Scores variants with a masked language model and its tokenizer, the way ProteinGym scores ESM models. A higher score means the model finds the variant more plausible.
More...
#include <variant_scoring.hpp>
|
| | VariantScorer (EncoderLM &model, const TextTokenizer &tokenizer, DeviceBackend *backend, VariantScoringOptions options={}) |
| |
| ResidueLogProbs | masked_marginals (std::string_view sequence) |
| | Each residue's distribution with that residue masked.
|
| |
| ResidueLogProbs | wild_type_marginals (std::string_view sequence) |
| | Each residue's distribution from one pass over the unmasked sequence.
|
| |
| double | pseudo_log_likelihood (std::string_view sequence) |
| | The sum over residues of log p(residue) with it masked. Every residue is scored.
|
| |
| double | score (const ResidueLogProbs &marginals, std::string_view sequence, const std::vector< Mutation > &mutations, int64_t offset=1) const |
| | The summed log-ratio of mutations in marginals computed on sequence.
|
| |
| std::vector< float > | single_mutant_scan (const ResidueLogProbs &marginals, std::string_view sequence) const |
| | Every single substitution: row i, column j is log p(kAminoAcids[j]) - log p(wild
type) at residue i, row-major, length x 20. The wild type's own column is 0.
|
| |
| int64_t | token_of (char residue) const |
| |
Scores variants with a masked language model and its tokenizer, the way ProteinGym scores ESM models. A higher score means the model finds the variant more plausible.
- Masked marginals (ProteinGym's choice for ESM-2) mask each residue in turn. A substitution scores
log p(mutant) - log p(wild type) at its masked position, and a multiple mutant the sum of its substitutions. The cost is one forward pass per residue, whatever the number of variants.
- Wild-type marginals read the same log-ratio from one unmasked pass. They are faster but less accurate.
- Pseudo-log-likelihood sums
log p(residue) with each residue masked in turn: one pass per residue, for each sequence scored.
◆ VariantScorer()
- Parameters
-
| backend | Where the model runs. The model, tokenizer and backend must outlive this. |
◆ masked_marginals()
| ResidueLogProbs pulsatrix::VariantScorer::masked_marginals |
( |
std::string_view |
sequence | ) |
|
Each residue's distribution with that residue masked.
- Exceptions
-
| std::invalid_argument | for an empty sequence or a character the tokenizer doesn't know. |
◆ pseudo_log_likelihood()
| double pulsatrix::VariantScorer::pseudo_log_likelihood |
( |
std::string_view |
sequence | ) |
|
The sum over residues of log p(residue) with it masked. Every residue is scored.
◆ score()
| double pulsatrix::VariantScorer::score |
( |
const ResidueLogProbs & |
marginals, |
|
|
std::string_view |
sequence, |
|
|
const std::vector< Mutation > & |
mutations, |
|
|
int64_t |
offset = 1 |
|
) |
| const |
The summed log-ratio of mutations in marginals computed on sequence.
- Exceptions
-
| std::invalid_argument | if a mutation doesn't match the sequence, or names a residue that isn't a single vocabulary token. |
◆ single_mutant_scan()
| std::vector< float > pulsatrix::VariantScorer::single_mutant_scan |
( |
const ResidueLogProbs & |
marginals, |
|
|
std::string_view |
sequence |
|
) |
| const |
Every single substitution: row i, column j is log p(kAminoAcids[j]) - log p(wild
type) at residue i, row-major, length x 20. The wild type's own column is 0.
◆ token_of()
| int64_t pulsatrix::VariantScorer::token_of |
( |
char |
residue | ) |
const |
◆ wild_type_marginals()
| ResidueLogProbs pulsatrix::VariantScorer::wild_type_marginals |
( |
std::string_view |
sequence | ) |
|
Each residue's distribution from one pass over the unmasked sequence.
The documentation for this class was generated from the following file: