In biological research, the "style" of an experiment often screams louder than the biology itself. When a researcher treats cells with a drug in one laboratory, the resulting images or genetic profiles will look fundamentally different from the same experiment performed in another lab. These discrepancies—known as batch effects—arise from subtle differences in pipetting, temperature, staining protocols, or sequencing hardware.
To a standard machine learning model, these technical artifacts are often more salient than the actual biological signal. If a model is trained to distinguish between "healthy" and "diseased" cells, it might inadvertently learn to distinguish between "Lab A" and "Lab B."
A growing body of research is solving this by treating experimental noise not as a nuisance to be removed, but as a source of supervision. By leveraging the inherent structure of biological experiments, researchers are training models to "see through" technical artifacts to find the invariant biological truth.
Replicates as Natural Data Augmentations
In computer vision, we often create "invariance" by showing a model different versions of the same image—flipped, rotated, or recolored—and telling the model they represent the same object. In biology, we have a more powerful version of this: the biological replicate.
A replicate is the same experiment performed multiple times, often across different physical plates or days. Juan C. Caicedo and colleagues at the Broad Institute pioneered the use of these replicates as "weak labels." By telling a model that two different samples share the same treatment, the model is forced to extract features that persist across different technical environments while ignoring site-specific artifacts.
This idea has been refined into sophisticated contrastive learning frameworks. For example, Set-DINO, developed by Heming Yao and the team at Genentech, uses replicates from distinct experimental batches as "natural" data augmentations. Similarly, the CWA-MSN framework by Yu-Hsuan Liao aligns images of the same treatment from different physical wells.
The intuition is simple: if the model is forced to map two visually different images (due to batch noise) to the same point in a latent space, it must discard the noise and keep the signal. This approach has proven remarkably efficient; Liao demonstrated that this cross-well alignment allows models to achieve state-of-the-art performance with 90% less data.
Adversarial Decoupling of Technical Noise
While contrastive learning "pulls" similar biological signals together, adversarial training "pushes" technical noise out. This approach uses a "minimax" game: one part of the network (the encoder) tries to learn useful biological features, while another part (the adversary) tries to guess which experimental batch the data came from.
If the adversary can easily guess the batch, the encoder is still leaking technical noise. The encoder is therefore optimized to make the adversary fail. Carlos Arteta and Andrew Zisserman applied this to high-throughput imaging in the HTX tool, using an adversarial branch to filter out technical noise.
This "information factorization" has been extended to genomic data as well. Lily Monnier developed Adversarial Information Factorization (AIF) to strip batch effects from single-cell RNA sequencing data. By penalizing the inclusion of batch-identifying information, these models create "neural fingerprints" that represent the intrinsic biological state of a cell, regardless of which animal or lab produced the data Wei Wu.
Self-Supervised Biological Priors
Sometimes, we don't even need explicit replicate labels to learn invariance. We can use the physical or biological structure of the data itself.
Alex Xijie Lu and Alan M. Moses developed paired cell inpainting, which asks a model to predict the protein distribution in one cell based on the appearance of its neighbor. Because neighboring cells in a dish share the same environment but have different individual shapes, the model learns to isolate underlying biological principles from the incidental physical distortions of a single cell.
In the realm of genomics, the CLEAR framework by Wenkai Han uses synthetic distortions—like random gene masking and "dropout" events—to simulate the noise inherent in sequencing. By training the model to recognize the identity of a cell across these distortions, the representation space naturally aligns data from different experimental batches, prioritizing the shared biological identity over the technical "style" of the experiment.
Toward Biological Foundation Models
The ultimate goal of this research direction is the creation of Single-Cell Foundation Models (scFMs). As Seungbyn Baek and Insuk Lee argue, by exposing models to massive, heterogeneous datasets from thousands of different studies, the models eventually learn to distinguish the universal "language" of biology from the cacophony of laboratory noise.
Models like RedDino, developed by Andrea Loddo and Carsten Marr, demonstrate this potential. By training on millions of red blood cell images across multiple sites, the model becomes robust to varying staining protocols, allowing it to cluster cells based on their actual health status rather than the specific lab where the blood was drawn.
By turning the "problem" of batch effects into a "solution" for supervision, these researchers are building a future where AI can integrate the world's biological data into a single, unified map of life.