AI Disease Gene Discovery: Using Evolution as the Clue

The most useful trick in AI disease gene discovery right now is not reading medical papers. It’s reading 1,000 other species. An algorithm called EvORanker, developed by Dr. Christina Canavati and Prof. Yuval Tabach at the Faculty of Medicine of the Hebrew University of Jerusalem, ranks candidate disease-causing genes by comparing genetic patterns across more than 1,000 species. In testing, it put the correct gene at the top of the list in nearly 70% of cases and within the top five 95% of the time, and it performed best in exactly the hardest situation: genes nobody has studied much. The work was published in Genetics in Medicine, with collaborators at Oslo University Hospital, Vall d’Hebron in Barcelona, and Shaare Zedek Medical Center.

Key Takeaways

• The core signal is co-evolution: genes that work together in a cell tend to be kept or lost together across species. If two genes vanish in the same lineages, they probably share a job.
• EvORanker uses that pattern to rank candidate genes from a patient’s genome, moving the likely culprit to the top of a list that can otherwise hold thousands of entries.
• Performance: correct gene ranked first in nearly 70% of cases, in the top five in 95%.
• Its advantage is largest for poorly characterized genes, where literature-based methods have almost nothing to go on.
• It is a ranking tool, not a diagnosis. A clinician still evaluates the top candidates against the patient.

The problem: a sequencing result is a list, not an answer

Sequence a child’s genome and you get variants. Thousands of them look unusual in some way. Most are harmless. Somewhere in that list, maybe, is the one that explains the child’s symptoms.

Existing tools rank those candidates mostly by reading what has been published: which genes are known to cause which diseases, which are expressed in which tissues, which appear in which databases. That works well for genes people have studied. It fails for the roughly one-third of human genes that have almost no published functional characterization at all, which is exactly where undiagnosed cases tend to hide. The NIH’s rare disease center counts more than 10,000 rare diseases, with fewer than 5% having an approved treatment; a large share of undiagnosed patients are stuck against poorly understood genes.

So the question becomes: if the literature is silent about a gene, where else could information about its function come from?

AI disease gene discovery mechanism: what evolution already computed

Here is the actual idea, and it’s elegant enough to teach at a dinner table.

Step 1: Build a presence/absence matrix. Take a human gene. Now check all 1,000-plus species with sequenced genomes: does each one have a recognizable version of this gene, and how similar is it? You end up with a long row of numbers, one per species. That row is called a phylogenetic profile.

Step 2: Compare profiles. Do this for every human gene and you have a giant matrix. Now look for genes whose rows match: present in the same species, absent in the same species, weakened in the same species.

Step 3: Infer shared function. This is the inference that makes the method work. If two genes were independently lost in the same set of lineages, the most likely explanation is that they belonged to the same biological system, and once that system was abandoned, both became unnecessary together. Classic example: species that lost the ability to build cilia lost the whole set of cilia-related genes as a block. Co-loss is evidence of co-function.

Step 4: Rank candidates using that signal. Given a patient’s candidate gene list plus their symptoms, EvORanker asks which candidates co-evolve with genes already linked to that symptom pattern. A gene that evolutionarily travels with known cilia genes is a strong suspect in a patient with a ciliopathy phenotype, even if nobody has ever published a functional study of it.

What the model computed, precisely. Not a prediction of protein structure. Not a language-model guess about text. A similarity score between rows of a presence/absence-and-conservation matrix, integrated with phenotype information, used to sort a list.

What it got right

Nearly 70% top-1 accuracy and 95% top-5 accuracy in the reported evaluation, outperforming existing tools, with the largest margins on poorly understood genes. That last part is the real contribution: it adds information where other methods have none.

What the limits are

The coverage published in March 2026 does not lay out failure modes in detail, which is itself worth noting. But the method’s constraints follow from its logic. Genes present in nearly every species have flat profiles and therefore carry little signal. Genes unique to a narrow lineage have almost no comparison set. Co-evolution suggests shared pathway membership, which is not the same as proving a specific variant causes a specific disease. And ranking accuracy measured on known disease genes, where the answer is already in the literature, is an optimistic estimate of performance on genuinely novel cases. The broader picture is sobering: an August 2026 scoping review in the Journal of Medical Internet Research found that of 81 AI studies in pediatric rare disease, only 25.9% included external validation and 14.8% prospective evaluation.

The analogy: the lunch-table method

Suppose you want to know which kids at a school belong to the same club, but the club list is missing.

You could ask around, which works for the popular clubs and fails for the obscure ones. Or you could check attendance records across 1,000 days: which kids are absent on the same days? If six kids are always missing together on Thursdays, they probably do something together on Thursdays. You’ve inferred group membership from a pattern of absences, without being told anything about the group.

That’s phylogenetic profiling. The 1,000 days are 1,000 species. Absence is gene loss. And the inference is: things that disappear together probably belonged together.

Cross-species evidence: what different organisms contribute

Organism group	Why it helps the comparison	What it reveals about human genes
Other mammals (mouse, dog)	Very similar gene sets	Fine-grained sequence conservation; poor at showing loss patterns
Fish and amphibians	Shared vertebrate systems, different environments	Genes tied to vision, hearing, and limb development
Insects and worms	Distant but well-studied genetics	Core cell machinery conserved for 600 million years
Fungi and plants	Lack whole animal systems	Sharp signal: which human genes are animal-specific
Single-celled eukaryotes with and without cilia	Some lineages lost cilia entirely	The textbook co-loss block, the method’s clearest example
Bacteria and archaea	Only the deepest-shared machinery	Which human genes trace to the origin of cells

The method’s power comes from the spread, not from any single species. A gene that is present in ciliated protists and absent in non-ciliated ones is telling you something no mouse study would.

