🧬 TEMPO: Temporal Expression Model for Profiling and Optimization

Deep Learning Framework for Optimal Time-Point Selection and Semi-Profiling in Time-Series Single-Cell Studies

Fig. 1 — Overview of the TEMPO framework

📘 Overview

TEMPO (Temporal Expression Model for Profiling and Optimization) is a deep learning framework for optimal time-point selection and semi-profiling of time-series single-cell datasets.
It builds upon and generalizes the earlier DTPSP framework with:

• A modular encoder–GCN–regressor neural architecture
• A comprehensive beam-search strategy for optimal time-point selection
• A clean Python API (tempo_sc) packaged for reproducibility
• Support for single-cell profiling and inference (semi-profiling)

TEMPO addresses a central challenge in time-series biology:

How can we choose the most informative time points to collect single-cell data, while accurately inferring the missing ones?

---

🔍 Method Summary

a. Reference time-point selection

Given bulk or pseudobulk gene expression across T time points:

1. A beam search explores candidate subsets of time points ( S )
2. For each subset, a deep model predicts non-selected time points
3. Input features include:
   • Masked gene expression
   • Autoencoder latent embeddings
   • Graph Convolutional Network (GCN) neighborhood embeddings
   • Sinusoidal positional embeddings (PE) for time
4. The best subset ( S_B ) is selected according to validation MAE/R².

---

b. Single-cell profiling at selected time points

Cells are sequenced only at ( S_B ), reducing cost while preserving key biological dynamics.

Standard single-cell QC and processing (Scanpy) follow.

---

c. Single-cell inference at unmeasured time points

A generative model (e.g., VAE-GAN) learns from real scRNA-seq at ( S_B ) and bulk trajectories, enabling semi-profiling:

• Single-cell matrices are inferred for unmeasured time points
• A complete temporal single-cell atlas is produced

---

d. Downstream analysis

With full single-cell data available across all times, users can perform:

• Temporal trajectory visualization
• Differential expression
• Pseudotime inference
• Pathway enrichment
• Temporal deconvolution
• Multi-omics integration

---

🎯 Key Features

1. Optimal time-point selection

Maximizes information captured across a temporal study while minimizing profiling cost.

2. Improved neural architecture

Combines:

• Masked autoencoder
• GCN layer for gene topology
• Deep regressor
• Joint fine-tuning
• Positional embeddings

3. Scalable

Works for:

• Bulk RNA-seq
• Pseudobulk (single-cell aggregated)
• Microarray or temporal proteomics

4. Semi-profiling of missing time points

Generates realistic single-cell matrices for unmeasured time points.

5. Fully packaged and reproducible

Installable with pip directly from PyPI.

---

📦 Installation

Install the latest version

pip install tempo-sc