Documentation for the tasks

scprint.tasks.cell_emb

Embedder

Embedder a class to embed and annotate cells using a model

Parameters:
  • batch_size (int, default: 64 ) –

    The size of the batches to be used in the DataLoader. Defaults to 64.

  • num_workers (int, default: 8 ) –

    The number of worker processes to use for data loading. Defaults to 8.

  • how (str, default: 'random expr' ) –

    The method to be used for selecting valid genes. Defaults to "most expr".

  • max_len (int, default: 2000 ) –

    The maximum length of the gene sequence. Defaults to 1000.

  • add_zero_genes (int, default: 0 ) –

    The number of zero genes to add to the gene sequence. Defaults to 100.

  • precision (str, default: '16-mixed' ) –

    The precision to be used in the Trainer. Defaults to "16-mixed".

  • pred_embedding (List[str], default: ['cell_type_ontology_term_id', 'disease_ontology_term_id', 'self_reported_ethnicity_ontology_term_id', 'sex_ontology_term_id'] ) –

    The list of labels to be used for plotting embeddings. Defaults to [ "cell_type_ontology_term_id", "disease_ontology_term_id", "self_reported_ethnicity_ontology_term_id", "sex_ontology_term_id", ].

  • doclass (bool, default: True ) –

    Whether to perform classification. Defaults to True.

  • doplot (bool, default: True ) –

    Whether to generate plots. Defaults to True.

  • keep_all_cls_pred (bool, default: False ) –

    Whether to keep all class predictions. Defaults to False.

  • dtype (dtype, default: float16 ) –

    Data type for computations. Defaults to torch.float16.

  • output_expression (str, default: 'none' ) –

    The method to output expression data. Options are "none", "all", "sample". Defaults to "none".
Source code in scprint/tasks/cell_emb.py 
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
def __init__(
    self,
    batch_size: int = 64,
    num_workers: int = 8,
    how: str = "random expr",
    max_len: int = 2000,
    doclass: bool = True,
    add_zero_genes: int = 0,
    precision: str = "16-mixed",
    pred_embedding: List[str] = [
        "cell_type_ontology_term_id",
        "disease_ontology_term_id",
        "self_reported_ethnicity_ontology_term_id",
        "sex_ontology_term_id",
    ],
    plot_corr_size: int = 64,
    doplot: bool = True,
    keep_all_cls_pred: bool = False,
    dtype: torch.dtype = torch.float16,
    output_expression: str = "none",
):
    """
    Embedder a class to embed and annotate cells using a model

    Args:
        batch_size (int, optional): The size of the batches to be used in the DataLoader. Defaults to 64.
        num_workers (int, optional): The number of worker processes to use for data loading. Defaults to 8.
        how (str, optional): The method to be used for selecting valid genes. Defaults to "most expr".
        max_len (int, optional): The maximum length of the gene sequence. Defaults to 1000.
        add_zero_genes (int, optional): The number of zero genes to add to the gene sequence. Defaults to 100.
        precision (str, optional): The precision to be used in the Trainer. Defaults to "16-mixed".
        pred_embedding (List[str], optional): The list of labels to be used for plotting embeddings. Defaults to [ "cell_type_ontology_term_id", "disease_ontology_term_id", "self_reported_ethnicity_ontology_term_id", "sex_ontology_term_id", ].
        doclass (bool, optional): Whether to perform classification. Defaults to True.
        doplot (bool, optional): Whether to generate plots. Defaults to True.
        keep_all_cls_pred (bool, optional): Whether to keep all class predictions. Defaults to False.
        dtype (torch.dtype, optional): Data type for computations. Defaults to torch.float16.
        output_expression (str, optional): The method to output expression data. Options are "none", "all", "sample". Defaults to "none".
    """
    self.batch_size = batch_size
    self.num_workers = num_workers
    self.how = how
    self.max_len = max_len
    self.add_zero_genes = add_zero_genes
    self.pred_embedding = pred_embedding
    self.keep_all_cls_pred = keep_all_cls_pred
    self.plot_corr_size = plot_corr_size
    self.precision = precision
    self.doplot = doplot
    self.dtype = dtype
    self.doclass = doclass
    self.output_expression = output_expression

__call__

call function to call the embedding

Parameters:
  • model (Module) –

    The scPRINT model to be used for embedding and annotation.

  • adata (AnnData) –

    The annotated data matrix of shape n_obs x n_vars. Rows correspond to cells and columns to genes.

  • cache (bool, default: False ) –

    Whether to use cached results if available. Defaults to False.
Raises:
  • ValueError –

    If the model does not have a logger attribute.

  • ValueError –

    If the model does not have a global_step attribute.
Returns:
  • AnnData –

    The annotated data matrix with embedded cell representations.

  • –

    List[str]: List of gene names used in the embedding.

  • –

    np.ndarray: The predicted expression values if output_expression is not "none".

  • dict –

    Additional metrics and information from the embedding process.
Source code in scprint/tasks/cell_emb.py 
 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
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
def __call__(self, model: torch.nn.Module, adata: AnnData, cache=False):
    """
    __call__ function to call the embedding

    Args:
        model (torch.nn.Module): The scPRINT model to be used for embedding and annotation.
        adata (AnnData): The annotated data matrix of shape n_obs x n_vars. Rows correspond to cells and columns to genes.
        cache (bool, optional): Whether to use cached results if available. Defaults to False.

    Raises:
        ValueError: If the model does not have a logger attribute.
        ValueError: If the model does not have a global_step attribute.

    Returns:
        AnnData: The annotated data matrix with embedded cell representations.
        List[str]: List of gene names used in the embedding.
        np.ndarray: The predicted expression values if output_expression is not "none".
        dict: Additional metrics and information from the embedding process.
    """
    # one of "all" "sample" "none"
    try:
        mdir = (
            model.logger.save_dir if model.logger.save_dir is not None else "data"
        )
    except:
        mdir = "data"
    try:
        file = (
            mdir
            + "/step_"
            + str(model.global_step)
            + "_predict_part_"
            + str(model.counter)
            + "_"
            + str(model.global_rank)
            + ".h5ad"
        )
        hasfile = os.path.exists(file)
    except:
        hasfile = False