How to Teach Your Kid About Finding Disease Genes With Evolution

Ages 5–8: Same and Different Animals

Get pictures of a human, a fish, a bird, a fly, and a tree. Ask which ones have eyes, which have bones, which have leaves. Make a grid with checkmarks. Then point at the pattern: things with eyes tend to also have the parts that make eyes work. Congratulations, they just built a presence/absence matrix without knowing the words.

Ages 9–12: The Absence Pattern Puzzle

Make a 10-by-10 grid: ten “genes” as rows, ten “species” as columns, filled with checks and blanks that you design so exactly three rows match each other perfectly. Ask your kid to find the matching rows. Then tell them those three genes probably do the same job, and ask why that would follow. This is the entire EvORanker idea, and a 10-year-old can execute it in fifteen minutes.

Ages 13+: Real Conservation Data

Take a gene your teen has heard of (say, the one for lactase, or a hemoglobin gene) and look it up in the NCBI Gene database, then click through to the ortholog list to see which species have it. Do the same for a second gene in the same pathway. Ask them to describe how similar the species lists are, and what that similarity would imply if the two genes’ functions were unknown. If they’re ambitious, Ensembl’s comparative genomics view shows gene trees across species directly.

The question to ask: “If two genes disappeared in the same twenty species, what’s the most boring explanation? And is boring more likely than exciting here?”

What to actually do at home

Use this to make evolution concrete instead of abstract

Most kids learn evolution as a story about the past. This is evolution as a working instrument: 1,000 genomes are a dataset, and comparing them locates disease genes in children today. That framing tends to stick better than finch beaks.

Name the inference out loud

“Things that vanish together probably worked together” is a pattern of reasoning, not a biology fact. It shows up in debugging code, in diagnosing a car, in figuring out which app is draining a phone battery. Kids who can name it start using it elsewhere.

Keep ranking and diagnosis separate

EvORanker sorts a list. That’s genuinely valuable and it isn’t a diagnosis. Our piece on the Boston Children’s rare disease cases shows what a full diagnostic workflow looks like, and human review sits at the end of both.

Tie it to the DNA basics your kid may already have

If your family has done the strawberry DNA extraction or talked about gene editing, this is the natural next layer: from “DNA exists” to “DNA is comparable across species, and the comparison is useful.” Our CRISPR and genetic engineering guide covers the editing side of the same territory.

What not to do

Don’t let “AI found a disease gene” imply the AI understood the disease. EvORanker computed similarity between rows of a matrix. The biology that makes that similarity meaningful was done by a billion years of evolution and by the scientists who chose to look there. That’s a better story anyway, and it’s true.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can state the core inference: genes lost in the same species probably share a job.
• Month 2 red flags: They treat “ranked first” as “proven.” Re-read the 70% figure with them and ask what happened in the other 30%.
• Month 3 self-check: Ask them to invent a case where the method would fail. A gene present in all species, or one unique to humans, are both correct answers. Being able to name a method’s blind spot is the marker of real understanding.

Frequently Asked Questions

What is phylogenetic profiling in simple terms?

It’s building a row of yes/no (and how-similar) entries for one gene across many species, then comparing rows between genes. Genes with matching rows, especially matching absences, tend to participate in the same biological process. EvORanker applies this across more than 1,000 species.

How accurate is EvORanker, really?

In the reported evaluation it ranked the correct disease-causing gene first in nearly 70% of cases and within the top five in 95%. Those numbers come from testing on cases where the answer was already known, which tends to be a favorable setting. Performance on genuinely unsolved cases is the open question, and prospective evaluation is rare across this whole field.

Why is it better for poorly understood genes?

Because competing methods lean on published literature, and for an unstudied gene there is no literature. Evolutionary conservation data exists for every gene with recognizable relatives, studied or not. The method draws on a source of information that doesn’t depend on anyone having written a paper.

Does this replace geneticists?

No. It reorders a candidate list so the promising entries are near the top. A clinical geneticist still has to evaluate whether a top-ranked variant actually explains a specific patient, check inheritance patterns, and often order functional tests. The tool shortens the search; the judgment stays human.

Can my kid explore real cross-species gene data?

Yes, and it’s free. The NCBI Gene database shows ortholog lists for most human genes, and Ensembl draws gene trees across species. For a teenager who already likes biology, comparing which species have a given gene is a genuinely engaging afternoon and requires no software installation.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Canavati, C., Tabach, Y., et al. (2026). “EvORanker: evolutionary-signal-based ranking of candidate disease genes.” Genetics in Medicine. Coverage: https://www.bfhu.org/2026/03/30/a-new-ai-breakthrough-could-change-the-long-search-for-rare-disease-diagnoses
2. Zhao, J., Luo, J., Li, Q., & Chen, Y. (2026, August 28). “Machine Learning, Large Language Models, and Multimodal AI for Diagnosing Pediatric Rare Diseases: Scoping Review.” Journal of Medical Internet Research. https://pmc.ncbi.nlm.nih.gov/articles/PMC13524366/
3. National Institutes of Health, Genetic and Rare Diseases Information Center. “About GARD.” https://rarediseases.info.nih.gov/about
4. National Center for Biotechnology Information. Gene database (ortholog data). https://www.ncbi.nlm.nih.gov/gene/
5. Ensembl. Comparative genomics and gene trees. https://www.ensembl.org/
6. NBC News. (2026, June 18). “AI helped Boston Children’s Hospital diagnose rare diseases in kids” (context on AI-assisted rare disease workflows). https://www.nbcnews.com/tech/innovation/ai-boston-childrens-hospital-diagnose-rare-diseases-kids-openai-rcna350387
7. DeepRare research team. (2026). Agentic diagnostic framework for rare disease. Nature. https://www.nature.com/articles/s41586-025-10097-9