pulsatrix 1.1.0
Loading...
Searching...
No Matches
variant_scoring.hpp
Go to the documentation of this file.
1
6#pragma once
7
8#include <cstdint>
9#include <string>
10#include <string_view>
11#include <vector>
12
15
16namespace pulsatrix {
17
19struct Mutation {
20 char wild_type = 0;
21 int64_t position = 0;
22 char mutant = 0;
23};
24
29[[nodiscard]] std::vector<Mutation> ParseMutations(std::string_view text);
30
36[[nodiscard]] std::string ApplyMutations(std::string_view sequence, const std::vector<Mutation>& mutations, int64_t offset = 1);
37
40 int64_t length = 0;
41 int64_t vocab = 0;
42 std::vector<float> values;
43 [[nodiscard]] float at(int64_t residue, int64_t token) const { return values[static_cast<size_t>(residue * vocab + token)]; }
44};
45
47inline constexpr const char* kAminoAcids = "ACDEFGHIKLMNPQRSTVWY";
48
54 int64_t window = 1024;
56 int64_t batch_size = 8;
62 int64_t max_pass_bytes = int64_t{4} << 30;
63};
64
78public:
80 VariantScorer(EncoderLM& model, const TextTokenizer& tokenizer, DeviceBackend* backend, VariantScoringOptions options = {});
81
84 [[nodiscard]] ResidueLogProbs masked_marginals(std::string_view sequence);
86 [[nodiscard]] ResidueLogProbs wild_type_marginals(std::string_view sequence);
88 [[nodiscard]] double pseudo_log_likelihood(std::string_view sequence);
89
95 [[nodiscard]] double score(const ResidueLogProbs& marginals, std::string_view sequence, const std::vector<Mutation>& mutations,
96 int64_t offset = 1) const;
97
102 [[nodiscard]] std::vector<float> single_mutant_scan(const ResidueLogProbs& marginals, std::string_view sequence) const;
103
104 [[nodiscard]] int64_t token_of(char residue) const;
105
106private:
108 [[nodiscard]] std::vector<int64_t> Tokens(std::string_view sequence) const;
111 [[nodiscard]] int64_t BatchSize(int64_t length) const;
113 [[nodiscard]] std::vector<float> LogSoftmax(const std::vector<std::vector<int64_t>>& batch);
114
115 EncoderLM& model_;
116 const TextTokenizer& tokenizer_;
117 DeviceBackend* backend_;
118 VariantScoringOptions options_;
119 int64_t mask_id_ = 0;
120};
121
128[[nodiscard]] int64_t ScoringPassBytes(const EncoderLMConfig& config, int64_t length);
129
132[[nodiscard]] std::pair<int64_t, int64_t> OptimalWindow(int64_t position, int64_t length, int64_t window);
133
134} // namespace pulsatrix
Vendor-agnostic compute/memory backend. CPUBackend, CUDABackend (Phase 1.5), and HIPBackend (Phase 1....
Definition device_backend.hpp:248
A masked language model: token embeddings (with ESM's token-dropout scaling, when the config has it),...
Definition encoder_lm.hpp:109
A tokenizer: added tokens are split out of the raw text first (longest match, earliest first); each r...
Definition text_tokenizer.hpp:166
Scores variants with a masked language model and its tokenizer, the way ProteinGym scores ESM models....
Definition variant_scoring.hpp:77
ResidueLogProbs masked_marginals(std::string_view sequence)
Each residue's distribution with that residue masked.
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 resi...
double pseudo_log_likelihood(std::string_view sequence)
The sum over residues of log p(residue) with it masked. Every residue is scored.
int64_t token_of(char residue) const
ResidueLogProbs wild_type_marginals(std::string_view sequence)
Each residue's distribution from one pass over the unmasked sequence.
VariantScorer(EncoderLM &model, const TextTokenizer &tokenizer, DeviceBackend *backend, VariantScoringOptions options={})
Masked-language-model encoders: ESM-2 protein language models (PLM-2), Hugging Face EsmForMaskedLM ch...
Definition acquisition_functions.hpp:16
std::vector< Mutation > ParseMutations(std::string_view text)
Parses a variant, A23G, or several joined by :, A23G:L45P (ProteinGym's format).
@ Tokens
Token ids into a language model.
std::string ApplyMutations(std::string_view sequence, const std::vector< Mutation > &mutations, int64_t offset=1)
sequence with mutations applied; position p is sequence[p - offset].
int64_t ScoringPassBytes(const EncoderLMConfig &config, int64_t length)
Roughly the most memory one sequence of length tokens takes in a forward pass of config's model that ...
std::pair< int64_t, int64_t > OptimalWindow(int64_t position, int64_t length, int64_t window)
ProteinGym's get_optimal_window: the token range [begin, end) around position in a sequence of length...
constexpr const char * kAminoAcids
The 20 standard amino acids, in the order SingleMutantScan's columns use.
Definition variant_scoring.hpp:47
The settings of a masked LM, from its Hugging Face config.json.
Definition encoder_lm.hpp:35
One substitution, numbered from 1 as variant tables write it: A23G.
Definition variant_scoring.hpp:19
int64_t position
Definition variant_scoring.hpp:21
char mutant
Definition variant_scoring.hpp:22
char wild_type
Definition variant_scoring.hpp:20
Log-probabilities of every token at each residue: length rows of vocab values.
Definition variant_scoring.hpp:39
int64_t vocab
Definition variant_scoring.hpp:41
int64_t length
Definition variant_scoring.hpp:40
float at(int64_t residue, int64_t token) const
Definition variant_scoring.hpp:43
std::vector< float > values
Definition variant_scoring.hpp:42
Definition variant_scoring.hpp:49
int64_t batch_size
Masked copies run together in one forward pass, at most. It changes speed, not results.
Definition variant_scoring.hpp:56
int64_t max_pass_bytes
The memory one forward pass may use, by ScoringPassBytes. A batch of long windows shrinks to fit; a s...
Definition variant_scoring.hpp:62
int64_t window
The longest token window the model sees. A longer sequence is cut, for each masked position,...
Definition variant_scoring.hpp:54
The tokenizer interface (TOK-1): a pipeline of normalizer, pre-tokenizer, model, post-processor and d...