How I integrate

The evidence streamline

What would change my conclusion

A target nomination is only as good as the evidence that could have overturned it. Two genes from the bone CRISPRi screen, scored the same way:

SupportingConflictingMixedNot in this summary

Had we required chromatin contact, we would have discarded NCOA3, one of the four genes that held up in every assay. When layers disagree, the disagreement decides the next experiment.

Source: Conery, Pippin, Wagley, Trang et al., Genome Biology 2025. The bone story →

The clinical and population lens

Choose the question MD

Clinical training shapes the phenotype before any model runs: what counts as a case, which comparison matters to a patient, which outcome would change care.

Design for the population MPH

Reference scores calibrated across populations. Endpoint transformations chosen for the estimand. Sample sizes set from the analysis, not from convenience.

Protect participants

Consent and analysis plans that state what genetic information is and is not returned, how long data are kept, and who may access them.

Is the difference real?

Tissue studies compare diseased and healthy samples, but disease can shift a whole organ, not just the lesion. Move the sliders. Synthetic data, for illustration.

Difference-in-differences assumes that, without the disease, lesion and adjacent tissue would differ by the same amount in cases as in controls. When that holds, it isolates the lesion-specific change; the naive comparison cannot.

Practices

• QC rules for single-cell and spatial data written as automated pass/warn/fail checks, so a pipeline stops rather than quietly passes.
• Reference panels identity-checked before any LD-dependent estimate.
• Every quoted number traced to the file that produced it.
• Ongoing work: atlas-scale single-nucleus and spatial transcriptomics, ambient-RNA and doublet modelling, deconvolution benchmarking, drug-target Mendelian randomisation, paired chromatin-accessibility and 3D-genome analyses.