2026-04-14
Before I encountered batch effects in my own studies, I considered batch correction as just another preprocessing step—something you apply when data is measured in fragments: normalize, log-transform, run a batch correction method, move on.
That framing turns out to be wrong.
Batch effects are not just unwanted variation we failed to control. They are often unavoidable consequences of how studies are designed and executed.
They show up when you scale, when you deliberately sample diversity, and when you operate under constraints. In each case, the variation is not incidental; it is a byproduct of what you are trying to do.
Three projects. Three different problems.
Project 1: When batch and biology are the same thing.
In a COVID-19 metabolomics study, we had a clean case-control design until we examined how the samples were processed. All controls had been run in one batch. All cases in another.
When batch maps perfectly onto a condition, you cannot statistically separate them. Any correction that removes batch variation also removes the biological signal. The two are not just correlated; they are effectively the same variable.
Instead, we used a third batch with anchor samples (the same biological material run across batches) to estimate per-metabolite shifts and align the data to restore comparability. The anchors gave us something external to hold onto: a way to align batches without assuming that condition and batch are separable, because they weren’t.
The shift that mattered here: we weren’t removing batch effects. We were making the problem solvable.
This wasn’t just a technical artifact; it was a consequence of working under urgency, where perfect balance in study design isn’t always feasible.
Reference: M. Buyukozkan, S. Alvarez-Mulett, A. C. Racanelli, F. Schmidt, R. Batra, et al. “Integrative metabolomic and proteomic signatures define clinical outcomes in severe COVID-19” iScience, 2022, doi: 10.1016/j.isci.2022.104612
Project 2: When batch overlays biology but doesn't define it.
In our postmortem brain metabolomics work across diverse ethnoracial groups, samples were collected from multiple sites and processed under different conditions.
There were clear systematic differences across sites. Overall, the cohort structure wasn’t perfectly aligned with any single biological variable of interest. Batch was structured variation, real, but not inseparable from biology.
We used ComBat here. It works because it assumes that biological and technical variation can be decomposed. That assumption held here.
When batch is separable, you can model and subtract it. When it is entangled with your outcome, you can’t
The cohort structure we saw was not noise; it was the result of deliberately sampling across populations.
Reference: A. Schweickart*, R. Batra*, et al. “Sphingolipid and ceramide associations with tau vary across diverse ethnoracial groups in postmortem brain tissue” Under Review, 2025, doi: 10.1101/2025.11.04.25339489
Project 3: When everything is standardized, batch still happens.
Quantitative NMR-based metabolomics is about as standardized as metabolomics gets. Tightly controlled protocols, reference spectra, quantitative outputs.
At scale (500K UKBB samples), even on the same platform, different instruments, calibration drift, and throughput constraints introduce systematic variation. In spectroscopic data, this shows up as shifts and scaling differences that standard normalization doesn’t fully resolve.
The takeaway is simple, standardization reduces batch effects. It does not eliminate them. Moreover, this is what scaling looks like in practice: variation that emerges from the logistics of measurement itself.
Reference: N. Amin, J. Liu, W. Sproviero, M. Arnold, R. Batra, et al. “Interplay between age, APOE Ɛ4 and the metabolome in plasma and brain in Alzheimer’s disease” Translational Psychiatry, 2025, doi: 10.1038/s41398-025-03625-8
Types of batch effects
Across these cases, what changed wasn’t the method. It was recognizing that these were not nuisances, but different regimes created by the constraints of the study itself.
Confounded: batch is entangled with biology → not identifiable without external structure (anchors)
Separable: batch overlays biology → can be modeled and removed
Intrinsic (platform-level): variation emerges from measurement at scale → requires signal-level alignment
Batch correction is not just about removing unwanted variation. It is about preserving the variation that matters for your question. Batch correction is not a preprocessing step. It is a modeling decision.
Batch effects are not something you eliminate. They are something you work around, given the constraints of your study. And those constraints—scale, diversity, and urgency—are often exactly what make the study worth doing in the first place.
2026-03-24
Three preprocessing steps that solve different problems.
When I was a PhD student at the University of Heidelberg working on image-based siRNA knockdown screens, I remember being confused by preprocessing steps that kept appearing in analysis pipelines.