    if not cache or not hasfile:
        model.predict_mode = "none"
        model.keep_all_cls_pred = self.keep_all_cls_pred
        # Add at least the organism you are working with
        if self.how == "most var":
            sc.pp.highly_variable_genes(
                adata, flavor="seurat_v3", n_top_genes=self.max_len
            )
            curr_genes = adata.var.index[adata.var.highly_variable]
        adataset = SimpleAnnDataset(
            adata, obs_to_output=["organism_ontology_term_id"]
        )
        col = Collator(
            organisms=model.organisms,
            valid_genes=model.genes,
            how=self.how if self.how != "most var" else "some",
            max_len=self.max_len,
            add_zero_genes=self.add_zero_genes,
            genelist=[] if self.how != "most var" else curr_genes,
        )
        dataloader = DataLoader(
            adataset,
            collate_fn=col,
            batch_size=self.batch_size,
            num_workers=self.num_workers,
            shuffle=False,
        )
        model.eval()
        model.on_predict_epoch_start()
        device = model.device.type
        model.doplot = self.doplot
        with (
            torch.no_grad(),
            torch.autocast(device_type=device, dtype=self.dtype),
        ):
            for batch in tqdm(dataloader):
                gene_pos, expression, depth = (
                    batch["genes"].to(device),
                    batch["x"].to(device),
                    batch["depth"].to(device),
                )
                model._predict(
                    gene_pos,
                    expression,
                    depth,
                    predict_mode="none",
                    pred_embedding=self.pred_embedding,
                )
                torch.cuda.empty_cache()
        model.log_adata(name="predict_part_" + str(model.counter))
        try:
            mdir = (
                model.logger.save_dir
                if model.logger.save_dir is not None
                else "data"
            )
        except:
            mdir = "data"
        file = (
            mdir
            + "/step_"
            + str(model.global_step)
            + "_"
            + model.name
            + "_predict_part_"
            + str(model.counter)
            + "_"
            + str(model.global_rank)
            + ".h5ad"
        )

    pred_adata = sc.read_h5ad(file)
    if self.output_expression == "all":
        adata.obsm["scprint_mu"] = model.expr_pred[0]
        adata.obsm["scprint_theta"] = model.expr_pred[1]
        adata.obsm["scprint_pi"] = model.expr_pred[2]
        adata.obsm["scprint_pos"] = model.pos.cpu().numpy()
    elif self.output_expression == "sample":
        adata.obsm["scprint_expr"] = (
            utils.zinb_sample(
                model.expr_pred[0],
                model.expr_pred[1],
                model.expr_pred[2],
            )
            .cpu()
            .numpy()
        )
        adata.obsm["scprint_pos"] = model.pos.cpu().numpy()
    elif self.output_expression == "old":
        expr = np.array(model.expr_pred[0])
        expr[
            np.random.binomial(
                1,
                p=np.array(
                    torch.nn.functional.sigmoid(
                        model.expr_pred[2].to(torch.float32)
                    )
                ),
            ).astype(bool)
        ] = 0
        expr[expr <= 0.3] = 0
        expr[(expr >= 0.3) & (expr <= 1)] = 1
        adata.obsm["scprint_expr"] = expr.astype(int)
        adata.obsm["scprint_pos"] = model.pos.cpu().numpy()
    else:
        pass
    pred_adata.obs.index = adata.obs.index
    try:
        adata.obsm["scprint_umap"] = pred_adata.obsm["X_umap"]
    except:
        print("too few cells to embed into a umap")
    try:
        adata.obsm["scprint_leiden"] = pred_adata.obsm["leiden"]
    except:
        print("too few cells to compute a clustering")
    adata.obsm["scprint"] = pred_adata.X
    pred_adata.obs.index = adata.obs.index
    adata.obs = pd.concat([adata.obs, pred_adata.obs], axis=1)
    if self.keep_all_cls_pred:
        allclspred = model.pred
        columns = []
        for cl in model.classes:
            n = model.label_counts[cl]
            columns += [model.label_decoders[cl][i] for i in range(n)]
        allclspred = pd.DataFrame(
            allclspred, columns=columns, index=adata.obs.index
        )
        adata.obs = pd.concat(adata.obs, allclspred)

    metrics = {}
    if self.doclass and not self.keep_all_cls_pred:
        for cl in model.classes:
            res = []
            if cl not in adata.obs.columns:
                continue
            class_topred = model.label_decoders[cl].values()

            if cl in model.labels_hierarchy:
                # class_groupings = {
                #    k: [
                #        i.ontology_id
                #        for i in bt.CellType.filter(k).first().children.all()
                #    ]
                #    for k in set(adata.obs[cl].unique()) - set(class_topred)
                # }
                cur_labels_hierarchy = {
                    model.label_decoders[cl][k]: [
                        model.label_decoders[cl][i] for i in v
                    ]
                    for k, v in model.labels_hierarchy[cl].items()
                }
            else:
                cur_labels_hierarchy = {}

            for pred, true in adata.obs[["pred_" + cl, cl]].values:
                if pred == true:
                    res.append(True)
                    continue
                if len(cur_labels_hierarchy) > 0:
                    if true in cur_labels_hierarchy:
                        res.append(pred in cur_labels_hierarchy[true])
                        continue
                    elif true not in class_topred:
                        raise ValueError(
                            f"true label {true} not in available classes"
                        )
                    elif true != "unknown":
                        res.append(False)
                elif true not in class_topred:
                    raise ValueError(f"true label {true} not in available classes")
                elif true != "unknown":
                    res.append(False)
                # else true is unknown
                # else we pass
            if len(res) == 0:
                # true was always unknown
                res = [1]
            if self.doplot:
                print("    ", cl)
                print("     accuracy:", sum(res) / len(res))
                print(" ")
            metrics.update({cl + "_accuracy": sum(res) / len(res)})
    return adata, metrics

compute_classification

Compute classification metrics for the given annotated data.

Parameters:
  • adata (AnnData) –

    The annotated data matrix of shape n_obs x n_vars. Rows correspond to cells and columns to genes.

  • classes (List[str]) –

    List of class labels to be used for classification.

  • label_decoders (Dict[str, Any]) –

    Dictionary of label decoders for each class.

  • labels_hierarchy (Dict[str, Any]) –

    Dictionary representing the hierarchy of labels.

  • metric_type (List[str], default: ['macro', 'micro', 'weighted'] ) –

    List of metric types to compute. Defaults to ["macro", "micro", "weighted"].
Returns:
  • Dict[str, Dict[str, float]] –

    Dict[str, Dict[str, float]]: A dictionary containing classification metrics for each class.
Source code in scprint/tasks/cell_emb.py 
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
def compute_classification(
    adata: AnnData,
    classes: List[str],
    label_decoders: Dict[str, Any],
    labels_hierarchy: Dict[str, Any],
    metric_type: List[str] = ["macro", "micro", "weighted"],
) -> Dict[str, Dict[str, float]]:
    """
    Compute classification metrics for the given annotated data.

    Args:
        adata (AnnData): The annotated data matrix of shape n_obs x n_vars. Rows correspond to cells and columns to genes.
        classes (List[str]): List of class labels to be used for classification.
        label_decoders (Dict[str, Any]): Dictionary of label decoders for each class.
        labels_hierarchy (Dict[str, Any]): Dictionary representing the hierarchy of labels.
        metric_type (List[str], optional): List of metric types to compute. Defaults to ["macro", "micro", "weighted"].

