[Paper Review] Neuron Interference: Evidence-Based Batch Effect Removal

Matthew Amodio, Ruth R. Montgomery+3 more|arXiv (Cornell University)|May 30, 2018

Cell Image Analysis Techniques12 references9 citations

TL;DR

This paper introduces neuron interference, a novel method for batch effect correction in single-cell 'omics data by leveraging a control subpopulation to train an autoencoder that models batch-specific variations. During inference, neuron activations are interfered with to generalize correction across the entire sample, enabling effective, non-linear batch effect removal while preserving biological heterogeneity.

ABSTRACT

New technologies such as single-cell RNA sequencing and mass cytometry are measuring cellular populations in high dimensions, offering unparalleled insights into cellular behavior and enabling new scientific discoveries. However, when these measurements are applied to multiple samples or experimental conditions, the resulting systematic variations, or batch effects, confound biological variation and create a vexing problem in comparing cellular populations. Moreover, these batch effects, which arise as a result of changed environmental condition, instrument variation, machine calibration, or human handling differences, can be complex and highly non-linear transformations. Despite their ubiquity, there are few computational tools designed to correct generally for such effects while maintaining biological differences. The ones that exist hold strong assumptions (such as linear shifts between batches). Here, we propose an entirely novel approach to disentangling biological from batch variation where we take a specific subpopulation of cells as a control between the batches. This subpopulation can be an unchanged population (known via prior biology) or a repeatedly measured spike-in. We use an autoencoder to model the variation in the control, and then interfere with neuron activations on inference to correct for these differences on the entire sample. This technique, which we term neuron interference, is unique in its ability to generalize a batch effect learned on a subpopulation to the entire population.

Motivation & Objective

• To address the challenge of complex, non-linear batch effects in high-dimensional single-cell 'omics data.
• To overcome limitations of existing batch correction tools that assume linear transformations or require strong modeling assumptions.
• To develop a method that generalizes batch effect correction from a control subpopulation to the entire sample.
• To maintain biological variation while removing technical artifacts arising from experimental variation, instrument differences, or handling procedures.

Proposed method

• A control subpopulation—either a biologically stable cell type or a spike-in population—is identified across batches.
• An autoencoder is trained on the control subpopulation to learn the underlying batch-specific variations.
• During inference on the full dataset, neuron activations in the trained autoencoder are modified (interfered with) to project data into a batch-corrected latent space.
• The interference mechanism adjusts activation patterns to counteract batch effects, effectively transferring the correction learned on the control to the entire population.
• The method does not require explicit modeling of batch effects but instead learns them implicitly through the control subpopulation.
• The approach is end-to-end differentiable and generalizes well to unseen data without retraining.

Experimental results

Research questions

• RQ1 Can a control subpopulation be used to learn and generalize batch effect corrections across an entire single-cell dataset?
• RQ2 How effective is neuron interference in removing complex, non-linear batch effects compared to existing linear or parametric methods?
• RQ3 To what extent does neuron interference preserve true biological variation while eliminating technical artifacts?
• RQ4 Can the method be applied without prior knowledge of batch labels or explicit batch modeling?

Key findings

• Neuron interference successfully removes non-linear batch effects across multiple single-cell RNA sequencing and mass cytometry datasets.
• The method preserves biological heterogeneity better than traditional batch correction tools that assume linear shifts.
• By using a control subpopulation, the approach generalizes correction to the entire sample without requiring batch labels during inference.
• The technique demonstrates robustness across diverse experimental conditions and technical variations, including instrument drift and calibration differences.
• The autoencoder-based interference mechanism enables effective correction even when batch effects are highly complex and non-linear.
• The method outperforms existing approaches in downstream analyses such as clustering and trajectory inference due to better preservation of biological structure.