At the time, they felt interchangeable. All seemed to involve “adjusting the data” before analysis. It took a while before I understood that they addressed different sources of variation.
Years later, I still see the same confusion among students. And the reason is understandable: these steps often appear together in workflows but are rarely explained conceptually.
A simple way to clarify them is to think about the structure of the data.
Most datasets in biology and machine learning can be represented as a matrix:
Rows = samples
Columns = features
Different preprocessing steps operate on different parts of this matrix.
Cross-Sample Normalization - Make samples comparable
In many biological experiments, samples differ not only because of biology, but also because of technical variation.
For example:
sequencing depth in RNA-seq
total signal intensity in metabolomics
global brightness differences in imaging assays
These differences can make one sample appear globally “larger” than another.
Cross-sample normalization adjusts for these global shifts, making samples comparable.
Importantly, this step does not change the distribution of individual features.
It simply removes systematic differences between samples.
Memory aid:
Normalize → level the samples
Log Transformation - Adjust the shape of feature distributions
Many biological measurements span orders of magnitude and are strongly right-skewed.
A small number of large values can dominate the data.
Log transformation helps by:
compressing large values
reducing skew
stabilizing variance
turning multiplicative relationships into additive ones
For example:
1, 10, 100, 1000 → log10 → 0, 1, 2, 3
The data become easier to model and interpret.
This step does not make samples comparable.
It changes how values are distributed within each feature.
Memory aid:
Log → fix the shape
Feature Scaling (Standardization) - Make features comparable
Different features often have different ranges or variability.
Without adjustment, features with larger variance can dominate analyses such as:
PCA
clustering
regression
many machine-learning models
Standardization (z-scoring) rescales each feature: (x−μ)/σ
After scaling:
mean = 0
standard deviation = 1
This ensures features contribute on a comparable footing.
Memory aid:
Scaling → use the same ruler
Typical Order in Many Workflows
When all three steps are needed, the order often looks like this:
Cross-sample normalization – remove technical bias
Log transformation – stabilize distributions
Feature scaling – standardize feature variance
They correct different sources of variation:
sample-level differences
distributional geometry
feature-level comparability
Understanding which source of variation you are correcting is what makes preprocessing intentional.
Preprocessing is often presented as a checklist of steps.
In practice, it is about recognizing which type of variation is present in the data.
That realization took me some time as a student.
And judging by the questions I still hear from students today, it remains a useful distinction to make explicit.
2026-03-10
The real bottleneck in multimodal biology may be understanding what integrated models mean.
The promise of multi-omics
Multi-omics integration is often presented as the natural next step in systems biology. If biological systems operate across multiple molecular layers, genes, proteins, metabolites, lipids, phenotypes, then integrating these layers should reveal mechanisms that single-omic analyses might miss.
Over the past decade, numerous methods have been developed for this purpose: network-based integration, latent embedding models, and multimodal deep learning approaches. Much of the recent progress in the field has therefore focused on integrating heterogeneous datasets computationally.
In principle, integrating multiple modalities should yield deeper biological insight. In practice, however, most studies remain bi-modal rather than truly multimodal. And the limiting factor is rarely computational. The harder challenge often begins after the data have been combined: understanding what the integrated model is telling us and how it relates to the biological question at hand.
Different ways to integrate omics layers
Integration strategies can be organized around different conceptual anchors. Across several projects, we explored a range of such approaches.
Network-anchored integration
One early approach was network module detection, for example, using KeyPathwayMiner 4.0 (Alcaraz N*, Pauling J*, Batra R*, et al., 2014). In this framework, heterogeneous molecular entities can be represented within a biological network, while omics measurements are mapped as node attributes. The algorithm searches for subnetworks that exhibit consistent perturbation patterns across datasets.
This was among the first methods to enable analysis of multiple omics layers within a single network framework. But the algorithm does not generate biological meaning. It produces candidate subnetworks, and determining which modules are relevant requires contextual biological interpretation.
Phenotype-anchored integration
Another strategy is to anchor integration on phenotypic attributes.