    Returns:
        Dict[str, Dict[str, float]]: A dictionary containing classification metrics for each class.
    """
    metrics = {}
    for label in classes:
        res = []
        if label not in adata.obs.columns:
            continue
        labels_topred = label_decoders[label].values()
        if label in labels_hierarchy:
            parentdf = (
                bt.CellType.filter()
                .df(include=["parents__ontology_id"])
                .set_index("ontology_id")[["parents__ontology_id"]]
            )
            parentdf.parents__ontology_id = parentdf.parents__ontology_id.astype(str)
            class_groupings = {
                k: get_descendants(k, parentdf) for k in set(adata.obs[label].unique())
            }
        for pred, true in adata.obs[["pred_" + label, label]].values:
            if pred == true:
                res.append(true)
                continue
            if label in labels_hierarchy:
                if true in class_groupings:
                    res.append(true if pred in class_groupings[true] else "")
                    continue
                elif true not in labels_topred:
                    raise ValueError(f"true label {true} not in available classes")
            elif true not in labels_topred:
                raise ValueError(f"true label {true} not in available classes")
            res.append("")
        metrics[label] = {}
        metrics[label]["accuracy"] = np.mean(np.array(res) == adata.obs[label].values)
        for x in metric_type:
            metrics[label][x] = f1_score(
                np.array(res), adata.obs[label].values, average=x
            )
    return metrics

compute_corr

Compute the correlation between the output and target matrices.

Parameters:
  • out (ndarray) –

    The output matrix.

  • to (ndarray) –

    The target matrix.

  • doplot (bool, default: True ) –

    Whether to generate a plot of the correlation coefficients. Defaults to True.

  • compute_mean_regress (bool, default: False ) –

    Whether to compute mean regression. Defaults to False.

  • plot_corr_size (int, default: 64 ) –

    The size of the plot for correlation. Defaults to 64.
Returns:
  • dict( dict ) –

    A dictionary containing the computed metrics.
Source code in scprint/tasks/cell_emb.py 
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
def compute_corr(
    out: np.ndarray,
    to: np.ndarray,
    doplot: bool = True,
    compute_mean_regress: bool = False,
    plot_corr_size: int = 64,
) -> dict:
    """
    Compute the correlation between the output and target matrices.

    Args:
        out (np.ndarray): The output matrix.
        to (np.ndarray): The target matrix.
        doplot (bool, optional): Whether to generate a plot of the correlation coefficients. Defaults to True.
        compute_mean_regress (bool, optional): Whether to compute mean regression. Defaults to False.
        plot_corr_size (int, optional): The size of the plot for correlation. Defaults to 64.

    Returns:
        dict: A dictionary containing the computed metrics.
    """
    metrics = {}
    corr_coef, p_value = spearmanr(
        out,
        to.T,
    )
    corr_coef[p_value > 0.05] = 0
    # corr_coef[]
    # only on non zero values,
    # compare a1-b1 corr with a1-b(n) corr. should be higher

    # Plot correlation coefficient
    val = plot_corr_size + 2 if compute_mean_regress else plot_corr_size
    metrics.update(
        {"recons_corr": np.mean(corr_coef[val:, :plot_corr_size].diagonal())}
    )
    if compute_mean_regress:
        metrics.update(
            {
                "mean_regress": np.mean(
                    corr_coef[
                        plot_corr_size : plot_corr_size + 2,
                        :plot_corr_size,
                    ].flatten()
                )
            }
        )
    if doplot:
        plt.figure(figsize=(10, 5))
        plt.imshow(corr_coef, cmap="coolwarm", interpolation="none", vmin=-1, vmax=1)
        plt.colorbar()
        plt.title('Correlation Coefficient of expr and i["x"]')
        plt.show()
    return metrics

default_benchmark

Run the default benchmark for embedding and annotation using the scPRINT model.

Parameters:
  • model (Module) –

    The scPRINT model to be used for embedding and annotation.

  • default_dataset (str, default: 'pancreas' ) –

    The default dataset to use for benchmarking. Options are "pancreas", "lung", or a path to a dataset. Defaults to "pancreas".

  • do_class (bool, default: True ) –

    Whether to perform classification. Defaults to True.

  • coarse (bool, default: False ) –

    Whether to use coarse cell type annotations. Defaults to False.
Returns:
  • dict( dict ) –

    A dictionary containing the benchmark metrics.
Source code in scprint/tasks/cell_emb.py 
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
def default_benchmark(
    model: torch.nn.Module,
    default_dataset: str = "pancreas",
    do_class: bool = True,
    coarse: bool = False,
) -> dict:
    """
    Run the default benchmark for embedding and annotation using the scPRINT model.

    Args:
        model (torch.nn.Module): The scPRINT model to be used for embedding and annotation.
        default_dataset (str, optional): The default dataset to use for benchmarking. Options are "pancreas", "lung", or a path to a dataset. Defaults to "pancreas".
        do_class (bool, optional): Whether to perform classification. Defaults to True.
        coarse (bool, optional): Whether to use coarse cell type annotations. Defaults to False.

    Returns:
        dict: A dictionary containing the benchmark metrics.
    """
    if default_dataset == "pancreas":
        adata = sc.read(
            FILE_LOC + "/../../data/pancreas_atlas.h5ad",
            backup_url="https://figshare.com/ndownloader/files/24539828",
        )
        adata.obs["cell_type_ontology_term_id"] = adata.obs["celltype"].replace(
            COARSE if coarse else FINE
        )
        adata.obs["assay_ontology_term_id"] = adata.obs["tech"].replace(
            COARSE if coarse else FINE
        )
    elif default_dataset == "lung":
        adata = sc.read(
            FILE_LOC + "/../../data/lung_atlas.h5ad",
            backup_url="https://figshare.com/ndownloader/files/24539942",
        )
        adata.obs["cell_type_ontology_term_id"] = adata.obs["cell_type"].replace(
            COARSE if coarse else FINE
        )
    else:
        adata = sc.read_h5ad(default_dataset)
        adata.obs["batch"] = adata.obs["assay_ontology_term_id"]
        adata.obs["cell_type"] = adata.obs["cell_type_ontology_term_id"]
    preprocessor = Preprocessor(
        use_layer="counts",
        is_symbol=True,
        force_preprocess=True,
        skip_validate=True,
        do_postp=False,
    )
    adata.obs["organism_ontology_term_id"] = "NCBITaxon:9606"
    adata = preprocessor(adata.copy())
    embedder = Embedder(
        pred_embedding=["cell_type_ontology_term_id"],
        doclass=(default_dataset not in ["pancreas", "lung"]),
        devices=1,
    )
    embed_adata, metrics = embedder(model, adata.copy())

    bm = Benchmarker(
        embed_adata,
        batch_key="tech" if default_dataset == "pancreas" else "batch",
        label_key="celltype" if default_dataset == "pancreas" else "cell_type",
        embedding_obsm_keys=["scprint"],
        n_jobs=6,
    )
    bm.benchmark()
    metrics.update({"scib": bm.get_results(min_max_scale=False).T.to_dict()["scprint"]})
    metrics["classif"] = compute_classification(
        embed_adata, model.classes, model.label_decoders, model.labels_hierarchy
    )
    return metrics

scprint.tasks.grn

GNInfer

GNInfer a class to infer gene regulatory networks from a dataset using a scPRINT model.

