aeacus 0.1.0

mixture of experts single cell malignancy classifier

The aeacus/ folder contains inference code. The model/ folder contains the trained weights and normalization artifacts.

Install

From the repository root:

python -m venv venv
source venv/bin/activate
pip install -e .

Input Format

Input can be:

• .h5ad
• .txt / .tsv
• .csv
• an AnnData object

For text files, rows should be genes and columns should be cells. For .h5ad, genes should be in adata.var_names and cells in adata.obs_names.

Normalization

To use the same expression scale expected by the model:

• Already normalized: norm_type="False"
• Raw UMI counts: norm_type="cpm_log1p"
• TPM values: norm_type="tpm_log1p"

Do not z-score on new data. The model internally applies the training-set mean and standard deviation

Inference

For raw UMI count data:

from aeacus import Profiler

profiler = Profiler(
    test_input="query.h5ad",
    pretrain_dir="model",
    norm_type="cpm_log1p",
)

result_adata = profiler.load().profile()

For TPM data:

from aeacus import Profiler

result_adata = (
    Profiler(
        test_input="query.h5ad",
        pretrain_dir="model",
        norm_type="tpm_log1p",
    )
    .load()
    .profile()
)

For data already normalized:

from aeacus import Profiler

result_adata = (
    Profiler(
        test_input="query.h5ad",
        pretrain_dir="model",
        norm_type="False",
    )
    .load()
    .profile()
)

Output

Predictions are added to result_adata.obs:

malignancy_call      Normal or Malignant
malignancy_score     malignancy probability-like score
primary_expert       expert with the highest gate weight
primary_expert_label Normal or Malignant label for primary_expert
gate_entropy         uncertainty/spread of gate weights
normal_expert_weight gate weight assigned to the normal expert
malignant_expert_weight gate weight assigned to the malignant expert
expert_weight_0      same value as normal_expert_weight
expert_weight_1      same value as malignant_expert_weight
normal_expert_logit raw logit from the normal expert
malignant_expert_logit raw logit from the malignant expert

View results:

result_adata.obs["malignancy_call"].head()

View all other aeacus output columns:

prediction_cols = [
    "malignancy_call",
    "malignancy_score",
    "primary_expert",
    "primary_expert_label",
    "gate_entropy",
    "normal_expert_weight",
    "malignant_expert_weight",
    "expert_weight_0",
    "expert_weight_1",
    "normal_expert_logit",
    "malignant_expert_logit",
]

result_adata.obs[prediction_cols].head()

Note

• geneorder.tsv controls the model gene order. Missing genes are filled with zero after alignment.