Request
Generate a canonical ATCG semantic trace, then explain what canonical means.
Maple Brain Lab · DNA Research
Two paths toward DNA-native intelligence
DOGMA and Hermon DNA are separate research programs. DOGMA asks whether regulated local computation can replace the transformer. Hermon DNA asks how far genome-native encoders can learn transferable biological sequence representations. Neither project claims AGI; both publish the evidence required to move toward it.
Start Here
From DNA to AGI, second-edition research preview
A 509-page undergraduate guide connects molecular biology, DNA computing, machine learning, the canonical non-transformer DOGMA architecture, recursive evaluation, laboratories, and open research questions.
509 pages19 chapters12-week courseEvidence-labeled claims
Measured DOGMA
Cycle 129DecisionRejected by evidence gate
Test perplexity1.1198
Generation probesNot passed
Inspect evidenceMeasured Hybrid Sample
Watch one request cross the genomic boundary
This is a replay of a held-out evaluation, not simulated typing. The bootstrap compiler emits a canonical 64-base trace; a compact realization adapter produces the answer.
ATCG trace
ATGATGCCCGTGCGCCCTAGCGACCTCACGCCAGAACGACAGAACGATCGACCGTGTAATAAAT
Transcript
Generate a canonical ATCG se
Natural language
Canonical refers to a standardized, consistent sequence of A/T/C/G in DNA/RNA strands, ensuring uniformity in molecular structure and function. It defines expected patterns in genetic sequences, enabling accurate comparison and analysis across different biological contexts.
Native DOGMACycle 129 · rejected29.8M parameters; malformed generation was not promoted.
Hybrid bridge4/4 held-out probes passedDeterministic ATCG compiler + 0.6B LoRA realization layer.
Scientific boundaryBridge validated, hypothesis openThe trace is not yet proven to improve language over a matched baseline.
Non-transformer genomic computation network
DOGMA
Can intelligence emerge from regulated local structure, persistent state, and expression instead of self-attention?
- Substrate
- Byte and nucleotide sequences
- Learning
- Autoregressive modeling, structural objectives, curriculum evolution, and gated architecture search
- Evidence state
- Latest 29.8M native candidate rejected; transformer realization bridge passes 4/4 narrow held-out probes
DNA encoder and transformer research line
Hermon DNA
How far can genome-native representations support reliable sequence understanding and scientific reasoning?
- Substrate
- Nucleotide sequences, annotations, tasks, and retrieved evidence
- Learning
- Masked sequence modeling, supervised task heads, contrastive objectives, and instruction adaptation
- Evidence state
- DNABERT-2 classifier path implemented; broader foundation evaluation in progress
The Boundary
DNA-inspired is not the same as a DNA language model
DOGMA borrows organization principles such as local structure, regulation, persistent marks, transcription, and expression to define a new computational architecture. Its primary benchmark substrate can be text, code, symbolic sequences, or genomic data.
Hermon DNA models actual nucleotide sequences and scientific tasks. Its architecture is allowed to use transformers, bidirectional encoders, state-space models, and retrieval. The biology is in the data and evaluation target, not in a metaphor for the network.
Research Ladder
DNA-base AGI is a sequence of falsifiable milestones
The phrase is a research direction, not a product claim. Progress means clearing increasingly difficult evidence gates while retaining prior capability, calibration, and diversity.
Representation
Learn compact sequence states that transfer across tasks and lengths.
Linear probes, retrieval, reverse-complement consistency, out-of-distribution transfer.Dynamics
Maintain and revise latent state under changing rules and delayed consequences.
Long-horizon recall, rule switches, counterfactual interventions, causal ablations.Composition
Build reusable programs from local operators, memory, and conditional expression.
Systematic generalization to unseen combinations and lengths.Recursive improvement
Turn diagnosed failures into better curricula or bounded architecture proposals.
Multi-cycle improvement against frozen suites without loss of diversity or prior skills.Open-ended agency
Form hypotheses, select experiments, and revise models under explicit resource and safety constraints.
Independent replication on new environments; no AGI claim before this evidence exists.Recursive Learning
Self-improvement must be earned by independent evidence
Every cycle starts from measured failures, trains an isolated candidate, and ends at a frozen evaluation gate. Real-data anchors remain in the corpus; synthetic data is provenance-labeled and capped.
Primary Literature
Research foundations and competing baselines
These papers are references and baselines, not evidence that MapleAI has reproduced their results.
DNABERT-2ICLR 2024
BPE tokenization, efficient multi-species pretraining, and the Genome Understanding Evaluation benchmark.
Nucleotide TransformerNature Methods 2025
A family of DNA foundation encoders evaluated across human and multi-species genomic tasks.
HyenaDNANeurIPS 2023
Single-nucleotide, long-context genomic modeling using implicit convolutions instead of dense attention.
CaduceusICML 2024
Bidirectional state-space modeling with reverse-complement equivariance for DNA.
MambaCOLM 2024
Selective state-space sequence modeling with linear scaling in sequence length.
EvoScience 2024
A long-context biological sequence model based on StripedHyena, spanning nucleotide to genome scales.