2026 · Nature
Genome modelling and design across all domains of life with Evo 2
Verdict
UNSCIENTIFIC
Neither the paradigm nor the study inside it passes. The verdict is unscientific.
1·Study methodology · inside the paradigm
UNSCIENTIFIC
Does not follow the scientific method within Genomics / DNA Model.
2·Paradigm · Genomics / DNA Model
UNSCIENTIFIC
Fails the tests. The worse layer decides.
How the verdict is decided
The verdict grades both layers: the paradigm this study assumes, and how the study was carried out inside it. Both must be scientific for the whole thing to be scientific. A clean method inside an unscientific paradigm is unscientific. An unscientific method inside a better paradigm is unscientific. A split on either layer is mixed. The worse layer decides.
1·Study methodology
Did this study test its claim with methods that are independent, falsifiable, and non-circular?
UNSCIENTIFIC
This study claims that an AI trained on 9 trillion letters of DNA sequence can predict how genetic changes affect living things and can design new genome-scale sequences that work. All three tests fail.
Independently Verifiable
Fail
The computer produced probability numbers, and the DNA model told them those numbers meant harmful mutations and working genes.
Falsifiable
Fail
If the scores had matched nothing, the field would blame model size or training data, never the claim that letter strings carry biological function.
Non-Circular
Fail
The model was trained on sequences assembled by the DNA model and then graded against lists of mutations that the same model had already labeled harmful.
Why
They compiled 9.3 trillion letters from machines and software that assume those letters are DNA, and trained a computer to predict the next letter in a string. Then they asked whether changes in the computer's probability scores matched lists of mutations that clinicians and labs had already labeled harmful using the same framework. They called the matches proof that the computer understands biology. The exon test was graded against genome annotations produced by the same framework. The fitness tests were graded against lab readouts whose meaning, which variant is fit and which is broken, is supplied by the same model. The one genuinely negative result, no correlation with human-infecting virus proteins, was reported as proof the safety exclusion worked, not as a failure. At no point did anyone observe a mutation causing anything. They observed numbers moving and the model told them what the movement meant.
2·Paradigm · Genomics / DNA Model
Does the framework this study assumes pass the three tests?
UNSCIENTIFIC
Independently Verifiable
Fail
A sequencer gives off flashes of light and the model says the flashes are DNA letters sitting inside living cells as a code; nobody has ever watched that.
Falsifiable
Fail
When predicted effects fail to appear, the field says missing heritability, non-coding regions, or imperfect annotation, so the gene model can never lose.
Non-Circular
Fail
The machines, the assembly software, and the reference sequences were all built from the DNA model they are now used to confirm.
Why
The DNA model claims a molecule sits inside every cell, carries a code, and that changes in the code cause disease. Every instrument in this paper presupposes that claim. The sequencer produces light; software trained on the model turns the light into letters. The assemblies, the annotations, and the lists of harmful variants were all produced by tools built from the model. Evo 2 is a computer trained on those letters, and its predictions are graded against labels the model produced earlier. Nobody in this chain saw a gene do anything. They saw machines produce signals and software name the signals. When the model's predictions fall short, the field blames scale, data, or annotation quality, never the claim that the letters are a code that runs life. The premise rode through untouched and came out looking like a discovery.