A genome is no longer hard to read. A clinical exome comes back in days, and the raw cost of sequencing has fallen faster than almost any technology in medicine. What has not fallen is the cost of deciding what the result means for the patient sitting in front of you.
That gap — between generating variants and interpreting them — is where bioinformatics actually lives, and where machine learning has started to earn its place.
Where models help in the pipeline
A clinical genomics pipeline is a chain of judgement calls dressed up as file formats. Models are useful at several of the links, though for different reasons.
Variant calling
Distinguishing a true variant from a sequencing artefact used to be the domain of hand-tuned statistical filters. Deep-learning callers now treat the pileup as an image-like tensor and learn the difference from labelled truth sets. The gain is largest exactly where the old filters were weakest: indels, homopolymer runs, and low-complexity regions.
Variant effect prediction
Predicting whether a missense change disrupts a protein is a classic supervised problem, and protein language models have moved the state of the art considerably. But the honest framing matters: these tools produce a prior, not a diagnosis. A high pathogenicity score is a reason to look harder, not a reason to report.
Prioritisation and phenotype matching
This is where the practical time is won. Given a patient's phenotype and a list of several hundred plausible variants, a model that ranks them against curated gene–disease relationships turns a day of manual triage into an afternoon of review.
- Ingest variants alongside structured phenotype terms.
- Rank candidates against curated knowledge, not just raw scores.
- Explain every ranking with the evidence that produced it.
- Log the pipeline version and reference build behind each result.
The three things that break clinical genomics AI
Reference and version drift. A variant interpreted against one reference build and one transcript set is not the same variant interpreted against another. Pipelines that do not pin and record their references will eventually produce results nobody can reproduce — and an irreproducible genomic result is not a clinical result.
Population bias. Most large variant databases over-represent individuals of European ancestry. A model trained on them will be systematically less useful, and less safe, for everyone else. In Türkiye, where consanguinity patterns and founder variants differ meaningfully from those reference populations, this is not a theoretical concern.
Silent reclassification. Gene–disease relationships change as evidence accumulates. A variant reported as uncertain in 2024 may be benign today. Any serious platform needs to know which patients were told what, and when, so a reclassification can be surfaced rather than buried.
A pathogenicity score is a hypothesis about biology. The report is a statement to a family. Never let the pipeline blur the two.
What good looks like
The genomics platforms that survive contact with a real clinical laboratory tend to share the same unglamorous properties. Every result is traceable to a pipeline version, a reference build, and a knowledge-base snapshot. Every model output carries calibrated uncertainty rather than a bare number. Variant curation stays with the clinical scientist, and the software's job is to bring them the right five variants instead of the wrong five hundred.
None of that is about better models. It is about building the scaffolding that lets a good model be trusted — which, in genomic medicine, is the whole job.