Scvi-tools

Added

- Add {class}`scvi.external.Decipher` for dimensionality reduction and interpretable
representation learning in single-cell RNA sequencing data {pr}`3015`, {pr}`3091`.
- Add multiGPU support for {class}`~scvi.model.SCVI`, {class}`~scvi.model.SCANVI`,
{class}`~scvi.model.CondSCVI` and {class}`~scvi.model.LinearSCVI`, {class}`~scvi.model.TOTALVI`,
{class}`~scvi.model.MULTIVI` and {class}`~scvi.model.PEAKVI`. {pr}`3125`.
- Add an exception callback to {class}`scvi.train._callbacks.SaveCheckpoint` in order to save
optimal model during training, in case of failure because of Nan's in gradients. {pr}`3159`.
- Add {meth}`~scvi.model.SCVI.get_normalized_expression` for models: {class}`~scvi.model.PEAKVI`,
{class}`~scvi.external.POISSONVI`, {class}`~scvi.model.CondSCVI`, {class}`~scvi.model.AUTOZI`,
{class}`~scvi.external.CellAssign` and {class}`~scvi.external.GIMVI`. {pr}`3121`.
- Add {class}`scvi.external.RESOLVI` for bias correction in single-cell resolved spatial
transcriptomics {pr}`3144`.
- Add semisupervised training mixin class {class}`scvi.model.base.SemisupervisedTrainingMixin` {pr}`3164`.
- Add scib-metrics support for {class}`scvi.autotune.AutotuneExperiment` and
{class}`scvi.train._callbacks.ScibCallback` for autotune for scib metrics {pr}`3168`.
- Add Support of dask arrays in AnnTorchDataset. {pr}`3193`.
- Add a {doc}`/user_guide/use_case` section in the docs, {pr}`3200`.
- Add {class}`scvi.external.SysVI` for cycle consistency loss and VampPrior {pr}`3195`.

Fixed

- Fixed bug in distributed {class}`scvi.dataloaders.ConcatDataLoader` {pr}`3053`.
- Fixed bug when loading Pyro-based models and scArches support for Pyro {pr}`3138`
- Fixed disable vmap in {class}`scvi.external.MRVI` for large sample sizes to avoid
out-of-memory errors. Store distance matrices as numpy array in xarray to reduce
memory usage {pr}`3146`.
- Fixed {class}`scvi.external.MRVI` MixtureSameFamily log probability calculation {pr}`3189`.

Changed

- Updated the CI workflow with multiGPU tests {pr}`3053`.
- Set `mode="change"` as default DE method. Compute positive and negative LFC separately
by default (`test_mode="three"`). Corrected computation of pseudocounts and make if
default to add a pseudocounts for genes not expressed (`pseudocount=None`). According to
Eq. 10 of Boyeau _et al_, _PNAS_ 2023 {pr}`2826`

Removed

Added

- Add MuData Minification option to {class}`~scvi.model.TOTALVI` {pr}`3061`.
- Add Support for MPS usage in mac {pr}`3100`.
- Add support for torch.compile before train (EXPERIMENTAL) {pr}`2931`.
- Add support for Numpy 2.0 {pr}`2842`.
- Changed scvi-hub ModelCard and add criticism metrics to the card {pr}`3078`.
- MuData support for {class}`~scvi.model.MULTIVI` via the method
{meth}`~scvi.model.MULTIVI.setup_mudata` {pr}`3038`.

Fixed

- Fixed batch_size pop to get in {class}`scvi.dataloaders.DataSplitter` {pr}`3128`.

Changed

- Updated the CI workflow with internet, private and optional tests {pr}`3082`.
- Changed loompy stored files to anndata {pr}`2842`.
- Address AnnData >= 0.11 deprecation warning for {class}`anndata.experimental` by replacing
instances to {class}`anndata.abc` and {class}`anndata.io` {pr}`3085`.

Removed

- Removed the support for loompy and local mde function {pr}`2842`.

Added

- Added adaptive handling for last training minibatch of 1-2 cells in case of
`datasplitter_kwargs={"drop_last": False}` and `train_size = None` by moving them into
validation set, if available. {pr}`3036`.
- Add `batch_key` and `labels_key` to `scvi.external.SCAR.setup_anndata`. {pr}`3045`.
- Implemented variance of ZINB distribution. {pr}`3044`.
- Support for minified mode while retaining counts to skip the encoder.
- New Trainingplan argument `update_only_decoder` to use stored latent codes and skip training of
the encoder.
- Refactored code for minified models. {pr}`2883`.
- Add {class}`scvi.external.METHYLVI` for modeling methylation data from single-cell
bisulfite sequencing (scBS-seq) experiments {pr}`2834`.

