How artificial intelligence is changing each step of genetic testing, and why human judgement still has the final say.

Authors: Sameer Malik, Sana

IN BRIEF

For two decades the headline story in genetics was the falling price of sequencing. The quieter and more important story is what happens after sequencing finishes: turning noisy signals into letters, those letters into a list of variants, and that list into a sentence a clinician can act on. In the last few years, artificial intelligence has moved into every one of those steps. It now decides which letter a sequencer most likely saw, separates true variants from machine noise, predicts what an unstudied mutation does to a protein, reads a patient's face for clues to a syndrome, mines the scientific literature for evidence, and sometimes drafts the first version of the report as well. This article follows that journey from the raw read to the signed report, explains simply what AI does at each stage, and is just as clear about what it still cannot do. In genetic testing, the question is not only about how fast you can produce an answer but whether you can trust it.

THE MAP: THREE TIERS OF GENOMIC ANALYSIS

Genomic analysis is conventionally split into three tiers: Primary analysis: The first step is reading itself. The sequencer's signal is converted into DNA letters, each with a confidence score (the output is a FASTQ file). Secondary analysis: The second step is finding the differences. Those letters are matched against a standard reference genome to work out where the patient differs from it, producing a list of variants (a VCF file). Tertiary analysis: The third step is making sense of that list. Everything that turns that list into meaning: annotation, prediction, phenotype matching, classification, and the written report. In everyday language, the first step is sequencing, and the second and third together are interpretation. This article follows the same order. Part One looks at AI in the reading step. Part Two follows it through the two tiers of interpretation. AI is now part of all three steps, but its role, and the level of caution it needs, changes as we move from simply counting signals to making a clinical statement about a person.

Part One - Sequencing: Teaching a Machine to Read

It is tempting to think of a sequencer as a device that simply reads DNA, the way a scanner reads a barcode. It does not. A sequencer measures a physical proxy for each DNA letter and then has to guess the letter from that measurement. On an Illumina machine the proxy is a flash of coloured light; on an Oxford Nanopore device it is a tiny dip in an electrical current as the strand squeezes through a protein pore. Neither a colour nor a current is a letter. Something has to translate one into the other, and that translation step called basecalling is the first place where artificial intelligence enters the story.

Basecalling: from a wiggle to an A, C, G, or T

Early basecallers used hand-written rules and statistics. The hardest cases were long runs of the same letter (homopolymers), such as AAAAAA, where the signal barely changes and counting is difficult. Modern basecallers replaced those rules with neural networks trained on millions of examples where the true sequence was already known. The network learns, from the data itself, the subtle link between the shape of the signal and the letter that produced it. Oxford Nanopore's current basecallers (called the Dorado family) and Illumina's real-time analysis software are both, under the hood, deep learning models.

The gain is not cosmetic. Careful comparisons have found that AI basecallers clearly improved the accuracy of each read. The best modern configurations, paired with deep-learning variant callers, can now match or beat older methods even in regions that used to be very hard to read. This lets a laboratory trust these reads for variants it once could only confirm with a second, separate technology.

WHY THE READ IS NOT A GIVEN

Two reports can claim the same sequencing depth, meaning the same number of times each position was read, and still differ in quality. Depth counts how many reads you have, while basecalling decides whether each read is correct. A better basecaller improves accuracy without adding more reads. This is the first lesson of AI in genetics: much of the improvement is invisible, built into a step the clinician never sees that simply makes the letters more likely to be correct before anyone looks at a variant.

There is a right answer, the real DNA letter, and a model is judged on how often it matches that answer. No clinical judgement is involved, no ethics, no patient. It is pattern recognition in its simplest form, and it is now done better by a trained model than by a human-designed rule. Keep that in mind, because as we move into interpretation, the right answer becomes less clear and the stakes get higher.

Part Two - Interpretation: From Reads to Meaning

Once the sequencer has produced reads, the genome has been read but not yet understood. A whole-genome run yields four to five million positions where a person’s DNA differs from the reference DNA, and only a handful will matter for the question that prompted the test. Interpretation is the work of getting down from millions to a few, and then from those few to a defensible clinical answer. It splits into the two remaining tiers of analysis: secondary, which finds the variants, and tertiary, which decides what they mean. AI behaves very differently in each.

