RaCoon Web Server — Help Guide
Sequence
Example: MSLQMVTVSN…
Valid protein sequence (wild-type) used to compute the variant’s log-likelihood ratio (LLR) score with ESM1b. RaCoon uses this sequence to obtain the model’s raw prediction and identify residue-level attributes (disorder, PPI, etc.).
Variant
Example: M443T
The missense mutation you are evaluating.
Format: W123M, where:
- W = wild-type amino acid
- 123 = position in the protein (1-based)
- M = mutant amino acid
Raw LLR score
Example: -7.70
The uncalibrated log-likelihood ratio from ESM1b.
- Lower (more negative) values generally indicate a higher likelihood of pathogenicity.
- RaCoon does not interpret this score directly; it is mapped to a calibrated probability based on residue context.
Mutation residue class
Example: ('short', 'ordered', 'sulfur', 'non_ppi')
This shows the contextual attributes of the mutated residue used to determine which calibration node applies. In this example:
- short — mutation is in a protein <1022 aa (no sliding window)
- ordered — residue is predicted to be structured
- sulfur — cysteine or methionine residue
- non_ppi — residue is not predicted to be in a protein–protein interface
These attributes define the calibration subgroup whose feature distributions most closely match this variant.
Calibrated pathogenicity score
Example: 75.98 ± 3.42
RaCoon's final pathogenicity probability, expressed as a percent.
- 75.98% = estimated chance this variant is pathogenic.
- ±3.42% = uncertainty across 75 model replicates in the ensemble.
- This score is context-aware, meaning it depends not just on the LLR but also on the residue’s subgroup (disorder, interface, sulfur, etc.).
Node calibration scores
Example: ECE = 0.04, MCE = 0.04
These metrics describe the calibration quality of the node used for this variant.
- ECE (Expected Calibration Error): 0.04
- Measures overall deviation between predicted probabilities and observed frequencies.
Values <0.1 indicate excellent calibration. - Measures overall deviation between predicted probabilities and observed frequencies.
- MCE (Maximal Calibration Error): 0.04
- Largest per-bin difference between predicted and observed pathogenicity.
Lower values mean no individual score range is substantially miscalibrated. - Largest per-bin difference between predicted and observed pathogenicity.
Together, these scores reflect the reliability of the probability estimate for this specific residue subgroup.