AUSTRALIS: Illuminating Antarctic Genomic Dark Matter with Environment-Aware AI
Interactive explorer of Southern Ocean genes, foundation-model representations, structural evidence, ecological alignment, and open-set detection in a leakage-controlled benchmark.
Overview — From the ACE catalogue to the AUSTRALIS pilot benchmark
From Southern Ocean metagenomes to a leakage-controlled, multimodal benchmark for studying the ecological organization of functionally unresolved genes.
Pilot benchmark construction
How 89.7 million unigenes become 92,272 benchmark genes
Retain genes linked to exactly one AGC; exclude absent or ambiguous multi-AGC assignments.
- Strict-known or strict-unknown functional status
- Valid, non-DISC AGNOSTOS category
- Non-missing environmental predictability score
- Defined HIGH or LOW ecological association
120,754 strict-known and 260,761 strict-unknown. Nucleotide length must be divisible by three, without internal stops, ambiguous codons, or translated X residues.
Proteins are grouped at ≥30% sequence identity and ≥80% alignment coverage. Each group remains entirely within train, validation, or test.
Balance length, GC content, protein length, AGC size, prevalence, abundance, and taxonomy.
11,534 genes per stratum × 8 strata
FL/ATT × strict-known/strict-unknown × HIGH/LOW.
Original ACE catalogue
Southern Ocean sampling and ecological context
- 34 stations sampled during the Antarctic Circumnavigation Expedition.
- 218 metagenomes spanning free-living and particle-attached fractions.
- 207 model-ready samples: 127 FL and 80 ATT.
- Physical, nutrient, depth, chlorophyll, and water-mass descriptors.
Unigene: non-redundant representative coding sequence.
AGC: gene cluster used for functional and ecological organization.
CD-HIT: method that clusters highly similar genes and selects representatives.
Functional resources: KEGG KO, SEED, eggNOG, CAZy, BiGG, and Pfam provide pathway, orthology, enzyme, reaction, and domain evidence. AGNOSTOS classifies clusters as K (known), GU (genomic unknown), or EU (environmental unknown).
AUSTRALIS benchmark
A balanced view of resolved and unresolved genes
Strict-known 45,628 · 49.8%
Strict-unknown 46,035 · 50.2%
Functional dark matter is not random sequence.
Strict-known requires explicit annotation evidence and AGNOSTOS category K.
Strict-unknown requires no explicit annotation and category GU or EU. It means unresolved by current evidence, not proven biochemical novelty.
Molecular foundation models
Complementary views of the same biological sequence
GenomeOcean
Genomic language model (500 million parameters) producing context-aware DNA embeddings.
ESM2
Protein language model (650 million parameters) representing amino-acid sequences.
ESM3
Multimodal protein language model (1,400 million parameters) used to compare sequence and predicted-structure representations.
Explore physical gradients, nutrients, geography, and latent environmental variables for normalized ACE samples.
Antarctica-centered projection of ACE sampling locations. Point color encodes the selected environmental variable.
The distribution and marginal box plots show range, concentration, and variation between size fractions.
Samples and environmental metadata
Controlled comparison of baselines, GenomeOcean, ESM2, and fused representations across six ecological-functional tasks.
Can molecular representations distinguish genes with HIGH versus LOW ecological association?How much do DNA and protein foundation models contribute relative to simple properties, compositional descriptors, sequence-similarity baselines, and DNA-protein fusion strategies?
Mean AUROC across repeated seeds for the six benchmark tasks. Higher values indicate better discrimination between genes with HIGH and LOW ecological association. Paired comparisons use the same samples, environmental coverage, and homology-aware partitions.
Per-seed representation benchmark across six ecological-functional tasks
Each row records one seeded evaluation under controlled splits.
Incremental value of ESM3 sequence, ESM3 structure, and bi- or trimodal fusion.
Does the ESM3 structural representation add predictive information beyond DNA and protein representations, either on its own or through multimodal fusion?
The heatmap reports held-out performance for classifying genes as having HIGH or LOW ecological association. Each cell summarizes test-set performance for an L2-regularized logistic-regression classifier using DNA, protein-sequence, predicted-structure, or multimodal embeddings. Standardization and PCA are fitted only on training data, and MMseqs2 homology groups are kept within a single partition to prevent leakage. Within each column, all representations use the same genes, labels, and splits, enabling controlled comparison of their predictive information.
Structural benchmark results
Structure does not universally outperform sequence. Neutral and negative results are retained because they inform multimodal model design.
Masked-known retrieval evaluates nearest neighbors against KEGG KO, Pfam, and EggNOG without exposing the target label.
Do learned molecular representation spaces place genes with shared functional annotations near one another when the query label is hidden?
Recovery at increasing k measures whether neighboring genes share a functional annotation hidden from the query. Differences among KEGG KO, Pfam, and EggNOG reflect their distinct granularity and coverage.
Functional-label recovery by k
Stratified robustness and generalization
Pfam-family holdout: 95 in-distribution families, 20 validation OOD families, and 21 test OOD families evaluated across five seeds.
Can the evaluated representations distinguish proteins from unseen Pfam families while preserving discrimination among families observed during training?
OOD AUROC measures separation of unseen from observed Pfam families; ID accuracy and macro-F1 show whether known families remain classifiable.
Preliminary open-set results
Search strict-unknown genes prioritized through representational consensus, ecological coherence, and robustness controls.
Which strict-unknown genes remain the strongest candidates after integrating representational consensus, ecological coherence, and robustness evidence?
Filters subset the frozen candidate catalogue without recalculating evidence scores. The resulting table reports genes satisfying the selected criteria and their supporting neighborhood evidence.
Filtered candidates
Candidate counts summarize how many strict-unknown genes remain at each evidence tier. Tiers rank evidence consistency, not confirmed biochemical novelty.
Loading the frozen reviewer snapshot…