Active learning chooses which new data to collect so that each labeled example improves a predictive model as much as possible. In the ENIGMA genotype-to-phenotype project it is used to choose the next round of growth screens at the oak-ridge-field-research-center. The output is a proposal, not a measured outcome. The method selected 50 conditions for the next Oak Ridge experiments. It prioritizes organic acids (fumaric, melibionic, itaconic), nitrate and low-pH-relevant substrates, where the model is least confident and field relevance is highest 1.
The ranking scores each of 343 testable conditions by a combined score: error rate × model uncertainty × field relevance weight 1. Field relevance doubles the weight of conditions that match Oak Ridge geochemistry. These include nitrogen sources such as nitrate, organic acids associated with necromass decomposition, low-pH-compatible substrates and aromatic compounds 1. The project’s active-learning notebook (NB10) produces the 50-condition proposal ranked by error × uncertainty × field relevance, as top candidates for the next ENIGMA experiments 1.
The report lists the top 10 recommended conditions: fumaric acid, melibionic acid, fumarate, itaconic acid, 2-hydroxypropanoic acid (lactic acid), hydroxy-glutaric acid γ-lactone, difumarate, L-glutamic acid, nitrate and pyruvic acid 1. AUC (area under the ROC curve, a threshold-free measure of classifier discrimination) is the report’s reason for ranking these conditions highly. The report says they fall in the class where the full-corpus model performs worst: AUC 0.654 for “other” carbon metabolism and 0.435 for nitrogen. It adds that these are the classes where Oak Ridge contamination chemistry matters most, through a nitric-acid-driven pH drop and organic acid accumulation in plume sediments 1. The “performs worst” description fits nitrogen (AUC 0.435, “worse than random”) but not the “other” class. The same report’s full-corpus table gives lower AUCs for antibiotics (0.619) and metals (0.605) than for “other” (0.654). It notes that metals are only trivially “predicted” because of a 98% growth rate 1. The selection therefore targets nitrogen, the weakest class, and the mid-ranked “other” class. It does not show that experiments in those classes will raise performance.
The proposal also chooses which organisms to test. prescottella (16% growth rate across tested conditions) and microbacterium (23%) are named as the most informative test subjects. Their predictions are among the least reliable in the current model. The report expects that adding data for these genera would disproportionately improve cross-genus generalization 1. Together with the condition ranking, these target genera define a concrete 50-experiment proposal for ENIGMA’s next growth screens. The proposal emphasizes fumaric, melibionic and itaconic acids, nitrate and low-pH-relevant substrates 1. The expected generalization gain from these genera is the report’s expectation, not a measured result.
The benefit of this selection is unvalidated, and any claim of benefit is a hypothesis. The report links the proposed 50 experiments to 7,844 prediction failures in the current corpus. It suggests that resolving even a fraction of them would move overall accuracy from 65.1% toward the per-class ceiling of ~78% seen on amino acids and nucleosides. This is a projection, not a measured gain 1. Formal retrospective subsampling against random selection had not been done. The report calls it the next validation step and describes the ranking framework itself as in place and actionable 1. The formal test of hypothesis H6 would measure accuracy improvement per labeled pair when experiments are added in active-learning order rather than random order, using held-out current data. That test remains to be done 1. See Genotype × Condition to Phenotype Prediction from ENIGMA Growth Curves for the full project context.
Tensions
The report gives two accounts of where the model is weakest. Its active-learning rationale calls the “other” class (AUC 0.654) and nitrogen (AUC 0.435) the classes where the full-corpus model performs worst. Its own full-corpus table ranks antibiotics (0.619) and metals (0.605) below “other”, and only nitrogen is marked “worse than random” 1. The two accounts have not been reconciled. As a result, it is unclear whether the field-relevance weight, and not error alone, explains why organic acids rank above antibiotic and metal conditions.
Open Directions
- H6 retrospective subsampling: using held-out current ENIGMA growth data, add labeled condition–strain pairs in active-learning-ranked order and in random order, and compare accuracy gained per labeled pair. This would show whether the error × uncertainty × field-relevance ranking beats random selection at all. The report names this as the next validation step 1.
- Isolating the field-relevance term: repeat the same subsampling with the field-relevance weight removed, so the score uses error and uncertainty only. This would show whether doubling the weight for Oak Ridge-relevant conditions explains why organic acids rank above lower-AUC antibiotic and metal conditions. The report defines the weighting but does not test its separate effect 1.
- Genus-level held-out test: hold out prescottella and microbacterium observations and test whether adding them improves cross-genus generalization more than adding other genera. This would test the report’s expectation that these genera are the most informative 1.