Parameters:
  • layer (Optional[list[int]], default: None ) –

    List of layers to use for the inference. Defaults to None.

  • batch_size (int, default: 64 ) –

    Batch size for processing. Defaults to 64.

  • num_workers (int, default: 8 ) –

    Number of workers for data loading. Defaults to 8.

  • drop_unexpressed (bool, default: False ) –

    Whether to drop unexpressed genes. Defaults to False.

  • num_genes (int, default: 3000 ) –

    Number of genes to consider. Defaults to 3000.

  • precision (str, default: '16-mixed' ) –

    Precision type for computations. Defaults to "16-mixed".

  • cell_type_col (str, default: 'cell_type' ) –

    Column name for cell type information. Defaults to "cell_type".

  • how (str, default: 'random expr' ) –

    Method to select genes. Options are "random expr", "most var within", "most var across", "given". Defaults to "random expr".

  • preprocess (str, default: 'softmax' ) –

    Preprocessing method. Options are "softmax", "sinkhorn", "none". Defaults to "softmax".

  • head_agg (str, default: 'mean' ) –

    Aggregation method for heads. Options are "mean", "sum", "none". Defaults to "mean".

  • filtration (str, default: 'thresh' ) –

    Filtration method for the adjacency matrix. Options are "thresh", "top-k", "mst", "known", "none". Defaults to "thresh".

  • k (int, default: 10 ) –

    Number of top connections to keep if filtration is "top-k". Defaults to 10.

  • apc (bool, default: False ) –

    Whether to apply Average Product Correction. Defaults to False.

  • known_grn (optional, default: None ) –

    Known gene regulatory network to use as a reference. Defaults to None.

  • symmetrize (bool, default: False ) –

    Whether to symmetrize the adjacency matrix. Defaults to False.

  • doplot (bool, default: True ) –

    Whether to generate plots. Defaults to True.

  • max_cells (int, default: 0 ) –

    Maximum number of cells to consider. Defaults to 0.

  • forward_mode (str, default: 'none' ) –

    Mode for forward pass. Defaults to "none".

  • genes (list, default: [] ) –

    List of genes to consider. Defaults to an empty list.

  • loc (str, default: './' ) –

    Location to save results. Defaults to "./".

  • dtype (dtype, default: float16 ) –

    Data type for computations. Defaults to torch.float16.

  • locname (str, default: '' ) –

    Name for the location. Defaults to an empty string.

  • add_emb_in_model (int, default: 8 ) –

    Number of embeddings to add in the model. Defaults to 8.
Source code in scprint/tasks/grn.py 
 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
def __init__(
    self,
    layer: Optional[List[int]] = None,
    batch_size: int = 64,
    num_workers: int = 8,
    drop_unexpressed: bool = False,
    num_genes: int = 3000,
    precision: str = "16-mixed",
    cell_type_col: str = "cell_type",
    how: str = "random expr",  # random expr, most var within, most var across, given
    max_len: int = 3000,
    preprocess: str = "softmax",  # sinkhorn, softmax, none
    head_agg: str = "mean",  # mean, sum, none
    filtration: str = "thresh",  # thresh, top-k, mst, known, none
    k: int = 10,
    apc: bool = False,
    known_grn: Optional[any] = None,
    symmetrize: bool = False,
    doplot: bool = True,
    shared_qk: bool = True,
    max_cells: int = 0,
    forward_mode: str = "none",
    genes: List[str] = [],
    loc: str = "./",
    dtype: torch.dtype = torch.float16,
    locname: str = "",
    add_emb_in_model: int = 8,
):
    """
    GNInfer a class to infer gene regulatory networks from a dataset using a scPRINT model.

    Args:
        layer (Optional[list[int]], optional): List of layers to use for the inference. Defaults to None.
        batch_size (int, optional): Batch size for processing. Defaults to 64.
        num_workers (int, optional): Number of workers for data loading. Defaults to 8.
        drop_unexpressed (bool, optional): Whether to drop unexpressed genes. Defaults to False.
        num_genes (int, optional): Number of genes to consider. Defaults to 3000.
        precision (str, optional): Precision type for computations. Defaults to "16-mixed".
        cell_type_col (str, optional): Column name for cell type information. Defaults to "cell_type".
        how (str, optional): Method to select genes. Options are "random expr", "most var within", "most var across", "given". Defaults to "random expr".
        preprocess (str, optional): Preprocessing method. Options are "softmax", "sinkhorn", "none". Defaults to "softmax".
        head_agg (str, optional): Aggregation method for heads. Options are "mean", "sum", "none". Defaults to "mean".
        filtration (str, optional): Filtration method for the adjacency matrix. Options are "thresh", "top-k", "mst", "known", "none". Defaults to "thresh".
        k (int, optional): Number of top connections to keep if filtration is "top-k". Defaults to 10.
        apc (bool, optional): Whether to apply Average Product Correction. Defaults to False.
        known_grn (optional): Known gene regulatory network to use as a reference. Defaults to None.
        symmetrize (bool, optional): Whether to symmetrize the adjacency matrix. Defaults to False.
        doplot (bool, optional): Whether to generate plots. Defaults to True.
        max_cells (int, optional): Maximum number of cells to consider. Defaults to 0.
        forward_mode (str, optional): Mode for forward pass. Defaults to "none".
        genes (list, optional): List of genes to consider. Defaults to an empty list.
        loc (str, optional): Location to save results. Defaults to "./".
        dtype (torch.dtype, optional): Data type for computations. Defaults to torch.float16.
        locname (str, optional): Name for the location. Defaults to an empty string.
        add_emb_in_model (int, optional): Number of embeddings to add in the model. Defaults to 8.

    """
    self.batch_size = batch_size
    self.num_workers = num_workers
    self.layer = layer
    self.locname = locname
    self.how = how
    assert (
        self.how
        in [
            "most var within",
            "most var across",
            "random expr",
            "given",
            "most expr",
        ]
    ), "how must be one of 'most var within', 'most var across', 'random expr', 'given', 'most expr'"
    self.num_genes = num_genes
    self.preprocess = preprocess
    self.cell_type_col = cell_type_col
    self.filtration = filtration
    self.doplot = doplot
    self.genes = genes
    self.apc = apc
    self.dtype = dtype
    self.forward_mode = forward_mode
    self.k = k
    self.max_len = max_len
    self.symmetrize = symmetrize
    self.known_grn = known_grn
    self.head_agg = head_agg
    self.max_cells = max_cells
    self.curr_genes = None
    self.drop_unexpressed = drop_unexpressed
    self.precision = precision
    self.shared_qk = shared_qk
    self.add_emb_in_model = add_emb_in_model
    if self.filtration != "none" and self.head_agg == "none":
        raise ValueError("filtration must be 'none' when head_agg is 'none'")

__call__

call runs the method

Parameters:
  • model (Module) –

    The model to be used for generating the network

  • adata (AnnData) –

    Annotated data matrix of shape n_obs × n_vars. n_obs is the number of cells and n_vars is the number of genes.

