SENADVAE

SENA Discrepancy VAE implementation for scvi-tools

Details on the architecture and rationale can be found in the original paper: arXiv:2506.12439.

---

Overview

SENADVAE is a customized implementation of the SENA discrepancy variational autoencoder within the scvi-tools framework.
It is designed to handle the specific training dynamics and data structures required for perturb-seq transcriptomic experiments.

---

Repository structure

• module.py
  Defines the neural network architecture adapted to meet scvi-tools requirements.
  Includes generative and inference functions.

• training_plan.py
  Implements a custom training plan to handle SENA’s complex optimization scheme:

  • Dynamic scheduling of loss parameters.
  • Separate reconstruction losses for control and perturbed transcriptomic profiles.

• dataloader.py
  Provides utilities for pre-processing perturb-seq .h5ad files.
  Ensures input data is structured according to SENADVAE’s requirements and scvi expectations.

• model.py
  Contains the core SENADVAE class, responsible for:

  • Initializing model components.
  • Wrapping the architecture into a scvi-tools compatible model.
  • Defining custom inference methods for pertubation prediction.
  • Implementing metric generation utilities (e.g., reconstruction accuracy, discrepancy measures).
  • Incudes key internal layer extraction.

---