In development · Early access opening soon
Decoding the grammar of the human genome.
Aibilge is an AI foundation model that learns the language of DNA. It reads regulatory sequence the way language models read text, to reveal how genes are switched on, off and tuned.
01 · The challenge
We can read every letter. We still can't read the sentences.
The human genome has been sequenced, but most of it is not genes. It is the regulatory code that decides when, where and how much each gene is used. Most disease-linked variants sit in this code, and its rules are still largely unknown.
base pairs in a single human genome
protein-coding genes, a small fraction of the sequence
of the genome does not code for protein, and much of its function is unresolved
02 · Genome grammar
DNA has a language. Aibilge learns its rules.
Short motifs act like words. Their order, spacing and combination act like syntax. Together they form instructions that cells read to control every gene.
- Vocabulary: binding motifs, promoters, splice signals
- Syntax: spacing, orientation and combinations of motifs
- Context: interactions across thousands of base pairs
GGCCAGTTCAGGCTGACCTGGACCTATAAAAGCTCAGGTCCCAGCTGAGCCACGTGAGGTTCGAGGACTTACCAGTCCGCGGCGCCGCGTTGACCAATGGCTCAGGAAGCCGTAAGTCCTGACTTCAGGCTTCCAGG
- TATA box · core promoter25–32
- E-box · TF binding motif51–56
- CpG island · methylation78–89
- 5′ splice donor site112–117
Illustrative example. Not model output.
03 · Approach
From sequence to function, in three steps.
- 01
Read
Raw DNA is tokenised into the units the genome actually uses, so the model sees sequence the way a linguist sees text.
- 02
Learn the grammar
Self-supervised training across genomes teaches the model regulatory vocabulary, syntax and long-range dependencies, without hand-written rules.
- 03
Predict
From sequence alone, the model estimates where genes are regulated, how strongly they are expressed, and what a change in a single letter could do.
04 · Applications
Built for the questions genomics can't yet answer.
Variant interpretation
Prioritise non-coding variants of uncertain significance by their predicted effect on gene regulation.
Regulatory discovery
Map enhancers, silencers and binding sites to see how cells switch genes on, off and in between.
Sequence design
Design regulatory DNA with intended activity for gene therapy and synthetic biology.
Target discovery
Connect disease-associated regions to the genes and mechanisms they control.
Early access
Help shape the first model of genome grammar.
We're opening access to a small group of researchers, clinicians and partners. Tell us who you are and we'll be in touch.