marginal
EQTLPredictor
Bases: Module
Predictor for eQTL models.
The challenge with eQTL analysis is that the SNF predictors can vary from gene to gene and that we must support donor-level effects as well as cell-level features. This requires a different form of forward function to keep us from running separate regressions across genes or copying the same patient-level information across all cells. For modeling the mean, the basic structure is
eta = W @ coefs["mean"] # shared block
+ U[donor_idx] # per-individual random intercept
+ gather(einsum('kjm,jm->kj', D, beta)) # genotype main effect
+ gather(einsum('kjm,jml->kjl', D, zeta)) # genotype interaction
Args: n_outcomes: number of genes (J) feature_dims: mapping with keys "mean" and "dispersion" -> number of shared covariate columns for each dosage: (K, J, K_max) individual x gene x SNP position tensor tracking genotypes mask: (J, K_max) indicator for valid SNP entries per gene. The issue is that some genes might only be located close to a small number of SNPs. interaction_dim: number of columns (L) in the genotype-interaction design, not counting the intercept column. 0 means no interactions. accumulator_names: Names of additional statistical summary statistics to store while accumulating each epoch. These are necessary for our posterior moment updates described in donor_moments.py. Appropriate values are,
score = +l' observed = -l''
third = -l''' fourth = -l''''
Source code in scdesigner/src/scdesigner/base/marginal.py
787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 | |
accumulators
property
Return the epoch accumulators as a dictionary keyed by derivative.
forward(obs_dict)
Forward pass to assemble mean prediction
obs_dict must contain: "mean": (B, p) shared mean design "dispersion": (B, p_disp) shared dispersion design "genotype": (B, 1) integer-valued donor code column "interaction": (B, L + 1) genotype-interaction design (only read when interaction_dim > 0). Its first column is the intercept, which we drop so that zeta isn't collinear with beta.
Source code in scdesigner/src/scdesigner/base/marginal.py
900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 | |
zero_accumulators()
Reset all per-epoch accumulators.
Source code in scdesigner/src/scdesigner/base/marginal.py
890 891 892 893 894 | |
GLMPredictor
Bases: Module
GLM-style predictor with arbitrary named parameters.
Args: n_outcomes: number of model outputs (e.g. genes) feature_dims: mapping from param name -> number of covariate features link_fns: optional mapping from param name -> callable(link) applied to linear predictor
The module will create one coefficient matrix per named parameter with shape
(n_features_for_param, n_outcomes) and expose them as Parameters under
self.coefs[param_name].
Source code in scdesigner/src/scdesigner/base/marginal.py
722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 | |
forward(obs_dict)
Forward Pass for Given Covariates
obs_dict : Dict of (str -> torch.Tensor) A dictionary of tensors, with one key/value pair per parameter. These tensors are the conditioning information to pass to the .predict() function of this distribution class. They are the numerical design matrices implied by the initializing formulas.
Source code in scdesigner/src/scdesigner/base/marginal.py
769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 | |
Marginal
Bases: ABC
A Feature-wise Marginal Model
This is a class for handling feature-wise (e.g., gene-level) modeling that ignores any correlation between features. For example, it can be used to model the relationship between experimental design features, such as cell type or treatment, and the parameters of a collection of negative binomial models (one for each gene).
These marginals can be plugged into a copula to build the complete scDesign3 simulator. Methods are expected for estimating model parameters, evaluating likelihoods on new samples, and generating new samples given conditioning information. Since these marginal models are intended to be used within a copula, they should also provide utilities for evaluating quantiles and computing cumulative distribution functions.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
formula
|
dict or str
|
A dictionary or string specifying the relationship between the columns of an input data frame (adata.obs, adata.var, or similar attributes) and the parameters of the marginal model. If only a string is provided, then the means are allowed to depend on the design parameters, while all other parameters are treated as fixed. If a dictionary is provided, each key should correspond to a parameter. The string values should be in a format that can be parsed by the formulaic package. For example, '~ x' will ensure that the parameter varies linearly with X. |
required |
penalties
|
sequence of CoefficientPenalty
|
Penalties |
()
|
Attributes:
| Name | Type | Description |
|---|---|---|
formula |
dict or str
|
A dictionary or string specifying the relationship between the columns of an input data frame (adata.obs, adata.var, or similar attributes) and the parameters of the marginal model. If only a string is provided, then the means are allowed to depend on the design parameters, while all other parameters are treated as fixed. If a dictionary is provided, each key should correspond to a parameter. The string values should be in a format that can be parsed by the formulaic package. For example, '~ x' will ensure that the parameter varies linearly with X. |
penalties |
tuple of CoefficientPenalty
|
|
chunk_size |
int or None
|
For backed anndata objects, we read continguous blocks of rows at a
time. This parameter governs the size of those chunks and is used in
:meth: |
design |
Design
|
Design matrix factory initially adapted to the full |