  • cell_type (str, default: None ) –

    Specific cell type to filter the data. Defaults to None.
Returns:
  • AnnData –

    Annotated data matrix with predictions and annotations.

  • –

    np.ndarray: Filtered adjacency matrix.
Source code in scprint/tasks/grn.py 
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
def __call__(self, model: torch.nn.Module, adata: AnnData, cell_type=None):
    """
    __call__ runs the method

    Args:
        model (torch.nn.Module): The model to be used for generating the network
        adata (AnnData): Annotated data matrix of shape `n_obs` × `n_vars`. `n_obs` is the number of cells and `n_vars` is the number of genes.
        cell_type (str, optional): Specific cell type to filter the data. Defaults to None.

    Returns:
        AnnData: Annotated data matrix with predictions and annotations.
        np.ndarray: Filtered adjacency matrix.
    """
    # Add at least the organism you are working with
    if self.layer is None:
        self.layer = list(range(model.nlayers))
    subadata = self.predict(model, adata, self.layer, cell_type)
    adjacencies = self.aggregate(model.attn.get(), model.genes)
    if self.head_agg == "none":
        return self.save(
            adjacencies[self.add_emb_in_model :, self.add_emb_in_model :, :],
            subadata,
        )
    else:
        return self.save(
            self.filter(adjacencies)[
                self.add_emb_in_model :, self.add_emb_in_model :
            ],
            subadata,
        )

default_benchmark

default_benchmark function to run the default scPRINT GRN benchmark

Parameters:
  • model (Any) –

    The scPRINT model to be used for the benchmark.

  • default_dataset (str, default: 'sroy' ) –

    The default dataset to use for benchmarking. Defaults to "sroy".

  • cell_types (List[str], default: ['kidney distal convoluted tubule epithelial cell', 'kidney loop of Henle thick ascending limb epithelial cell', 'kidney collecting duct principal cell', 'mesangial cell', 'blood vessel smooth muscle cell', 'podocyte', 'macrophage', 'leukocyte', 'kidney interstitial fibroblast', 'endothelial cell'] ) –

    List of cell types to include in the benchmark. Defaults to [ "kidney distal convoluted tubule epithelial cell", "kidney loop of Henle thick ascending limb epithelial cell", "kidney collecting duct principal cell", "mesangial cell", "blood vessel smooth muscle cell", "podocyte", "macrophage", "leukocyte", "kidney interstitial fibroblast", "endothelial cell",

  • maxlayers (int, default: 16 ) –

    Maximum number of layers to use from the model. Defaults to 16.

  • maxgenes (int, default: 5000 ) –

    Maximum number of genes to consider. Defaults to 5000.

  • batch_size (int, default: 32 ) –

    Batch size for processing. Defaults to 32.

  • maxcells (int, default: 1024 ) –

    Maximum number of cells to consider. Defaults to 1024.
Returns:
  • dict –

    A dictionary containing the benchmark metrics.
Source code in scprint/tasks/grn.py 
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
def default_benchmark(
    model: Any,
    default_dataset: str = "sroy",
    cell_types: List[str] = [
        "kidney distal convoluted tubule epithelial cell",
        "kidney loop of Henle thick ascending limb epithelial cell",
        "kidney collecting duct principal cell",
        "mesangial cell",
        "blood vessel smooth muscle cell",
        "podocyte",
        "macrophage",
        "leukocyte",
        "kidney interstitial fibroblast",
        "endothelial cell",
    ],
    maxlayers: int = 16,
    maxgenes: int = 5000,
    batch_size: int = 32,
    maxcells: int = 1024,
):
    """
    default_benchmark function to run the default scPRINT GRN benchmark

    Args:
        model (Any): The scPRINT model to be used for the benchmark.
        default_dataset (str, optional): The default dataset to use for benchmarking. Defaults to "sroy".
        cell_types (List[str], optional): List of cell types to include in the benchmark. Defaults to [
            "kidney distal convoluted tubule epithelial cell",
            "kidney loop of Henle thick ascending limb epithelial cell",
            "kidney collecting duct principal cell",
            "mesangial cell",
            "blood vessel smooth muscle cell",
            "podocyte",
            "macrophage",
            "leukocyte",
            "kidney interstitial fibroblast",
            "endothelial cell",
        ].
        maxlayers (int, optional): Maximum number of layers to use from the model. Defaults to 16.
        maxgenes (int, optional): Maximum number of genes to consider. Defaults to 5000.
        batch_size (int, optional): Batch size for processing. Defaults to 32.
        maxcells (int, optional): Maximum number of cells to consider. Defaults to 1024.

    Returns:
        dict: A dictionary containing the benchmark metrics.
    """
    metrics = {}
    layers = list(range(model.nlayers))[max(0, model.nlayers - maxlayers) :]
    clf_omni = None
    if default_dataset == "sroy":
        preprocessor = Preprocessor(
            is_symbol=True,
            force_preprocess=True,
            skip_validate=True,
            do_postp=False,
            min_valid_genes_id=5000,
            min_dataset_size=64,
        )
        clf_self = None
        todo = [
            ("han", "human", "full"),
            ("mine", "human", "full"),
            ("han", "human", "chip"),
            ("han", "human", "ko"),
            ("tran", "mouse", "full"),
            ("zhao", "mouse", "full"),
            ("tran", "mouse", "chip"),
            ("tran", "mouse", "ko"),
        ]
        for da, spe, gt in todo:
            if gt != "full":
                continue
            print(da + "_" + gt)
            preadata = get_sroy_gt(get=da, species=spe, gt=gt)
            adata = preprocessor(preadata.copy())
            grn_inferer = GNInfer(
                layer=layers,
                how="most var within",
                preprocess="softmax",
                head_agg="none",
                filtration="none",
                forward_mode="none",
                num_genes=maxgenes,
                num_workers=8,
                max_cells=maxcells,
                doplot=False,
                batch_size=batch_size,
                devices=1,
            )
            grn = grn_inferer(model, adata)
            grn.varp["all"] = grn.varp["GRN"]
            grn.var["ensembl_id"] = grn.var.index
            grn.var["symbol"] = make_index_unique(grn.var["symbol"].astype(str))
            grn.var.index = grn.var["symbol"]
            grn.varp["GRN"] = grn.varp["all"].mean(-1).T
            metrics["mean_" + da + "_" + gt] = BenGRN(
                grn, do_auc=True, doplot=False
            ).compare_to(other=preadata)
            grn.varp["GRN"] = grn.varp["GRN"].T
            if spe == "human":
                metrics["mean_" + da + "_" + gt + "_base"] = BenGRN(
                    grn, do_auc=True, doplot=False
                ).scprint_benchmark()

