diff --git a/snapatac2-python/src/embedding.rs b/snapatac2-python/src/embedding.rs
index fe147d3d..5ac2cc40 100644
--- a/snapatac2-python/src/embedding.rs
+++ b/snapatac2-python/src/embedding.rs
@@ -1,7 +1,9 @@
 use crate::utils::AnnDataLike;
 
 use anndata::{
-    data::{array::utils::to_csr_data, SelectInfoElem, SelectInfoElemBounds},
+    data::{
+        array::utils::to_csr_data, ArrayConvert, DynCsrMatrix, SelectInfoElem, SelectInfoElemBounds,
+    },
     AnnDataOp, ArrayData, ArrayElemOp, Backend, Selectable,
 };
 use anndata_hdf5::H5;
@@ -32,7 +34,11 @@ pub(crate) fn spectral_embedding<'py>(
     macro_rules! run {
         ($data:expr) => {{
             let slice = pyanndata::data::to_select_elem(selected_features, $data.n_vars())?;
-            let mut mat: CsrMatrix<f64> = $data.x().slice_axis(1, slice)?.unwrap();
+            let mut mat: CsrMatrix<f64> = $data
+                .x()
+                .slice_axis::<DynCsrMatrix, _>(1, slice)?
+                .unwrap()
+                .try_convert()?;
             if let Some(weights) = feature_weights {
                 normalize(&mut mat, &weights);
             } else {
@@ -71,12 +77,10 @@ pub(crate) fn spectral_embedding_nystrom<'py>(
             let weights = if let Some(weights) = feature_weights {
                 weights
             } else {
-                idf_from_chunks(
-                    $data
-                        .x()
-                        .iter(5000)
-                        .map(|x: (CsrMatrix<f64>, _, _)| x.0.select_axis(1, &selected_features)),
-                )
+                idf_from_chunks($data.x().iter(5000).map(|x: (DynCsrMatrix, _, _)| {
+                    let mat: CsrMatrix<f64> = x.0.try_convert().unwrap();
+                    mat.select_axis(1, &selected_features)
+                }))
             };
 
             let n_obs = $data.n_obs();
@@ -91,8 +95,9 @@ pub(crate) fn spectral_embedding_nystrom<'py>(
             let selected_samples = SelectInfoElem::from(idx);
             let mut seed_mat: CsrMatrix<f64> = $data
                 .x()
-                .slice(&[selected_samples, selected_features.clone()])?
-                .unwrap();
+                .slice::<DynCsrMatrix, _>(&[selected_samples, selected_features.clone()])?
+                .unwrap()
+                .try_convert()?;
 
             // feature weighting and L2 norm normalization.
             normalize(&mut seed_mat, &weights);
@@ -106,8 +111,9 @@ pub(crate) fn spectral_embedding_nystrom<'py>(
                 &v,
                 &mut u,
                 &d,
-                $data.x().iter(chunk_size).map(|x: (CsrMatrix<f64>, _, _)| {
-                    let mut mat = x.0.select_axis(1, &selected_features);
+                $data.x().iter(chunk_size).map(|x: (DynCsrMatrix, _, _)| {
+                    let mut mat: CsrMatrix<f64> = x.0.try_convert().unwrap();
+                    mat = mat.select_axis(1, &selected_features);
                     normalize(&mut mat, &weights);
                     mat
                 }),
@@ -175,7 +181,8 @@ fn spectral_mf(
             random_state,
         );
         let result = fun.call1(py, args)?;
-        let (evals, evecs): (PyReadonlyArray1<'_, f64>, PyReadonlyArray2<'_, f64>) = result.extract(py)?;
+        let (evals, evecs): (PyReadonlyArray1<'_, f64>, PyReadonlyArray2<'_, f64>) =
+            result.extract(py)?;
 
         anyhow::Ok((evals.to_owned_array(), evecs.to_owned_array()))
     })?;
@@ -396,8 +403,13 @@ pub(crate) fn multi_spectral_embedding<'py>(
                 ($data:expr) => {{
                     let slice = pyanndata::data::to_select_elem(&s, $data.n_vars())
                         .expect("Invalid feature selection");
-                    let mut mat: CsrMatrix<f64> =
-                        $data.x().slice_axis(1, slice).unwrap().expect("X is None");
+                    let mut mat: CsrMatrix<f64> = $data
+                        .x()
+                        .slice_axis::<DynCsrMatrix, _>(1, slice)
+                        .unwrap()
+                        .expect("X is None")
+                        .try_convert()
+                        .unwrap();
                     let feature_weights = idf(&mat);
 
                     // feature weighting and L2 norm normalization.