Fixed

- Breaking Change: Fix `get_outlier_cell_sample_pairs` function in {class}`scvi.external.MRVI`
to correctly compute the maxmimum log-density across in-sample cells rather than the
aggregated posterior log-density {pr}`3007`.
- Fix references to `scvi.external` in `scvi.external.SCAR.setup_anndata`.
- Fix gimVI to append mini batches first into CPU during get_imputed and get_latent operations {pr}`3058`.

Changed

Removed

Added

- Add support for Python 3.12 {pr}`2966`.
- Add support for categorial covariates in scArches in `scvi.model.archesmixin` {pr}`2936`.
- Add assertion error in cellAssign for checking duplicates in celltype markers {pr}`2951`.
- Add `scvi.external.poissonvi.get_region_factors` {pr}`2940`.
- {attr}`scvi.settings.dl_persistent_workers` allows using persistent workers in
{class}`scvi.dataloaders.AnnDataLoader` {pr}`2924`.
- Add option for using external indexes in data splitting classes that are under `scvi.dataloaders`
by passing `external_indexing=list[train_idx,valid_idx,test_idx]` as well as in all models
available {pr}`2902`.
- Add warning if creating data splits in `scvi.dataloaders` that create last batch with less than 3
cells {pr}`2916`.
- Add new experimental functional API for hyperparameter tuning with
{func}`scvi.autotune.run_autotune` and {class}`scvi.autotune.AutotuneExperiment` to replace
{class}`scvi.autotune.ModelTuner`, {class}`scvi.autotune.TunerManager`, and
{class}`scvi.autotune.TuneAnalysis` {pr}`2561`.
- Add experimental class {class}`scvi.nn.Embedding` implementing methods for extending embeddings
{pr}`2574`.
- Add experimental support for representing batches with continuously-valued embeddings by passing
in `batch_representation="embedding"` to {class}`scvi.model.SCVI` {pr}`2576`.
- Add experimental mixin classes {class}`scvi.model.base.EmbeddingMixin` and
{class}`scvi.module.base.EmbeddingModuleMixin` {pr}`2576`.
- Add option to generate synthetic spatial coordinates in {func}`scvi.data.synthetic_iid` with
argument `generate_coordinates` {pr}`2603`.
- Add experimental support for using custom {class}`lightning.pytorch.core.LightningDataModule`s
in {func}`scvi.autotune.run_autotune` {pr}`2605`.
- Add {class}`scvi.external.VELOVI` for RNA velocity estimation using variational inference
{pr}`2611`.
- Add `unsigned` argument to {meth}`scvi.hub.HubModel.pull_from_s3` to allow for unsigned
downloads of models from AWS S3 {pr}`2615`.
- Add support for `batch_key` in {meth}`scvi.model.CondSCVI.setup_anndata` {pr}`2626`.
- Add support for {meth}`scvi.model.base.RNASeqMixin` in {class}`scvi.model.CondSCVI` {pr}`2915`.
- Add `load_best_on_end` argument to {class}`scvi.train.SaveCheckpoint` to load the best model
state at the end of training {pr}`2672`.
- Add experimental class {class}`scvi.distributions.BetaBinomial` implementing the Beta-Binomial
distribution with mean-dispersion parameterization for modeling scBS-seq methylation data
{pr}`2692`.
- Add support for custom dataloaders in {class}`scvi.model.base.VAEMixin` methods by specifying
the `dataloader` argument {pr}`2748`.
- Add option to use a normal distribution in the generative model of {class}`scvi.model.SCVI` by
passing in `gene_likelihood="normal"` {pr}`2780`.
- Add {class}`scvi.external.MRVI` for modeling sample-level heterogeneity in single-cell RNA-seq
data {pr}`2756`.
- Add support for reference mapping with {class}`mudata.MuData` models to
{class}`scvi.model.base.ArchesMixin` {pr}`2578`.
- Add argument `return_mean` to {meth}`scvi.model.base.VAEMixin.get_reconstruction_error`
and {meth}`scvi.model.base.VAEMixin.get_elbo` to allow computation
without averaging across cells {pr}`2362`.
- Add support for setting `weights="importance"` in
{meth}`scvi.model.SCANVI.differential_expression` {pr}`2362`.

