Autonomous AI agents discover reverse transcriptases with tandem repeat arrays
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 autonomous AI agents discovered a new family of reverse transcriptases in bacterial viruses that carry tandem repeat arrays upstream, producing abundant RNAs during infection. All three tests fail.
Independently Verifiable
Fail
The computers returned strings of letters and the model told them those letters were DNA sequences encoding reverse transcriptases, but a stranger looking at the same screen sees letters, not enzymes.
Falsifiable
Fail
The search used profiles built from known reverse transcriptases, and sequences that did not match were sorted into categories called novel or unplaced rather than treated as evidence the classification system is wrong.
Non-Circular
Fail
The profile HMMs were built from sequences the model already calls reverse transcriptases, and the database was constructed by software that predicts genes from assemblies using the same model.
Why
They took 15.8 billion predicted protein sequences from public databases and clustered them into about 1.9 billion groups. Software had already predicted these proteins from metagenomic assemblies, which were built from sequencing machines that produced fluorescence signals interpreted as DNA bases. The AI agents searched these clusters using 52 profile HMMs built from known reverse transcriptase sequences. The search returned about 200,000 hits, which were sorted into nine groups. When a worker agent read a string of letters upstream of one hit, it recognized a repeating pattern and called it a tandem repeat array. They named the family ART. They then aligned RNA-seq reads to a reference genome and counted coverage over the array region, calling the counts expression. They predicted protein structures with AlphaFold2 and compared them to solved structures. At no point did anyone see a reverse transcriptase, a repeat array in a cell, or an RNA molecule. They saw letters on screens and coverage plots, and the model told them what those things meant.
2·Paradigm · Genomics / DNA Model
Does the framework this study assumes pass the three tests?
UNSCIENTIFIC
Independently Verifiable
Fail
Nobody has seen DNA, let alone watched it encode a protein; the sequencer produces light and the model says the light means DNA bases.
Falsifiable
Fail
When sequences do not match any known profile, the field calls them novel or unplaced, not evidence that the DNA model is wrong.
Non-Circular
Fail
The reference genomes, gene predictions, and profile HMMs were all built from the DNA model they claim to validate.
Why
The genomics framework assumes that DNA carries a code, that sequencing machines read that code, and that sequence similarity reveals evolutionary relationships and function. Nobody has seen DNA inside a living cell, let alone watched it encode a protein or copy RNA. The sequencer produces flashes of light. Software turns those flashes into letters. The model says the letters are DNA bases. Profile HMMs are built from sequences the model already classified as reverse transcriptases. Those profiles search databases of proteins that other software predicted from assemblies built by the same model. Every tool in the chain was constructed from the premise it claims to confirm. When sequences do not match, the field creates new categories called novel or unplaced rather than questioning whether sequence similarity means what the model says it means. The model treats a tagged signal as a discovery instead of asking whether the signal corresponds to anything a person could witness without first accepting the model.
From the study
Genes were predicted with prodigal-gv 2.10.0 under default parameters in unannotated assemblies of Logan, ENA, and NCBI, and the source gene calls were retained for annotated assemblies of NCBI and JGI.
52 profile hidden Markov models (HMMs) were assembled from six Pfam profiles (RVT_1, RVT_2, RVT_3, RVT_N, GIIM, and Intron_maturas2), the 45 class-specific profiles of myRT, and one profile that was built from an alignment from Toro.
The L0050 (228,907 bp logan contig) flank is spectacular: I can see by eye a tandem repeat array: "CATGTGTATCGCATGTT" / "CATGTGTTTCGCATGT" repeating many times with ~100-180 bp spacers — that's a CRISPR-like or msDNA-like repeat array?!