aeacus 0.6.0

mixture of experts single cell malignancy classifier

pip install aeacus

from source:

git clone https://github.com/pandey-ps/aeacus.git
cd aeacus
pip install -e .

Usage

from aeacus import Profiler

result = Profiler(test_input="query.h5ad").load().profile()
result.obs["malignancy_call"].value_counts()

Input

Format	Notes
.h5ad	genes in var_names, cells in obs_names
.txt / .tsv	rows = genes, columns = cells
.csv	rows = genes, columns = cells
AnnData object	passed directly

Parameters

Parameter	Default	Description
test_input	required	path to data or AnnData object
pretrain_dir	auto	path to folder with moe.pt, geneorder.tsv, train_mean.npy, train_std.npy, config.json.
norm_type	False	normalization to apply before inference (see below)
use_raw	False	if True, reads from adata.raw.X instead of adata.X
batch_size	8192	cells per batch, lower if out of memory
device	auto	"cuda" or "cpu"

norm_type

model was trained on CPM + log1p normalized data, choose norm_type based on your input:

Input	norm_type
raw UMI counts (10x, etc.)	"cpm_log1p" (or True)
CPM + log1p normalized	False (default) or "already_normalized"
TPM data (smart-seq, etc.)	"tpm_log1p"

Examples:

# raw counts - normalize 
Profiler(test_input="raw_counts.h5ad", norm_type="cpm_log1p")

# normalized - skip
Profiler(test_input="normalized.h5ad", norm_type=False)

# TPM - log1p
Profiler(test_input="tpm_data.h5ad", norm_type="tpm_log1p")

use_raw

use_raw chooses which data slot to read from:

Input	use_raw	norm_type
raw counts in .X, no .raw	False	"cpm_log1p"
raw counts in .raw, normalized in .X	True	"cpm_log1p"
normalized in .X	False	False

Example - AnnData with raw counts stored in .raw:

Profiler(
    test_input="adata.h5ad",
    use_raw=True,          # read from adata.raw.X
    norm_type="cpm_log1p", # then normalize
)

Inference

result = Profiler(test_input="data.h5ad", norm_type="cpm_log1p").load().profile()

Output

all predictions are added to result.obs:

Column	Description
malignancy_call	"Malignant" or "Normal"
malignancy_score	probability per cell
normal_expert_weight	weight assigned to the normal expert
malignant_expert_weight	weight assigned to the malignant expert

result.obs[["malignancy_call", "malignancy_score"]].head()

Note

• missing genes are filled with zeros after aligning to geneorder.tsv; warning showed if >20% of model genes are missing.