feature_dims |
dict
|
A dictionary containing the number of predictors associated with each distributional parameter. Note that this number is repeated for every feature (e.g., gene) in the marginal model. This information is often useful for computing the complexity of the estimated model. |
loader |
DataLoader
|
A torch DataLoader object that returns batches of data for use during training. This loader is constructed internally within the setup_data method. Enumerating this loader returns a tuple: the first element contains a tensor of feature measurements (y), and the second element is a dictionary of tensors containing predictors to use for each parameter (x, for each parameter theta(x)). This design is useful because the design matrices may differ between parameters of the marginal model, y | x ~ F_(theta(x))(y) |
train_loader |
DataLoader
|
Loader used for marginal optimization. With early stopping enabled,
this is a subset of :attr: |
validation_loader |
DataLoader or None
|
Held-out loader used to monitor unpenalized validation likelihood. |
fit_history |
list of dict
|
Per-epoch training and validation losses, with best/stop indicators. |
_validation_split_config |
tuple or None
|
Cached |
n_outcomes |
int
|
The number of features modeled by this marginal model. For example, this corresponds to the number of genes being simulated. |
predict |
Module
|
A torch.nn.Module storing the relationship between predictors for each parameter and the predicted feature-wise outcomes. This module is expected to take the second element of the tuple defined by each batch and then predict a tensor with the same shape as the first element of the batch tuple. |
predictor_names |
dict of list of str
|
A dictionary whose keys are the parameter names associated with this marginal model. The values for each key are the names of predictors in the design matrix implied by the associated formula. Note that these names may have been expanded from the original formula specification. For example, if cell_type is included in the formula, then the predictor names will include the unique levels of cell_type as separate columns in the design matrix and therefore as separate elements in this list. |
parameters |
dict of pandas.DataFrame
|
A dictionary whose keys are the parameter names associated with this marginal model. The values for each key are pandas DataFrames storing the fitted parameter values. The rows of each DataFrame are the experimental features specified by the associated formula object (the rownames are the same as those in predictor_names). The columns are features that are being predicted. |
device |
device
|
The device (e.g., cpu, mps, cuda) used by the model and data loaders.
You can set this when initializing the model and change it with
:meth: |
Examples:
>>> class DummyModel(Marginal):
... def fit(self):
... pass
...
... def likelihood(self, batch: Tuple[torch.Tensor, Dict[str, torch.Tensor]]):
... pass
...
... def uniformize(self, y: torch.Tensor, x: Dict[str, torch.Tensor]):
... pass
...
... def invert(self, u: torch.Tensor, x: Dict[str, torch.Tensor]):
... pass
...
... def setup_optimizer(self):
... pass
...
>>> model = DummyModel("~ cell_type")
>>> model.fit()
Source code in scdesigner/src/scdesigner/base/marginal.py
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 | |
fit_history_df
property
Return the fit history as a pandas DataFrame.
fit(max_epochs=2000, verbose=True, val_frac=0.1, min_epochs=10, loss_tol=0.0001, patience=40, validation_seed=0, **kwargs)
Fit the marginal predictor using vanilla PyTorch training loop.
This method runs stochastic gradient optimization using the template dataset defined by the setup_data method. The specific optimizer used can be modified with the setup_optimizer method and defaults to Adam.
Note that, unlike fit in class Simulator, this method does not allow
the template dataset as input. This requires .setup_data() to be
called first. We want to give finer-grained control over the data
loading and optimization in this class relative to the specific
Simulator implementations, which are designed to be easy to run with
as few steps as possible.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
max_epochs
|
int
|
The maximum number of epochs. This is the number of times we feed through our cells in the dataset. |
2000
|
verbose
|
bool
|
Should we print intermediate training outputs? |
True
|
val_frac
|
float
|
Fraction of observations held out for validation. Set to 0 to disable early stopping and train on the full dataset. |
0.1
|
min_epochs
|
int
|
Minimum number of completed epochs before early stopping is allowed. |
10
|
loss_tol
|
float
|
Required relative decrease in validation loss to reset patience.
For example, |
0.0001
|
patience
|
int
|
Number of non-improving validation epochs allowed after the model is
trained for |
40
|
validation_seed
|
int
|
Seed controlling the deterministic validation split. |
0
|
Returns:
| Type | Description |
|---|---|
None
|
This method doesn't return anything but modifies the self.parameters attribute with the trained model parameters. |
Source code in scdesigner/src/scdesigner/base/marginal.py
416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 | |
format_parameters()
Convert fitted coefficient tensors into pandas DataFrames.
Returns:
dict: mapping from parameter name -> pandas.DataFrame with rows
corresponding to predictor column names (from
self.predictor_names[param]) and columns corresponding to
self.adata.var_names (gene names). The values are moved to
CPU and converted to numpy floats.
Source code in scdesigner/src/scdesigner/base/marginal.py
584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 | |
invert(u, x)
abstractmethod
Invert pseudoobservations.
