File size: 11,577 Bytes
238fe38
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
# Copyright 2026 InstaDeep Ltd and CentraleSupélec
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""High-level API: raw counts in, cell types / cell states out."""

from dataclasses import dataclass
from typing import Optional, Sequence, Union

import torch

from .postprocessing import extract_predicted_cell_type, parse_cell_state_prediction
from .preprocessing import GeneMapper, ScFoundationPreprocessor
from .prompts import (
    TASK_CELL_STATE,
    TASK_CELL_TYPING,
    build_in_context_prompt,
    build_prompt,
)

# Mixed precision of the predictor: None (fp32), "auto" or a torch dtype.
AutocastDtype = Union[str, torch.dtype, None]


@dataclass
class CellStatePrediction:
    programs: list[tuple[str, float]]  # (meta-program, score), strongest first
    hallmark: Optional[str]
    text: str


class CellScribePredictor:
    """
    >>> predictor = model.predictor(tokenizer)
    >>> predictor.predict_cell_type(counts, gene_names)
    """

    def __init__(
        self,
        model: torch.nn.Module,  # CellScribe (not imported: no import cycle)
        tokenizer,
        gene_list: Optional[Sequence[str]] = None,
        cell_types: Optional[Sequence[str]] = None,
    ) -> None:
        """``gene_list`` / ``cell_types``: default to the model's. ``cell_types``
        is the label vocabulary used to read cell-type answers (the training
        vocabulary, as in the paper's evaluation); without it, the free text
        after "The cell type is" is returned."""
        gene_list = gene_list or model.gene_list
        if gene_list is None:
            raise ValueError("no gene list: load the model with from_pretrained")
        cell_types = cell_types or model.cell_types
        self.model = model.eval()
        self.cell_types = list(cell_types) if cell_types else None
        self.tokenizer = tokenizer
        self.tokenizer.padding_side = "left"
        self.mapper = GeneMapper(gene_list)
        cfg = model.config.scfoundation
        self.preprocessor = ScFoundationPreprocessor(
            num_genes=len(gene_list) + 2,
            pad_token_id=cfg["pad_token_id"],
            target_res_str=cfg["tgthighres"],
        ).to(self.device)

    @property
    def device(self) -> torch.device:
        return next(self.model.parameters()).device

    def _autocast(self, autocast_dtype: AutocastDtype) -> torch.autocast:
        """Mixed-precision context. ``"auto"``: bfloat16 on GPU (the precision of
        the paper's evaluation, and half the encoder memory), fp32 on CPU."""
        if autocast_dtype == "auto":
            autocast_dtype = torch.bfloat16 if self.device.type == "cuda" else None
        if isinstance(autocast_dtype, str):
            raise ValueError(
                "autocast_dtype must be 'auto', None or a dtype, "
                f"got {autocast_dtype!r}"
            )
        if autocast_dtype is not None and self.device.type != "cuda":
            # On CPU, torch's TransformerEncoderLayer keeps its fused fast path
            # under autocast, which mixes reduced-precision inputs with fp32 weights.
            raise ValueError("mixed precision is only supported on GPU; use None")
        return torch.autocast(
            self.device.type,
            dtype=autocast_dtype or torch.float32,
            enabled=autocast_dtype is not None,
        )

    @torch.no_grad()
    def generate_text(
        self,
        counts,
        gene_names: Sequence[str],
        task: str = TASK_CELL_TYPING,
        batch_size: int = 32,
        autocast_dtype: AutocastDtype = None,
    ) -> list[str]:
        """Generated answer for each cell.

        Args:
            counts: ``(n_cells, n_genes)`` raw UMI counts (numpy or scipy sparse).
            gene_names: HGNC symbols of the columns.
            autocast_dtype: mixed precision. ``None`` (default): fp32;
                ``"auto"``: bfloat16 on GPU, as in the paper's evaluation, and fp32
                on CPU; or a dtype such as ``torch.bfloat16``.

        The scFoundation pooling runs over the padded batch, so a cell's embedding
        depends slightly on its batch-mates; results are reproducible for a
        fixed ``batch_size`` and cell order.
        """
        if len(gene_names) != counts.shape[1]:
            raise ValueError(f"{len(gene_names)} names for {counts.shape[1]} columns")
        # Genes are matched once; cells are densified one batch at a time, so a
        # large sparse or backed matrix never has to fit in memory as a whole.
        indices = self.mapper.column_indices(gene_names)
        prompt = self.tokenizer(
            [build_prompt(self.tokenizer, task)], return_tensors="pt"
        )
        texts: list[str] = []
        for start in range(0, counts.shape[0], batch_size):
            dense = self.mapper.map_columns(counts[start : start + batch_size], indices)
            cells = self.preprocessor(torch.from_numpy(dense).to(self.device))
            n = cells["x"].shape[0]
            with self._autocast(autocast_dtype):
                generated = self.model.generate(
                    input_ids=prompt.input_ids.expand(n, -1).to(self.device),
                    attention_mask=prompt.attention_mask.expand(n, -1).to(self.device),
                    scrnaseq_preprocessed_input=cells,
                )
            texts += self.tokenizer.batch_decode(generated, skip_special_tokens=True)
        return texts

    def predict_cell_type(
        self, counts, gene_names: Sequence[str], **kwargs
    ) -> list[str]:
        texts = self.generate_text(counts, gene_names, task=TASK_CELL_TYPING, **kwargs)
        return [extract_predicted_cell_type(t, self.cell_types) for t in texts]

    def predict_cell_state(
        self, counts, gene_names: Sequence[str], **kwargs
    ) -> list[CellStatePrediction]:
        texts = self.generate_text(counts, gene_names, task=TASK_CELL_STATE, **kwargs)
        predictions = []
        for text in texts:
            programs, hallmark = parse_cell_state_prediction(text)
            predictions.append(CellStatePrediction(programs, hallmark, text))
        return predictions