Tier Two (Secondary Analysis): Finding the Variants in the Noise

Secondary analysis has two jobs. The first is alignment: taking each short, broken read and finding where it belongs on the three-billion-letter reference genome. It is like emptying a shredded book onto a table and working out which page every scrap came from. The second is variant calling: once the reads are stacked at each position, deciding whether a difference from the reference genome is a genuine variant in the patient or just a sequencing error in one or two reads.

Variant calling used to be a statistics problem, solved with probability models that asked how likely a pile of reads was under “there is a variant here” versus “the machine slipped.” The breakthrough, from Google in 2018, was to stop treating it as statistics and start treating it as image recognition. Their tool, DeepVariant, draws the stacked reads at each position as a picture, coloured rows of letters piled on top of one another. It hands that picture to the same kind of neural network used to recognise faces in photographs. The network learns what a real variant looks like and what an error looks like, just as it would learn any other visual pattern.

The approach worked unusually well. On benchmark human samples, deep learning callers are highly accurate for single-letter changes, and the same network can be retrained for a completely different sequencing platform without redesigning the logic. For long reads, newer methods go further: they separate the two inherited copies of a chromosome before calling and recover high-quality variants in difficult regions where short-read methods simply failed. In several head-to-head benchmarks, deep learning callers on the best long-read data now exceed the accuracy of short-read standards, a reversal of a long-standing assumption in the field.

WHAT AI ADDS IN SECONDARY ANALYSIS

More variants recovered from hard regions. Repetitive and difficult stretches that defeated older methods are now within reach, shrinking the genome’s blind spots. Fewer false alarms. Learning the visual signature of errors reduces the false calls an analyst would otherwise have to check by hand. Works across machines. One architecture, retrained, serves different sequencing platforms like Illumina, PacBio, and Nanopore so improvements propagate across platforms instead of being rebuilt each time. A right answer still exists. As with basecalling, performance is measured against gold-standard reference samples, which is why this step has been adopted faster and with less anxiety than the ones that follow.

Tier Three (Tertiary Analysis): Turning Variants into Meaning

This is where genomics has always suffered bottlenecks and where AI is both most useful and also most fraught. The variant caller hands over a long, unlabeled list. Nothing on it is marked “disease-causing” or “harmless.” Tertiary analysis is the discipline of attaching meaning, and it breaks into four problems: filtering the list down, predicting what each remaining variant does, connecting variants to the patient's actual symptoms, and finally classifying, prioritising, and reporting. AI is now reaching into all four.

1. Filtering: from millions to a manageable few

The first cut is mostly rule-based and was already standardized before the current wave of AI. Discard variants too common in healthy populations to explain a rare disease, drop those in regions unlikely to affect a gene, and so on. Using population databases such as gnomAD and clinical archives such as ClinVar, this filtering reduces millions of differences to roughly ten to fifty candidates worth human attention. Machine learning increasingly sharpens this step by scoring how likely a variant is to break a gene, so the candidates arrive pre-ranked rather than in a flat list.

2. Prediction: What does an unstudied variant actually do?

For a large share of candidate variants, a direct lookup returns little or nothing useful, and not always for the same reason. Some are genuinely novel, never reported in any patient, paper, or database. Many more already sit inside gnomAD or ClinVar itself, observed but unresolved, too rare for the case reports, family data, or lab experiments a confident call needs to have accumulated yet. Either way, the question becomes a prediction: if this single letter change alters one amino acid in a protein, does the protein still work? The most striking recent advance here comes from protein language models. The idea borrows directly from the technology behind chatbots. A language model trained on ordinary text learns which words are plausible in a sentence. A protein language model trained on millions of natural protein sequences learns which amino acids are plausible at each position in a protein. If a variant swaps in an amino acid the model finds very unlikely, one that nature almost never allows in that spot, it is flagged as probably harmful.

DeepMind's AlphaMissense (2023), built on the protein-structure model AlphaFold, is the best-known example. It produces a pathogenicity score (prediction about whether a change is harmful or not) for nearly every possible single-amino-acid change in the human protein, and it performs strongly when tested against known cases. Notably, its makers deliberately avoided training it on clinician-assigned labels to keep it from simply memorising and repeating existing (and sometimes mistaken) database opinions. Related models can predict the effect of hundreds of millions of DNA changes, including across different protein isoforms, which are different versions of the same protein made from the same gene. Specialised tools such as SpliceAI check whether a variant may disturb splicing, the cell’s editing process that removes unnecessary gene parts and joins the useful parts before making a protein. If this editing process goes wrong, the final protein may be incomplete or faulty. CADD gives a score for how harmful a DNA change is likely to be across the whole genome.

