Technical Summary: Reference-guided pseudotime inference across species and biological contexts

Problem Statement

Single-cell RNA sequencing (scRNA-seq) captures cellular heterogeneity, yet cells collected at the same chronological age often exhibit significant variation in biological age due to differences in differentiation timing, maturation speed, and degeneration rates. Existing pseudotime inference methods face critical limitations in addressing this heterogeneity:

1. Unsupervised methods (e.g., Monocle, Palantir, Diffusion Pseudotime) rely on cell neighborhood graphs but may not optimally reflect distance along specific axes of interest (e.g., disease progression or specific developmental trajectories) and lack inherent directionality.
2. Supervised methods (e.g., Psupertime, Sceptic) require accurate time-series labels within the study of interest, which are often unavailable in human studies (due to heterogeneity or ethical constraints) or in vitro experiments (which often cover only a few timepoints).
3. Cross-context limitations: Current approaches struggle to anchor in vitro models to in vivo timelines or transfer temporal information across species, particularly when batch effects and context-specific transcriptional variations are present.

Consequently, there is a lack of robust frameworks to order cells along desired temporal axes in biological contexts where reliable time labels are absent, such as human embryonic development, disease progression in patients, or in vitro organoid differentiation.

Methodology: The Cavebear Framework

The authors propose Cavebear, a machine learning framework designed to infer pseudotime in a query species or condition by leveraging high-quality, time-series scRNA-seq profiles from a reference species or condition. The framework operates in three distinct steps:

Step 1: Unsupervised Cross-Species Alignment

Cavebear utilizes a conditional Variational Autoencoder (cVAE), adapted from their previous method Icebear, to learn species-agnostic cell embeddings.

• Input: Raw gene expression count vectors concatenated with one-hot encoded batch and context labels.
• Mechanism: The encoder maps inputs to a low-dimensional latent space, while the decoder reconstructs gene expression conditioned on context and batch. The model minimizes reconstruction loss and KL divergence to capture biological variation while removing batch and context-specific differences.
• Adversarial Alignment: For scenarios with substantial biological differences (e.g., in vivo vs. in vitro), an optional adversarial alignment step is employed. A discriminator attempts to predict the context label from the latent embedding, while the cVAE is trained to minimize this prediction accuracy, further reducing context-specific information in the latent space.
• Selection: Models are selected based on the highest cross-context Local Inverse Simpson's Index (LISI) score in the latent space.

Step 2: Supervised Pseudotime Prediction

Using the species-agnostic embeddings generated in Step 1, a temporal prediction model is trained.

• Architecture: A fully connected multilayer perceptron (MLP) takes the latent embedding as input and outputs a scalar representing predicted pseudotime.
• Training: The model is trained on the reference dataset using known time labels, optimizing for Mean Squared Error (MSE) loss.
• Hyperparameters: The model performs a grid search over learning rates, hidden layer counts, and dimensions to select the best model based on validation loss.

Step 3: Query Prediction

The trained MLP from Step 2 is applied directly to the latent embeddings of the query dataset (which lacks time labels) to generate per-cell pseudotime estimates. No additional training or fine-tuning is performed on the query data.

Key Applications and Results

The paper validates Cavebear across three diverse biological applications, demonstrating its utility in cross-species, cross-system, and cross-disease contexts.

1. Cross-Species Developmental Inference (Zebrafish vs. Mouse)

• Setup: Mouse embryonic brain development (E8.0–E18.75) served as the reference for predicting pseudotime in zebrafish embryos (8h–96h post-fertilization).
• Performance: Cavebear outperformed existing unsupervised methods (Monocle 2, Palantir, DPT) and the supervised competitor TemporalVAE in 12 out of 14 tested cell types.
• Robustness: The method remained effective even when the reference time-series was subsampled to as few as 2 timepoints, though performance improved with higher timepoint density. It successfully handled cell types absent in the reference (e.g., hatching gland) by maintaining distinct identities.
• Heterogeneity: Cavebear successfully resolved developmental heterogeneity within single timepoints, showing positive Spearman correlation with an "upper baseline" (a theoretical limit trained on query time labels), whereas unsupervised methods failed to consistently order cells without directionality.

2. Anchoring In Vitro to In Vivo Timelines

• Mouse Embryoid Bodies (EBs): Cavebear mapped in vitro mouse EB differentiation (days 0–11) onto an in vivo mouse gastrulation and organogenesis timeline (E3.5–E13.5). Predicted pseudotime increased monotonically with culture days, validating the alignment.
• Human Cortical Organoids: Using a mouse reference, Cavebear inferred in vivo timing for human organoids at Day 14 and Day 84. The model correctly assigned earlier pseudotimes to Day 14 and later to Day 84. Notably, it identified a specific donor (429591) with a delayed developmental trajectory, consistent with external phenotypic data indicating neurodevelopmental delay.

3. Disease Progression in Colorectal Cancer (CRC)

• Setup: A human CRC dataset (matched tumor and normal tissues from 10 patients) was analyzed using a mouse CRC model (normal to malignant stages) as a reference.
• Findings: Cavebear inferred higher pseudotimes in tumor samples compared to matched adjacent normals for solid tissue cell types (endothelial, epithelial, fibroblast).
• Granularity: The method detected malignancy-associated pseudotimes in blood-associated cell types within "normal" adjacent tissues, suggesting persistent systemic changes.
• Cross-Disease Generalization: When the mouse CRC reference was replaced with a Chronic Myeloid Leukemia (CML) mouse model (a different cancer type), Cavebear still successfully identified significant progression signals in solid tissue cell types of the human CRC patients, demonstrating the transferability of progression signals across distinct disease models.

Significance and Claims

The paper positions Cavebear as a significant advancement for studying biological variation in contexts where temporal annotations were previously unattainable. Its primary contributions include:

1. Transfer Learning for Pseudotime: It establishes a framework for transferring temporal information across species and biological conditions, overcoming the limitation that accurate time labels are often unavailable in human or in vitro studies.
2. Benchmarking In Vitro Models: It provides a quantitative method to anchor in vitro experiments (organoids, embryoid bodies) to in vivo developmental timelines, allowing for the assessment of maturation rates and the identification of developmental delays or abnormalities.
3. Cellular-Level Disease Characterization: It enables the reconstruction of disease progression trajectories in patient samples by leveraging step-wise progression data from model organisms, revealing heterogeneous cellular-level patterns that are invisible to static analysis.
4. Robustness to Reference Density: The study demonstrates that while dense time-series references yield optimal results, Cavebear remains effective even with sparse reference data, outperforming traditional unsupervised methods.

The authors conclude that Cavebear serves as a flexible tool for systematic investigation of biological variation, particularly in scenarios involving human development, disease progression, and the interpretation of in vitro models, by leveraging the rich temporal atlases available in model organisms.