            ## OMNI
            if clf_omni is None:
                grn.varp["GRN"] = grn.varp["all"]
                _, m, clf_omni = train_classifier(
                    grn,
                    C=1,
                    train_size=0.9,
                    class_weight={1: 800, 0: 1},
                    shuffle=True,
                    return_full=False,
                )
                joblib.dump(clf_omni, "clf_omni.pkl")
                metrics["omni_classifier"] = m
            grn.varp["GRN"] = grn.varp["all"][:, :, clf_omni.coef_[0] > 0].mean(-1)
            if spe == "human":
                metrics["omni_" + da + "_" + gt + "_base"] = BenGRN(
                    grn, do_auc=True, doplot=True
                ).scprint_benchmark()
            grn.varp["GRN"] = grn.varp["GRN"].T
            metrics["omni_" + da + "_" + gt] = BenGRN(
                grn, do_auc=True, doplot=False
            ).compare_to(other=preadata)

            ## SELF
            if clf_self is None:
                grn.varp["GRN"] = np.transpose(grn.varp["all"], (1, 0, 2))
                _, m, clf_self = train_classifier(
                    grn,
                    other=preadata,
                    C=1,
                    train_size=0.5,
                    class_weight={1: 40, 0: 1},
                    shuffle=False,
                    return_full=False,
                )
                metrics["self_classifier"] = m
            grn.varp["GRN"] = grn.varp["all"][:, :, clf_self.coef_[0] > 0].mean(-1).T
            metrics["self_" + da + "_" + gt] = BenGRN(
                grn, do_auc=True, doplot=False
            ).compare_to(other=preadata)
            if spe == "human":
                grn.varp["GRN"] = grn.varp["GRN"].T
                metrics["self_" + da + "_" + gt + "_base"] = BenGRN(
                    grn, do_auc=True, doplot=True
                ).scprint_benchmark()

            ## chip / ko
            if (da, spe, "chip") in todo:
                preadata = get_sroy_gt(get=da, species=spe, gt="chip")
                grn.varp["GRN"] = grn.varp["all"].mean(-1).T
                metrics["mean_" + da + "_" + "chip"] = BenGRN(
                    grn, do_auc=True, doplot=False
                ).compare_to(other=preadata)
                grn.varp["GRN"] = (
                    grn.varp["all"][:, :, clf_omni.coef_[0] > 0].mean(-1).T
                )
                metrics["omni_" + da + "_" + "chip"] = BenGRN(
                    grn, do_auc=True, doplot=False
                ).compare_to(other=preadata)
                grn.varp["GRN"] = (
                    grn.varp["all"][:, :, clf_self.coef_[0] > 0].mean(-1).T
                )
                metrics["self_" + da + "_" + "chip"] = BenGRN(
                    grn, do_auc=True, doplot=False
                ).compare_to(other=preadata)
            if (da, spe, "ko") in todo:
                preadata = get_sroy_gt(get=da, species=spe, gt="ko")
                grn.varp["GRN"] = grn.varp["all"].mean(-1).T
                metrics["mean_" + da + "_" + "ko"] = BenGRN(
                    grn, do_auc=True, doplot=False
                ).compare_to(other=preadata)
                grn.varp["GRN"] = (
                    grn.varp["all"][:, :, clf_omni.coef_[0] > 0].mean(-1).T
                )
                metrics["omni_" + da + "_" + "ko"] = BenGRN(
                    grn, do_auc=True, doplot=False
                ).compare_to(other=preadata)
                grn.varp["GRN"] = (
                    grn.varp["all"][:, :, clf_self.coef_[0] > 0].mean(-1).T
                )
                metrics["self_" + da + "_" + "ko"] = BenGRN(
                    grn, do_auc=True, doplot=False
                ).compare_to(other=preadata)
            del grn
    elif default_dataset == "gwps":
        if not os.path.exists(FILEDIR + "/../../data/perturb_gt.h5ad"):
            adata = get_perturb_gt()
            adata.write_h5ad(FILEDIR + "/../../data/perturb_gt.h5ad")
        else:
            adata = read_h5ad(FILEDIR + "/../../data/perturb_gt.h5ad")
        preprocessor = Preprocessor(
            force_preprocess=True,
            skip_validate=True,
            do_postp=False,
            min_valid_genes_id=maxgenes,
            min_dataset_size=64,
        )
        nadata = preprocessor(adata.copy())
        adata.var["isTF"] = False
        adata.var.loc[adata.var.gene_name.isin(grnutils.TF), "isTF"] = True
        adata.var["isTF"].sum()
        grn_inferer = GNInfer(
            layer=layers,
            how="most var within",
            preprocess="softmax",
            head_agg="none",
            filtration="none",
            forward_mode="none",
            num_genes=maxgenes,
            max_cells=maxcells,
            doplot=False,
            num_workers=8,
            batch_size=batch_size,
            devices=1,
        )
        grn = grn_inferer(model, nadata)
        grn.varp["all"] = grn.varp["GRN"]

        grn.varp["GRN"] = grn.varp["all"].mean(-1).T
        metrics["mean"] = BenGRN(grn, do_auc=True, doplot=False).compare_to(other=adata)
        grn.var["ensembl_id"] = grn.var.index
        grn.var.index = grn.var["symbol"]
        grn.varp["GRN"] = grn.varp["all"].mean(-1)
        metrics["mean_base"] = BenGRN(
            grn, do_auc=True, doplot=False
        ).scprint_benchmark()

        grn.varp["GRN"] = grn.varp["all"]
        grn.var.index = grn.var["ensembl_id"]
        _, m, clf_omni = train_classifier(
            grn,
            C=1,
            train_size=0.9,
            class_weight={1: 800, 0: 1},
            shuffle=True,
            doplot=False,
            return_full=False,
            use_col="gene_name",
        )
        grn.varp["GRN"] = grn.varp["all"][:, :, clf_omni.coef_[0] > 0].mean(-1).T
        metrics["omni"] = BenGRN(grn, do_auc=True, doplot=False).compare_to(other=adata)
        metrics["omni_classifier"] = m
        grn.var.index = grn.var["symbol"]
        grn.varp["GRN"] = grn.varp["GRN"].T
        metrics["omni_base"] = BenGRN(
            grn, do_auc=True, doplot=False
        ).scprint_benchmark()
        grn.varp["GRN"] = np.transpose(grn.varp["all"], (1, 0, 2))
        grn.var.index = grn.var["ensembl_id"]
        _, m, clf_self = train_classifier(
            grn,
            other=adata,
            C=1,
            train_size=0.5,
            class_weight={1: 40, 0: 1},
            doplot=False,
            shuffle=False,
            return_full=False,
            use_col="ensembl_id",
        )
        grn.varp["GRN"] = grn.varp["all"][:, :, clf_self.coef_[0] > 0].mean(-1).T
        metrics["self"] = BenGRN(grn, do_auc=True, doplot=False).compare_to(other=adata)
        metrics["self_classifier"] = m
        grn.var.index = grn.var["symbol"]
        grn.varp["GRN"] = grn.varp["GRN"].T
        metrics["self_base"] = BenGRN(
            grn, do_auc=True, doplot=False
        ).scprint_benchmark()
    else:
        # max_genes=4000
        adata = sc.read_h5ad(default_dataset)
        adata.var["isTF"] = False
        adata.var.loc[adata.var.symbol.isin(grnutils.TF), "isTF"] = True
        for celltype in cell_types:
            print(celltype)
            grn_inferer = GNInfer(
                layer=layers,
                how="random expr",
                preprocess="softmax",
                head_agg="max",
                filtration="none",
                forward_mode="none",
                num_workers=8,
                num_genes=2200,
                max_cells=maxcells,
                doplot=False,
                batch_size=batch_size,
                devices=1,
            )