In the Attribute Gene Expression Regularization (AuGER) framework (Batra*†, Stark*, Lauffer*, Jargosch* et al., 2021), clinical attributes serve as the organizing layer, linking gene expression profiles to patient phenotypes.
Instead of directly connecting omics datasets to each other, they are mapped through the phenotype layer: omics → phenotype → biological interpretation.
This approach enables the identification of gene signatures associated with specific clinical attributes, revealing both known and previously unexplored biological processes. The same framework can be extended to additional omics layers measured in the same individuals.
Yet, interpretation remains subjective. Once attribute-specific signatures are identified, investigators must still decide which biological programs warrant deeper investigation.
Cascade-based integration
In studies of acute respiratory distress syndrome (ARDS), we integrated metabolomics, lipidomics, and proteomics using two complementary approaches (Batra*, Whalen*, et al., 2022). One strategy attempted to reconstruct a putative molecular cascade linking metabolites, lipids, and proteins (Figure 3). A second strategy used partial correlation networks to identify molecular associations that differed between COVID-19-associated ARDS and bacterial ARDS (Figure 5).
These strategies illustrate two different integration philosophies:
putative cascade reconstruction
statistical dependency networks
Both produce interpretable structures, but neither eliminates the need for human judgement when translating patterns into biological hypotheses.
Biochemical pathway reconstruction
In an Alzheimer’s disease study (Batra*, Arnold* et al., 2022), metabolomic perturbations suggested a biochemical cascade. Proteomic data were then used to identify enzymes operating along that pathway (Figure 5).
This represents a classic metabolite-to-enzyme integration strategy: metabolic signals generate hypotheses that are anchored by enzymatic evidence.
Cross-omics co-expression networks
In another project examining immunometabolic stress during spaceflight (Houerbi et al., 2024), we constructed co-expression networks linking metabolomic and proteomic measurements (Figure 3).
Here, integration relied on correlated molecular programs across omics layers, revealing coordinated responses to the physiological stresses of space travel.
A pattern across studies
Looking across these studies, an interesting pattern emerges. Although the integration strategies differed across network modules, phenotype anchors, biochemical cascades, and cross-omics correlations, their structure was rarely determined solely by algorithms.
Instead, the choice of integration framework was shaped by a combination of practical and conceptual factors:
Which omics layers were available
Which biological questions emerged during analysis
Which representations remained interpretable
In many cases, the data helped shape the question, rather than the other way around. This may partly explain why most multi-omics studies remain bimodal even though integrating more modalities is technically feasible.
The interpretability bottleneck
As more omics layers are integrated, models increasingly rely on compressed representations of high-dimensional data. While these models can integrate heterogeneous datasets effectively, they may also obscure the reasoning chain required for biological interpretation.
Eventually, someone must still decide:
Which signals matter
Which pathways are plausible
Which findings justify experimental follow-up
Those decisions are rarely algorithmic. They depend on biological context, domain expertise, and sometimes subjective judgement.
In this sense, multi-omics integration is not only a computational problem.
It is also an interpretation problem.
In practice, interpretation often depends less on the integration algorithm itself and more on how clearly the analysis is anchored to the biological question being asked.
An open question
As multimodal datasets continue to grow, the field faces an interesting tension. Methods that integrate many modalities often produce compressed representations that are difficult to interpret. Methods that remain interpretable often integrate fewer modalities.
Where should that balance lie?
For those working on multi-omics integration: How do you decide where interpretation should anchor when integrating multiple molecular layers?
Questions shape integration
Integration methods are rarely developed in a vacuum. Most tools emerge from attempts to answer specific biological questions.
Network-based approaches often aim to identify coordinated perturbations within molecular pathways. Phenotype-anchored frameworks are designed to connect molecular variation to clinical or phenotypic attributes. Cascade reconstruction methods are typically motivated by the desire to trace potential biochemical relationships across molecular layers.
In other words, the structure of an integration method usually reflects the type of question it was originally designed to address.
Seen from this perspective, the interpretability challenge is not only methodological. It is also conceptual. When multiple modalities are integrated, biological insight ultimately depends on how clearly the analysis is anchored to the question being asked.
Integration methods combine data across modalities. Biological insight emerges only when those integrations are guided by a clear question.
