epibudget
A Python CLI for ranking protein variants that expose mutation interactions under a fixed experimental budget.
Pairwise epistasis-map recovery under fixed experimental budgets for TrpB and GB1
For TrpB, info meets the registered pairwise map-recovery rule relative to fitness and random; the corrective GB1 analysis is inconclusive. structural denotes loop-coverage allocation. Point estimates only; y-axis scales differ by row.
The scientific problem
A protein language model can rank thousands of variants before any wet-lab measurement. Yet the variants with the highest predicted fitness are not necessarily those that best reveal interactions among mutations. Under a budget of 48, 96, or 192 wells, that distinction determines whether an experiment concentrates on predicted winners or samples variants that constrain an epistasis map.
Method
epibudget selects measurements; it is neither a fitness optimizer nor an epistasis-inference package. Given a wild-type sequence, candidate positions, a budget B, and an ESM-2 checkpoint, it returns an ordered list of B single-, double-, and triple-mutant variants.
- Conjoint ESM-2 scoring applies every mutation in a variant before computing conditional scores, preserving context-dependent interaction signal.
- Factor-graph construction represents WT-referenced pairwise and third-order inclusion–exclusion loops.
- Label-blind allocation ranks candidates before measured fitness enters the pipeline:
structuraluses loop coverage alone, whereasinfoweights that coverage by ESM masking dispersion.
Loop coverage, masking dispersion, and predicted fitness remain separable signals, allowing their contributions to be evaluated independently.
Validation design
Selection remains label-blind: measured fitness is revealed only after the selected identities are fixed. The registered benchmark compares five methods—info, fitness, random, practice, and structural—at budgets 48, 96, and 192 on GB1 and TrpB.
Map recovery reports pairwise and third-order correlations separately. A separate downstream benchmark fits the same pairwise-ridge learner to each selected plate and evaluates held-out double and triple mutants.
The absence of intervals is a data-availability constraint: the public TrpB artifact does not include pointwise confidence intervals.
Current findings
On TrpB, info meets the registered pairwise map-recovery rule relative to fitness and random. The structural loop-count baseline yields higher Pearson and Spearman estimates at budgets 96 and 192, though not at 48; the registered results therefore do not support an incremental contribution from masking dispersion.
In the registered downstream benchmark, structural outperforms fitness in 20/20 partitions on both GB1 and TrpB. The info-versus-structural gate does not pass on either landscape.
Limitations and status
All comparative results remain provisional. GB1 map recovery remains inconclusive_zero_gpu with public_claim_eligible=false, and the TrpB source mirror contains 871 imputed fitness values that are not identified row by row. No general claim is made beyond these two landscapes and the fixed downstream learner.