apparent associations could not be confidently attributed to a shared causal variant. This is precisely the kind of auditable, transparent triage that distinguishes a disciplined hypothesis-generation framework from a fishing expedition. Every step, from the 1,900 planned tasks through the 2,584 SNP-level rows, heterogeneity analyses, weighted-median estimates and leave-one-out sensitivity checks, is documented in supplementary workbooks, allowing independent researchers to audit the provenance of each conclusion.

The broader significance of this work lies in its positioning within the drug development pipeline. Safety signals that emerge from human genetics are not deterministic predictions; they are prioritised hypotheses that can direct mechanistic validation in experimental systems and targeted evaluation in clinical safety datasets before large populations are exposed. By stratifying evidence by molecular proxy type, enforcing locus specificity through colocalisation, and excluding the confounded major histocompatibility complex region, the framework produces a short, defensible list of target-outcome pairs that merit attention. For a therapeutic area where biologics have transformed patient outcomes but questions remain about long-term risks such as demyelination, infection and cardiovascular events, the ability to generate such lists computationally, using only publicly available summary-level data, represents a meaningful step toward safer and faster drug development. The authors’ emphasis on auditability and reproducibility suggests the framework could be extended to other disease areas, turning the human genome into a routine preclinical screening tool for pharmacovigilance.

Subject of Research: Genetic prediction of safety outcomes for inflammatory bowel disease therapeutic targets using Mendelian randomisation and colocalisation

Article Title: A protein-prioritised, colocalisation-informed genetic framework for safety hypothesis generation across inflammatory bowel disease therapeutic pathways

Article References: Gao, X., Guo, X., Yang, S., & Zhang, Q. (2026). A protein-prioritised, colocalisation-informed genetic framework for safety hypothesis generation across inflammatory bowel disease therapeutic pathways. Journal of Translational Medicine. https://doi.org/10.1186/s12967-026-09068-z

Image Credits: AI Generated

DOI: 10.1186/s12967-026-09068-z

Keywords: inflammatory bowel disease, Mendelian randomisation, colocalisation, pQTL, eQTL, drug safety, FinnGen, ITGB7, TNFRSF1A, TYK2, IL23A, pharmacovigilance