            grn = grn_inferer(model, adata[adata.X.sum(1) > 500], cell_type=celltype)
            grn.var.index = make_index_unique(grn.var["symbol"].astype(str))
            metrics[celltype + "_scprint"] = BenGRN(
                grn, doplot=False
            ).scprint_benchmark()
            del grn
            gc.collect()
            grn_inferer = GNInfer(
                layer=layers,
                how="most var across",
                preprocess="softmax",
                head_agg="none",
                filtration="none",
                forward_mode="none",
                num_workers=8,
                num_genes=maxgenes,
                max_cells=maxcells,
                doplot=False,
                batch_size=batch_size,
                devices=1,
            )
            grn = grn_inferer(model, adata[adata.X.sum(1) > 500], cell_type=celltype)
            grn.var.index = make_index_unique(grn.var["symbol"].astype(str))
            grn.varp["all"] = grn.varp["GRN"]
            grn.varp["GRN"] = grn.varp["GRN"].mean(-1)
            metrics[celltype + "_scprint_mean"] = BenGRN(
                grn, doplot=False
            ).scprint_benchmark()
            if clf_omni is None:
                grn.varp["GRN"] = grn.varp["all"]
                _, m, clf_omni = train_classifier(
                    grn,
                    C=1,
                    train_size=0.6,
                    max_iter=300,
                    class_weight={1: 800, 0: 1},
                    return_full=False,
                    shuffle=True,
                    doplot=False,
                )
                joblib.dump(clf_omni, "clf_omni.pkl")
                metrics["classifier"] = m
            grn.varp["GRN"] = grn.varp["all"][:, :, clf_omni.coef_[0] > 0].mean(-1)
            metrics[celltype + "_scprint_class"] = BenGRN(
                grn, doplot=False
            ).scprint_benchmark()
            del grn
            gc.collect()
    return metrics

scprint.tasks.denoise

Denoiser

Denoiser class for denoising scRNA-seq data using a scPRINT model

Parameters:
  • batch_size (int, default: 10 ) –

    Batch size for processing. Defaults to 10.

  • num_workers (int, default: 1 ) –

    Number of workers for data loading. Defaults to 1.

  • max_len (int, default: 5000 ) –

    Maximum number of genes to consider. Defaults to 5000.

  • precision (str, default: '16-mixed' ) –

    Precision type for computations. Defaults to "16-mixed".

  • how (str, default: 'most var' ) –

    Method to select genes. Options are "most var". Defaults to "most var".

  • max_cells (int, default: 500000 ) –

    Number of cells to use for plotting correlation. Defaults to 10000.

  • doplot (bool, default: False ) –

    Whether to generate plots. Defaults to False.

  • predict_depth_mult (int, default: 4 ) –

    Multiplier for prediction depth. Defaults to 4.

  • downsample (Optional[float], default: None ) –

    Fraction of data to downsample. Defaults to None.

  • devices (List[int]) –

    List of device IDs to use. Defaults to [0].

  • dtype (dtype, default: float16 ) –

    Data type for computations. Defaults to torch.float16.
Source code in scprint/tasks/denoise.py 
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
def __init__(
    self,
    batch_size: int = 10,
    num_workers: int = 1,
    max_len: int = 5_000,
    precision: str = "16-mixed",
    how: str = "most var",
    max_cells: int = 500_000,
    doplot: bool = False,
    predict_depth_mult: int = 4,
    downsample: Optional[float] = None,
    dtype: torch.dtype = torch.float16,
):
    """
    Denoiser class for denoising scRNA-seq data using a scPRINT model

    Args:
        batch_size (int, optional): Batch size for processing. Defaults to 10.
        num_workers (int, optional): Number of workers for data loading. Defaults to 1.
        max_len (int, optional): Maximum number of genes to consider. Defaults to 5000.
        precision (str, optional): Precision type for computations. Defaults to "16-mixed".
        how (str, optional): Method to select genes. Options are "most var". Defaults to "most var".
        max_cells (int, optional): Number of cells to use for plotting correlation. Defaults to 10000.
        doplot (bool, optional): Whether to generate plots. Defaults to False.
        predict_depth_mult (int, optional): Multiplier for prediction depth. Defaults to 4.
        downsample (Optional[float], optional): Fraction of data to downsample. Defaults to None.
        devices (List[int], optional): List of device IDs to use. Defaults to [0].
        dtype (torch.dtype, optional): Data type for computations. Defaults to torch.float16.
    """
    self.batch_size = batch_size
    self.num_workers = num_workers
    self.max_len = max_len
    self.max_cells = max_cells
    self.doplot = doplot
    self.predict_depth_mult = predict_depth_mult
    self.how = how
    self.downsample = downsample
    self.precision = precision
    self.dtype = dtype

__call__

call calling the function

Parameters:
  • model (Module) –

    The scPRINT model to be used for denoising.

  • adata (AnnData) –

    The annotated data matrix of shape n_obs x n_vars. Rows correspond to cells and columns to genes.
Returns:
  • AnnData –

    The denoised annotated data matrix.
Source code in scprint/tasks/denoise.py 
 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
def __call__(self, model: torch.nn.Module, adata: AnnData):
    """
    __call__ calling the function

    Args:
        model (torch.nn.Module): The scPRINT model to be used for denoising.
        adata (AnnData): The annotated data matrix of shape n_obs x n_vars. Rows correspond to cells and columns to genes.