Changed

- Deprecate {func}`scvi.data.cellxgene`, to be removed in v1.3. Please directly use the
[cellxgene-census](https://chanzuckerberg.github.io/cellxgene-census/) instead {pr}`2542`.
- Deprecate {func}`scvi.nn.one_hot`, to be removed in v1.3. Please directly use the
`one_hot` function in PyTorch instead {pr}`2608`.
- Deprecate {class}`scvi.train.SaveBestState`, to be removed in v1.3. Please use
{class}`scvi.train.SaveCheckpoint` instead {pr}`2673`.
- Deprecate `save_best` argument in {meth}`scvi.model.PEAKVI.train` and
{meth}`scvi.model.MULTIVI.train`, to be removed in v1.3. Please pass in `enable_checkpointing`
or specify a custom checkpointing procedure with {class}`scvi.train.SaveCheckpoint` instead
{pr}`2673`.
- Move {func}`scvi.model.base._utils._load_legacy_saved_files` to
{func}`scvi.model.base._save_load._load_legacy_saved_files` {pr}`2731`.
- Move {func}`scvi.model.base._utils._load_saved_files` to
{func}`scvi.model.base._save_load._load_saved_files` {pr}`2731`.
- Move {func}`scvi.model.base._utils._initialize_model` to
{func}`scvi.model.base._save_load._initialize_model` {pr}`2731`.
- Move {func}`scvi.model.base._utils._validate_var_names` to
{func}`scvi.model.base._save_load._validate_var_names` {pr}`2731`.
- Move {func}`scvi.model.base._utils._prepare_obs` to
{func}`scvi.model.base._de_core._prepare_obs` {pr}`2731`.
- Move {func}`scvi.model.base._utils._de_core` to
{func}`scvi.model.base._de_core._de_core` {pr}`2731`.
- Move {func}`scvi.model.base._utils._fdr_de_prediction` to
{func}`scvi.model.base._de_core_._fdr_de_prediction` {pr}`2731`.
- {func}`scvi.data.synthetic_iid` now generates unique variable names for protein and
accessibility data {pr}`2739`.
- The `data_module` argument in {meth}`scvi.model.base.UnsupervisedTrainingMixin.train` has been
renamed to `datamodule` for consistency {pr}`2749`.
- Change the default saving method of variable names for {class}`mudata.MuData` based models
(_e.g._ {class}`scvi.model.TOTALVI`) to a dictionary of per-mod variable names instead of a
concatenated array of all variable names. Users may replicate the previous behavior by
passing in `legacy_mudata_format=True` to {meth}`scvi.model.base.BaseModelClass.save`
{pr}`2769`.
- Changed internal activation function in {class}`scvi.nn.DecoderTOTALVI` to Softplus to
increase numerical stability. This is the new default for new models. Previously trained models
will be loaded with exponential activation function {pr}`2913`.

Fixed

- Fix logging of accuracy for cases with 1 sample per class in scANVI {pr}`2938`.
- Disable adversarial classifier if training with a single batch.
Previously this raised a None error {pr}`2914`.
- {meth}`~scvi.model.SCVI.get_normalized_expression` fixed for Poisson distribution and
Negative Binomial with latent_library_size {pr}`2915`.
- Fix {meth}`scvi.module.VAE.marginal_ll` when `n_mc_samples_per_pass=1` {pr}`2362`.
- {meth}`scvi.module.VAE.marginal_ll` when `n_mc_samples_per_pass=1` {pr}`2362`.
- Enable option to drop_last minibatch during training by `datasplitter_kwargs={"drop_last": True}`
{pr}`2926`.
- Fix JAX to be deterministic on CUDA when seed is manually set {pr}`2923`.

Removed

- Remove {class}`scvi.autotune.ModelTuner`, {class}`scvi.autotune.TunerManager`, and
{class}`scvi.autotune.TuneAnalysis` in favor of new experimental functional API with
{func}`scvi.autotune.run_autotune` and {class}`scvi.autotune.AutotuneExperiment` {pr}`2561`.
- Remove `feed_labels` argument and corresponding code paths in {meth}`scvi.module.SCANVAE.loss`
{pr}`2644`.
- Remove {class}`scvi.train._callbacks.MetricsCallback` and argument `additional_val_metrics` in
{class}`scvi.train.Trainer` {pr}`2646`.