These tools are excellent at narrowing the field. They are not a verdict. A high pathogenicity score is evidence to be weighed, not a diagnosis. A model that is right ninety-four times in a hundred can be wrong the other six times. In a clinic, each error is a real person.

Their limits are worth stating plainly, because they are easy to forget behind an impressive number. AlphaMissense and its kin handle single-amino-acid substitutions but not most insertions, deletions, or complex rearrangements. Their scores are not specific to a particular disease; they perform worse in disordered, unstructured stretches of protein, and as deep models they are hard to interrogate. They give a number without showing their reasoning. The standard ACMG classification framework treats such predictions as supporting evidence of modest weight, never as a stand-alone basis for calling a variant pathogenic, and that is the right approach.

3. Phenotype: reading the patient, not just the genome

A variant only matters to a person. The same gene change can be central to one patient's illness and irrelevant to another's. So interpretation has to weigh the genome against the patient's actual features, known as their phenotype. AI has moved into this side of the problem too, from two directions.

The first is next-generation phenotyping from images. Many genetic syndromes leave a characteristic mark on the face, subtle enough that even experienced clinicians miss it. Tools such as DeepGestalt, delivered through the Face2Gene app, use a neural network trained on tens of thousands of patient photographs to suggest candidate syndromes from an ordinary photo. In some comparisons, it outperforms expert clinicians on the syndromes it knows. Its successor, GestaltMatcher, maps each face into a mathematical space so that patients can be matched to one another even when their disorder is so rare it was never in the training data. This creates a route to recognising, and sometimes defining, new syndromes.

The second is phenotype-driven prioritisation i.e. ranking of variants from structured clinical data. A clinician records a patient's signs as standardised terms from the Human Phenotype Ontology (HPO), and tools such as Exomiser combine that profile with variant evidence to push the variants that best explain the patient's condition toward the top of the list. The effect is to combine two streams of evidence, what the genome says and what the body shows that older pipelines handled separately.

4. Classification, literature, and the report

The final stretch is classification under the ACMG/AMP framework, weighing up to twenty-eight evidence criteria to place a variant on the five-point scale from benign to pathogenic, followed by writing the report. AI helps here in two ways. Machine learning and rule-based tools pre-apply many of the ACMG criteria automatically, so an analyst starts from a structured draft rather than a blank page. And large language models of the same family of systems behind general-purpose chatbots are beginning to mine the scientific literature for scattered evidence about a gene or variant, summarize it, and draft the written interpretation that goes into the report.

This is the highest-leverage and the highest-risk use of AI in the whole process. Highest leverage because literature review and report drafting are exactly the slow, manual work that creates the interpretation backlog. Highest risk because language models can produce fluent, confident text that is simply wrong, a phenomenon usually called hallucination. A fabricated citation in a clinical report is far more dangerous than a single misread signal in variant basecalling. The pattern emerging in practice is narrow and supervised: the model gathers and organises evidence and suggests wording, and a qualified human verifies every claim against its source before anything is signed off.

THE PATTERN ACROSS ALL OF TERTIARY ANALYSIS

Notice the trend. As we move from basecalling to variant calling to interpretation, the right answer becomes less clear and the cost of error grows. In tertiary analysis there is often no single right answer, the evidence is incomplete, and the output is a statement about a human being. That is exactly why the role of AI shifts from deciding to assisting: narrowing the field, ranking the candidates, surfacing the evidence, and drafting the report while a clinician or clinical scientist retains the judgment and the accountability.

Where AI Touches the Pipeline: A Map

Pulling the stages together, the table below shows what AI is doing at each step and how much human oversight the step demands. The pattern is consistent: the closer the step gets to a clinical statement about a person, the more the balance tips back toward the human.