    @torch.no_grad()
    def embed_cells(
        self,
        counts,
        gene_names: Sequence[str],
        batch_size: int = 32,
        autocast_dtype: AutocastDtype = None,
    ) -> torch.Tensor:
        """Conditioning vector of each cell, ``[n_cells, hidden_size]``."""
        if len(gene_names) != counts.shape[1]:
            raise ValueError(f"{len(gene_names)} names for {counts.shape[1]} columns")
        indices = self.mapper.column_indices(gene_names)
        vectors = []
        for start in range(0, counts.shape[0], batch_size):
            dense = self.mapper.map_columns(counts[start : start + batch_size], indices)
            cells = self.preprocessor(torch.from_numpy(dense).to(self.device))
            with self._autocast(autocast_dtype):
                vectors.append(self.model.embed_cells(cells).float())
        return torch.cat(vectors)

    @torch.no_grad()
    def generate_text_in_context(
        self,
        counts,
        gene_names: Sequence[str],
        example_counts,
        example_gene_names: Sequence[str],
        example_labels: Sequence[str],
        batch_size: int = 32,
        autocast_dtype: AutocastDtype = None,
    ) -> list[str]:
        """Generated cell-typing answer for each cell, given labelled example cells.

        In-context learning: the prompt lists the example cells with their cell
        types ("Example cell k: The cell type is <label>."), each example line
        being conditioned by its own cell, then asks for the type of the query
        cell. Example labels may be cell types absent from the training data.

        Args:
            counts, gene_names: query cells (as in :meth:`generate_text`).
            example_counts, example_gene_names: example cells, raw counts.
            example_labels: cell type of each example cell, in order.
        """
        if len(example_labels) != example_counts.shape[0]:
            raise ValueError(
                f"{len(example_labels)} labels for {example_counts.shape[0]} examples"
            )
        examples = self.embed_cells(
            example_counts,
            example_gene_names,
            batch_size=batch_size,
            autocast_dtype=autocast_dtype,
        )
        prompt_ids, spans = build_in_context_prompt(self.tokenizer, example_labels)
        input_ids = torch.tensor([prompt_ids], device=self.device)
        texts: list[str] = []
        for start in range(0, counts.shape[0], batch_size):
            queries = self.embed_cells(
                counts[start : start + batch_size],
                gene_names,
                batch_size=batch_size,
                autocast_dtype=autocast_dtype,
            )
            n = queries.shape[0]
            # One conditioning vector per prompt token: the query cell everywhere,
            # except on each example line, conditioned by its example cell.
            plan = queries[:, None, :].repeat(1, len(prompt_ids), 1)
            for k, (span_start, span_end) in enumerate(spans):
                plan[:, span_start:span_end] = examples[k]
            with self._autocast(autocast_dtype):
                generated = self.model.generate(
                    input_ids=input_ids.expand(n, -1),
                    attention_mask=torch.ones_like(input_ids).expand(n, -1),
                    conditioning=plan,
                    conditioning_default=queries,
                )
            texts += self.tokenizer.batch_decode(generated, skip_special_tokens=True)
        return texts

    def predict_cell_type_in_context(
        self,
        counts,
        gene_names: Sequence[str],
        example_counts,
        example_gene_names: Sequence[str],
        example_labels: Sequence[str],
        **kwargs,
    ) -> list[str]:
        """Cell type of each query cell, given labelled example cells (see
        :meth:`generate_text_in_context`). The answer is read as free text (the
        text after "The cell type is"), so it can be a type outside
        ``cell_types``."""
        texts = self.generate_text_in_context(
            counts=counts,
            gene_names=gene_names,
            example_counts=example_counts,
            example_gene_names=example_gene_names,
            example_labels=example_labels,
            **kwargs,
        )
        return [extract_predicted_cell_type(t) for t in texts]

    def predict_anndata(
        self, adata, task: str = TASK_CELL_TYPING, layer: Optional[str] = None, **kwargs
    ):
        """Same, for an AnnData with raw counts in ``layer`` (default ``.X``) and
        HGNC symbols as ``var_names``. Cells are read and processed one batch at a
        time, so a large AnnData can be opened with ``backed="r"``."""
        counts = adata.layers[layer] if layer else adata.X
        if task == TASK_CELL_TYPING:
            return self.predict_cell_type(counts, list(adata.var_names), **kwargs)
        return self.predict_cell_state(counts, list(adata.var_names), **kwargs)