    Returns:
        AnnData: The denoised annotated data matrix.
    """
    if os.path.exists("collator_output.txt"):
        os.remove("collator_output.txt")
    random_indices = None
    if self.max_cells < adata.shape[0]:
        random_indices = np.random.randint(
            low=0, high=adata.shape[0], size=self.max_cells
        )
        adataset = SimpleAnnDataset(
            adata[random_indices], obs_to_output=["organism_ontology_term_id"]
        )
    else:
        adataset = SimpleAnnDataset(
            adata, obs_to_output=["organism_ontology_term_id"]
        )
    if self.how == "most var":
        sc.pp.highly_variable_genes(
            adata, flavor="seurat_v3", n_top_genes=self.max_len, span=0.99
        )
        genelist = adata.var.index[adata.var.highly_variable]
        print(len(genelist))
    col = Collator(
        organisms=model.organisms,
        valid_genes=model.genes,
        max_len=self.max_len,
        how="some" if self.how == "most var" else self.how,
        genelist=genelist if self.how == "most var" else [],
        downsample=self.downsample,
        save_output=True,
    )
    dataloader = DataLoader(
        adataset,
        collate_fn=col,
        batch_size=self.batch_size,
        num_workers=self.num_workers,
        shuffle=False,
    )

    model.doplot = self.doplot
    model.on_predict_epoch_start()
    model.eval()
    device = model.device.type
    with torch.no_grad(), torch.autocast(device_type=device, dtype=self.dtype):
        for batch in tqdm(dataloader):
            gene_pos, expression, depth = (
                batch["genes"].to(device),
                batch["x"].to(device),
                batch["depth"].to(device),
            )
            model._predict(
                gene_pos,
                expression,
                depth,
                predict_mode="denoise",
                depth_mult=self.predict_depth_mult,
            )
    torch.cuda.empty_cache()
    self.genes = (
        model.pos.cpu().numpy()
        if self.how != "most var"
        else list(set(model.genes) & set(genelist))
    )
    tokeep = None
    metrics = None
    if self.downsample is not None:
        reco = model.expr_pred[0]
        reco = reco.cpu().numpy()
        tokeep = np.isnan(reco).sum(1) == 0
        reco = reco[tokeep]
        noisy = np.loadtxt("collator_output.txt")[tokeep]

        if random_indices is not None:
            true = adata.X[random_indices][tokeep]
        else:
            true = adata.X[tokeep]
        true = true.toarray() if issparse(true) else true
        if self.how == "most var":
            true = true[:, adata.var.index.isin(self.genes)]
            # noisy[true==0]=0
        else:
            true = np.vstack(
                [
                    true[
                        i,
                        adata.var.index.get_indexer(np.array(model.genes)[val]),
                    ].copy()
                    for i, val in enumerate(self.genes)
                ]
            )
        # reco[true==0] = 0
        # import pdb
        # pdb.set_trace()
        # reco[reco!=0] = 2
        # corr_coef = np.corrcoef(
        #    np.vstack([reco[true!=0], noisy[true!=0], true[true!=0]])
        # )
        corr_coef, p_value = spearmanr(
            np.vstack([reco[true != 0], noisy[true != 0], true[true != 0]]).T
        )
        metrics = {
            "reco2noisy": corr_coef[0, 1],
            "reco2full": corr_coef[0, 2],
            "noisy2full": corr_coef[1, 2],
        }
        # corr_coef[p_value > 0.05] = 0
        # if self.doplot:
        #    plt.figure(figsize=(10, 5))
        #    plt.imshow(
        #        corr_coef, cmap="coolwarm", interpolation="none", vmin=-1, vmax=1
        #    )
        #    plt.colorbar()
        #    plt.title("Expression Correlation Coefficient")
        #    plt.show()
        # metrics = {
        #    "reco2noisy": np.mean(
        #        corr_coef[
        #            self.max_cells : self.max_cells * 2, : self.max_cells
        #        ].diagonal()
        #    ),
        #    "reco2full": np.mean(
        #        corr_coef[self.max_cells * 2 :, : self.max_cells].diagonal()
        #    ),
        #    "noisy2full": np.mean(
        #        corr_coef[
        #            self.max_cells * 2 :,
        #            self.max_cells : self.max_cells * 2,
        #        ].diagonal()
        #    ),
        # }
    return (
        metrics,
        (
            random_indices[tokeep]
            if random_indices is not None and tokeep is not None
            else random_indices
        ),
        self.genes,
        (
            model.expr_pred[0][tokeep].cpu().numpy()
            if tokeep is not None
            else model.expr_pred[0].cpu().numpy()
        ),
    )

default_benchmark

default_benchmark function used to run the default denoising benchmark of scPRINT

Parameters:
  • model (Any) –

    The scPRINT model to be used for the benchmark.

  • default_dataset (str, default: FILE_DIR + '/../../data/r4iCehg3Tw5IbCLiCIbl.h5ad' ) –

    Path to the default dataset to use for benchmarking. Defaults to FILE_DIR + "/../../data/r4iCehg3Tw5IbCLiCIbl.h5ad".

  • max_len (int, default: 5000 ) –

    Maximum number of genes to consider. Defaults to 5000.
Returns:
  • dict –

    A dictionary containing the benchmark metrics.
Source code in scprint/tasks/denoise.py 
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
def default_benchmark(
    model: Any,
    default_dataset: str = FILE_DIR + "/../../data/r4iCehg3Tw5IbCLiCIbl.h5ad",
    max_len: int = 5000,
):
    """
    default_benchmark function used to run the default denoising benchmark of scPRINT

    Args:
        model (Any): The scPRINT model to be used for the benchmark.
        default_dataset (str, optional): Path to the default dataset to use for benchmarking. Defaults to FILE_DIR + "/../../data/r4iCehg3Tw5IbCLiCIbl.h5ad".
        max_len (int, optional): Maximum number of genes to consider. Defaults to 5000.

    Returns:
        dict: A dictionary containing the benchmark metrics.
    """
    adata = sc.read_h5ad(default_dataset)
    denoise = Denoiser(
        model,
        batch_size=40,
        max_len=max_len,
        max_cells=10_000,
        doplot=False,
        num_workers=8,
        predict_depth_mult=10,
        downsample=0.7,
        devices=1,
    )
    return denoise(adata)[0]

split_molecules

Splits molecules into two (potentially overlapping) groups. :param umis: Array of molecules to split :param data_split: Proportion of molecules to assign to the first group :param overlap_factor: Overlap correction factor, if desired :param random_state: For reproducible sampling :return: umis_X and umis_Y, representing split and ~(1 - split) counts sampled from the input array

Source code in scprint/tasks/denoise.py 
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
def split_molecules(
    umis: np.ndarray,
    data_split: float,
    overlap_factor: float = 0.0,
    random_state: np.random.RandomState = None,
) -> Tuple[np.ndarray, np.ndarray]:
    """Splits molecules into two (potentially overlapping) groups.
    :param umis: Array of molecules to split
    :param data_split: Proportion of molecules to assign to the first group
    :param overlap_factor: Overlap correction factor, if desired
    :param random_state: For reproducible sampling
    :return: umis_X and umis_Y, representing ``split`` and ``~(1 - split)`` counts
             sampled from the input array
    """
    if random_state is None:
        random_state = np.random.RandomState()

    umis_X_disjoint = random_state.binomial(umis, data_split - overlap_factor)
    umis_Y_disjoint = random_state.binomial(
        umis - umis_X_disjoint, (1 - data_split) / (1 - data_split + overlap_factor)
    )
    overlap_factor = umis - umis_X_disjoint - umis_Y_disjoint
    umis_X = umis_X_disjoint + overlap_factor
    umis_Y = umis_Y_disjoint + overlap_factor

    return umis_X, umis_Y