Stage	What AI does	Example tools/techniques	Human oversight
Primary: Sequencing	Converts raw signal into letters with confidence scores	Dorado, RTA (neural basecallers)	Low: clear right answer
Secondary: Finding variants	Aligns reads and separates true variants from machine noise	DeepVariant, Clair3, PEPPER-Margin	Low to moderate: benchmarked
Tertiary: Interpretation (predict)	Predicts whether a variant breaks a protein	AlphaMissense, ESM1b, SpliceAI, CADD	Moderate: supporting evidence only
Tertiary: Interpretation (phenotype)	Suggests syndromes from a face; ranks variants by symptoms	DeepGestalt, GestaltMatcher, Exomiser	Moderate to high
Tertiary: Interpretation (classify & report)	Applies ACMG criteria; mines literature; drafts text	Classification engines; large language models	High: every claim verified

Examples are for orientation only; the specific tools used vary by laboratory platform and assay. Oversight levels describe the degree of human review appropriate to each step, not the maturity of the technology.

What AI Still Cannot Do

A fair account of progress has to be equally honest about its edges. Several limits are not small difficulties that the next model will fix. They are built in, and they shape how AI should and should not be used in a clinical genetics laboratory.

The black-box problem

The most capable models are the least explainable. A deep model can produce a confident pathogenicity score without being able to show, in terms a clinician can check, why. For a regulated clinical decision that is a real obstacle. Clinical laboratories accredited under CAP/CLIA or ISO 15189 are expected to justify their conclusions and show exactly which evidence was used when a report is signed off. An answer that cannot be traced is hard to defend, however accurate it usually is. The AI systems that succeed in the clinic are increasingly the ones that show their work, that reveal the evidence behind a score rather than only the score.

Hallucination and the limits of fluency

Language models generate text that is plausible, not text that is guaranteed to be true. They can invent a reference, attach a finding to the wrong source, or state a relationship that does not exist, all in confident prose. In a clinical report which is a personal, heritable, consequential document, failure is unacceptable without verification. This is why generative AI in interpretation is being deployed as a drafting and evidence-gathering aid under mandatory human review, not as an autonomous author.

Bias, and the equity problem

An AI model is only as representative as the data it learned from, and the reference data behind clinical genomics is skewed. Population databases and variant archives have been built disproportionately from individuals of European ancestry. When tools trained on that data are applied to South Asian, African, Latin American, and other underrepresented populations, more variants come back as uncertain. This is not because those genomes are more pathogenic, but because they are not well catalogued. AI can inherit and even amplify this imbalance. For a country such as India, with an enormous and largely under-sequenced population, the implication is direct: importing models and databases built elsewhere is not enough. Closing the gap requires population-matched reference data and locally trained interpretation so that the tools see the diversity they are asked to read.

The VUS is not solved

Roughly half of the variants in ClinVar remain classified as Variants of Uncertain Significance, meaning there is not enough evidence to say whether they cause disease. AI has nudged that number by reclassifying some borderline cases, but it has not removed the problem, because the problem is fundamentally a shortage of evidence, not a shortage of computation. Where the underlying biology has never been studied, no model can create certainty from nothing. It can only estimate, and an estimate is not a fact. Resolving uncertainty still depends on slower work: functional experiments, family studies, and the steady accumulation of new cases. AI can speed that work up, but it cannot replace it.

THE ONE-LINE SUMMARY
Across the whole process, AI saves the expert's time. It does not replace the expert's judgment. The right test for any AI tool entering a clinical genetics workflow is not “is it impressive?” but “does it make a qualified human faster and more accurate without taking the decision, or the responsibility, out of their hands?”

Where This Leaves Us

The cost of generating genomic data fell long ago. What stayed expensive was understanding it, the time, expertise, and infrastructure needed to turn a signal into a sentence a clinician can trust. That gap, not the sequencing, has always been the real bottleneck in genomic medicine, and it is exactly the gap that artificial intelligence is now narrowing, one step at a time.

Seen across the whole process, the change is coherent rather than miraculous. AI makes the read more accurate before anyone looks at it. It pulls variants out of regions that used to be dark. It predicts the consequences of mutations no one has ever studied, reads a patient's face and symptoms alongside their genome, mines scattered literature, and drafts the report. Each step shaves time off a process that families used to wait years for. And at each step, as the work moves closer to a statement about a person, the technology's proper role narrows from making the decision to informing it. The clinician is the one who remains responsible. The future of genomic medicine is not simply faster sequencing; it is faster and safer understanding.

The answer to many patients' questions is already written in their DNA. The work that remains, the work AI is finally beginning to make manageable, is reading it correctly, in time, and for everyone.