Return a quantile from the distribution. This handles the link between the marginal and the copula model. The idea is that the copula will generate pseudo-observations on the unit cube. All the values will be between zero and one. By calling invert, we transform these zero-to-one valued pseudo-observations into observations in the original data space; values between zero and one can be thought of like quantiles.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
u
|
Tensor
|
Scalars between [0, 1] that specify the quantile level we want. |
required |
x
|
Dict of (str -> torch.Tensor)
|
A dictionary of tensors, with one key/value pair per parameter. These tensors are the conditioning information to pass to the .predict() function of this distribution class. They are the numerical design matrices implied by the initializing formulas. |
required |
Returns:
| Name | Type | Description |
|---|---|---|
z |
Tensor
|
A tensor with dimension dim(u) x num_genes. Each row gives the requested quantile for the marginal distributions across all genes. |
Source code in scdesigner/src/scdesigner/base/marginal.py
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 | |
likelihood(batch)
abstractmethod
Compute the log-likelihood for a batch.
The likelihood is used for maximum likelihood estimation. It is also used when computing AIC and BIC scores, which are important when choosing an appropriate model complexity.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
batch
|
tuple of (torch.Tensor, dict of str -> torch.Tensor)
|
A tuple of gene expression (y) and experimental factors (x) used to evaluate the model likelihood. The first element of the tuple is a cells x genes tensor. The second is a a dictionary of tensors, with one key/value pair per parameter. These tensors are the conditioning information to pass to the .predict() function of this distribution class. They are the numerical design matrices implied by the initializing formulas. |
required |
Returns:
| Type | Description |
|---|---|
Tensor
|
A scalar containing the log-likelihood of the batch under the current model parameters. |
Source code in scdesigner/src/scdesigner/base/marginal.py
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 | |
num_params()
Return the number of parameters.
Count the number of parameters in the marginal simulator. Usually this is just the number of predictors times the number of genes, because we use a linear model. However, in specific implementations, it’s possible to use more flexible models, in which case the number of parameters would increase.
Source code in scdesigner/src/scdesigner/base/marginal.py
603 604 605 606 607 608 609 610 611 612 613 614 | |
setup_data(adata, batch_size=1024, *, chunk_size=None, observation_transform=None, **kwargs)
Set up the dataloader for the AnnData object.
The simulator class definition doesn’t actually require any particular template dataset. This is helpful for reasoning about simulators abstractly, but when we actually want to estimate parameters, we need a template. This method takes an input template dataset and adds attributes to the simulator object for later estimation and sampling steps.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
adata
|
AnnData
|
This is the object on which we want to estimate the simulator. This serves as the template for all downstream fitting. |
required |
batch_size
|
int
|
The number of sample to return on each call of the data loader. Defaults to 1024. |
1024
|
chunk_size
|
int or None
|
For backed anndata objects, we read continguous blocks of rows at a
time. This parameter governs the size of those chunks and is used in
:meth: |
None
|
observation_transform
|
SpatialDesign
|
Transform that appends spatial basis columns to |
None
|
**kwargs
|
Any
|
Other keyword arguments passed to data loader construction. Any argument recognized by the PyTorch DataLoader can be passed in here. |
{}
|
Returns:
| Type | Description |
|---|---|
None
|
This method doesn't return anything but modifies the self.adata, self.design, self.loader, self.n_outcomes, self.feature_dims, and self.predictor_names attributes. |
Source code in scdesigner/src/scdesigner/base/marginal.py
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 | |
setup_validation_split(val_frac=0.1, validation_seed=0)
Create train and validation loaders that split the full dataset.
The full-data loader remains available as :attr:loader for downstream
workflows such as copula fitting. When validation is disabled or the
dataset is too small to split, :attr:train_loader points to
:attr:loader and :attr:validation_loader is None.
Source code in scdesigner/src/scdesigner/base/marginal.py
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 | |
to(device)
Move the predictor, optimizer state, and loaded data to device.
Source code in scdesigner/src/scdesigner/base/marginal.py
340 341 342 343 344 345 346 347 348 349 350 351 | |
uniformize(y, x)
abstractmethod
Uniformize using learned CDF.
Apply a quantile/CDF transformation to observations y, accounting for conditioning variables x. This step is used in training the copula model. Since the copula needs to operate on the unit cube, we need to transform the original data onto the unit cube. This can be done by applying a CDF transformation.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
y
|
Tensor
|
A cells x genes tensor with gene expression levels across all cells. |
required |
x
|
Dict of (str -> torch.Tensor)
|
A dictionary of tensors, with one key/value pair per parameter. These tensors are the conditioning information to pass to the .predict() function of this distribution class. They are the numerical design matrices implied by the initializing formulas. |
required |
Source code in scdesigner/src/scdesigner/base/marginal.py
699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 | |