| name | proteomics-heuristics |
| version | 0.1.0 |
| description | Domain heuristics for proteomics differential expression analysis. Covers data-type priors (LFQ, TMT, DIA), setting-specific top methods, ensemble strategies, and conditional rules derived from a 34576-experiment benchmark. Used by proposer and reflector to narrow the search space.
|
Proteomics Heuristics
Domain knowledge distilled from large-scale benchmarking of proteomics
differential expression analysis workflows.
Data-Type Priors
LFQ
| Axis | Recommended |
|---|
| DE | deqms, limma, rots, proda |
| FDR | ihw, bh |
| Imputation | none, mindet, seqknn, missforest |
| Normalization | none, rlr, median |
DEqMS is best when peptide counts are available; limma is a stable baseline.
TMT
| Axis | Recommended |
|---|
| DE | limma, proda, deqms |
| FDR | bh |
| Imputation | none, seqknn |
| Normalization | median, none |
TMT has lower missingness; limma is a strong baseline for TMT.
DIA
| Axis | Recommended |
|---|
| DE | limma, rots, proda, deqms |
| FDR | ihw |
| Imputation | none, mindet, knn |
| Normalization | none, quantile, rlr |
DIA missing data is MAR not MNAR; avoid MinProb.
Setting-Specific Top Methods
FG_DDA (FragPipe DDA)
- Best DE: deqms, rots, limma
- Best normalization: none, median
- Best imputation: mindet, seqknn
- Avoid DE: none
- DEqMS dominates DDA when peptide counts are available.
MQ_DDA (MaxQuant DDA)
- Best DE: deqms, limma
- Best normalization: none, median
- Best imputation: mindet, seqknn
- Avoid DE: none
DIANN_DIA (DIA-NN DIA)
- Best DE: limma, rots
- Best normalization: none
- Best imputation: mindet
- Avoid DE: none
- DIA-NN directLFQ + no norm + MinDet + limma is the top workflow.
spt_DIA (Spectronaut DIA)
- Best DE: rots, limma
- Best normalization: none
- Best imputation: mindet, knn
- ROTS is dominant on Spectronaut DIA (203/315 top-5 workflows).
FG_TMT (FragPipe TMT)
- Best DE: limma, proda
- Best normalization: median, none
- Best imputation: none, seqknn
- limma is the top DEA for FragPipe TMT.
MQ_TMT (MaxQuant TMT)
- Best DE: proda, limma
- Best normalization: median, none
- Best imputation: none, seqknn
- proDA is dominant on MaxQuant TMT (best_rank 28.8).
Ensemble Strategies
top_k_2of3
- Methods: limma, deqms, proda
- min_k: 2
- Conservative: protein must be DE in >=2/3 methods.
top_k_3of4
- Methods: limma, rots, deqms, proda
- min_k: 3
- Strict: protein must be DE in >=3/4 methods.
Condition Rules
| Condition | Action | Confidence | Evidence |
|---|
| missing_rate < 0.15 | prefer imputation=none | high | No high-performance workflow uses imputation when missing < 15% |
| has_peptide_counts and n_proteins > 200 | prefer de=deqms | high | DEqMS top-1 for FG_DDA (92 top-5) and MQ_DDA (81 top-5) |
| n_per_group < 3 | prefer de=rots or de=limma | medium | ROTS bootstrap handles small samples; limma EB shrinkage is robust |
| n_proteins > 500 | prefer fdr=ihw | high | IHW gains power over BH when many hypotheses are available |
| data_type == DIA | avoid imputation=minprob | high | DIA missing is MAR; MNAR imputation (MinProb) hurts performance |
| data_type == TMT and missing_rate < 0.05 | prefer imputation=none and normalization=median | high | TMT has low missingness; median norm + no imputation is top |