References

[1] Duong D, Solomon BD. Artificial intelligence in clinical genetics. Eur J Hum Genet. 2025;33(3):281-288. doi:10.1038/s41431-024-01782-w

[2] Wick RR, Judd LM, Holt KE. Performance of neural network basecalling tools for Oxford Nanopore sequencing. Genome Biol. 2019;20(1):129. doi:10.1186/s13059-019-1727-y

[3] Poplin R, Chang PC, Alexander DH, et al. A universal SNP and small-indel variant caller using deep neural networks. Nat Biotechnol. 2018;36(10):983-987. doi:10.1038/nbt.4235

[4] Shafin K, Pesout T, Chang PC, et al. Haplotype-aware variant calling with PEPPER-Margin-DeepVariant enables high accuracy in nanopore long-reads. Nat Methods. 2021;18(11):1322-1332. doi:10.1038/s41592-021-01299-w

[5] Zheng Z, Li S, Su J, et al. Symphonizing pileup and full-alignment for deep learning-based long-read variant calling (Clair3). Nat Comput Sci. 2022;2(12):797-803. doi:10.1038/s43588-022-00387-x

[6] Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583-589. doi:10.1038/s41586-021-03819-2

[7] Cheng J, Novati G, Pan J, et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science. 2023;381(6664):eadg7492. doi:10.1126/science.adg7492

[8] Brandes N, Goldman G, Wang CH, Ye CJ, Ntranos V. Genome-wide prediction of disease variant effects with a deep protein language model (ESM1b). Nat Genet. 2023;55(9):1512-1522. doi:10.1038/s41588-023-01465-0

[9] Jaganathan K, Kyriazopoulou Panagiotopoulou S, McRae JF, et al. Predicting splicing from primary sequence with deep learning (SpliceAI). Cell. 2019;176(3):535-548.e24. doi:10.1016/j.cell.2018.12.015

[10] Rentzsch P, Witten D, Cooper GM, Shendure J, Kircher M. CADD: predicting the deleteriousness of variants throughout the human genome. Nucleic Acids Res. 2019;47(D1):D886-D894. doi:10.1093/nar/gky1016

[11] Gurovich Y, Hanani Y, Bar O, et al. Identifying facial phenotypes of genetic disorders using deep learning (DeepGestalt). Nat Med. 2019;25(1):60-64. doi:10.1038/s41591-018-0279-0

[12] Hsieh TC, Bar-Haim A, Moosa S, et al. GestaltMatcher facilitates rare disease matching using facial phenotype descriptors. Nat Genet. 2022;54(3):349-357. doi:10.1038/s41588-021-01010-x

[13] Köhler S, Gargano M, Matentzoglu N, et al. The Human Phenotype Ontology in 2021. Nucleic Acids Res. 2021;49(D1):D1207-D1217. doi:10.1093/nar/gkaa1043

[14] Smedley D, Jacobsen JOB, Jäger M, et al. Next-generation diagnostics and disease-gene discovery with the Exomiser. Nat Protoc. 2015;10(12):2004-2015. doi:10.1038/nprot.2015.124

[15] Richards S, Aziz N, Bale S, et al. Standards and guidelines for the interpretation of sequence variants: ACMG/AMP joint consensus recommendation. Genet Med. 2015;17(5):405-424. doi:10.1038/gim.2015.30

[16] Karczewski KJ, Francioli LC, Tiao G, et al. The mutational constraint spectrum quantified from variation in 141,456 humans (gnomAD). Nature. 2020;581(7809):434-443. doi:10.1038/s41586-020-2308-7

[17] Landrum MJ, Lee JM, Benson M, et al. ClinVar: improving access to variant interpretations and supporting evidence. Nucleic Acids Res. 2018;46(D1):D1062-D1067. doi:10.1093/nar/gkx1153

[18] Popejoy AB, Fullerton SM. Genomics is failing on diversity. Nature. 2016;538(7624):161-164. doi:10.1038/538161a

[19] Mahmud MI, Banerjee T, et al. Artificial intelligence in genomics: a comprehensive survey of methods, resources, challenges, and prospects. Brief Bioinform. 2026;27(3):bbag229. doi:10.1093/bib/bbag229

[20] Nguyen E, Poli M, Durrant MG, et al. Sequence modeling and design from molecular to genome scale with Evo. Science. 2024;386(6723):eado9336. doi:10.1126/science.ado9336