2026-03-03
You can run a single-cell RNA-seq pipeline in 20 minutes. The tutorials are excellent. The pipeline is mature.
Agentic AI can now execute the entire analysis with a click: load data, perform quality control, normalize, cluster, generate visualizations, and even suggest annotations.
Execution is no longer the bottleneck.
So what exactly remains the responsibility of the scientist?
Recently, while walking through a single-cell pipeline with students, I paused at each step and asked:
What is this function actually doing?
What do these axes represent?
What assumption are we introducing?
I’ve been asking these questions for years across projects. Single-cell analysis simply makes the modeling layers unusually visible.
Here is what that discussion clarified.
Quality Control: Defining What Counts as Biology
Quality control is often described as “filtering bad cells.” That framing is misleading.
When we filter based on:
Number of detected genes
Total Unique Molecular Identifier (UMI) counts
Percentage of mitochondrial transcripts
we are not cleaning data.
We are defining what counts as a biologically valid observation.
High mitochondrial RNA often reflects stressed or dying cells. Extremely high total UMI counts may indicate doublets. Low feature counts may represent empty droplets.
Each threshold encodes a judgment about which observations represent meaningful biological units.
Change the threshold, and clusters may shift.
If clusters shift, the biological narrative may shift.
Quality control is not technical housekeeping.
It is biological boundary-setting.
Normalization: Deciding What Is Noise
Raw counts are not directly comparable across cells because sequencing depth varies.
Normalization corrects for library size, allowing expression levels to be compared across cells.
But this step assumes sequencing depth is technical noise rather than a biological signal.
That assumption often holds when depth reflects stochastic capture efficiency.
It fails when the total RNA content is biologically meaningful in itself.
For example:
Activated immune cells can globally increase transcription.
Large neurons may contain more RNA than small glial cells.
Cells in different cell cycle phases vary in total RNA abundance.
In these cases, “correcting” for sequencing depth may partially remove the real biological signal.
Interpretation begins by recognizing what you have chosen to treat as noise and whether that choice is defensible in your system.
Highly Variable Genes: Choosing the Signal
Not all genes are informative for distinguishing cell states.
Housekeeping genes are expressed broadly and contribute little to separation.
By selecting highly variable genes, those with "greater-than-expected variance relative to their mean," we prioritize structured heterogeneity.
In effect, we are declaring: variation defines identity.
This is a modeling choice. It elevates certain biological dimensions while down-weighting others.
Without this step, technical noise can dominate dimensional reduction.
With it, we amplify structured differences.
Scaling and PCA: Constructing Axes of Variation
Scaling (typically Z-scoring) ensures that genes with higher absolute expression do not dominate variance-based methods.
PCA then identifies orthogonal axes of maximal variance, linear combinations of genes weighted by their contribution.
A principal component is not a cell type.
It is an axis of structured variation.
The loadings indicate how strongly each gene contributes to that axis.
If a principal component separates T cells from B cells, it does so because the weighted genes encode that biology, not because PCA “discovered” discrete categories.
Interpretation requires examining those weights and relating them to known biological programs.
UMAP and Clustering: Geometry Before Biology
UMAP is a nonlinear embedding.
Its axes do not represent gene expression.
They represent distances between cells in compressed high-dimensional space.
Closer cells are more transcriptionally similar.
The axes themselves carry no intrinsic biological meaning.
Clustering then identifies communities in a similarity graph.
Clusters may represent:
Cell types
Activation states
Cell cycle
Stress responses
Technical artifacts
Geometry comes first.
Biology comes later.
Markers: From Structure to Meaning
Clusters are numeric labels until differential expression is performed.
Marker genes, those "enriched in one cluster relative to others," allow biological annotation.
Without markers, clustering is a structure without interpretation.
Markers translate geometry into biological hypotheses.
Why This Matters Now
In the age of agentic AI, all of this can be automated.
The pipeline runs.
The plots look polished.
The clusters appear coherent.
Annotations may even be suggested.
But AI does not understand your experimental design.
It does not know whether a cluster reflects lineage, a stress response, the cell cycle, or an artifact.
It does not reason about whether a mitochondrial threshold removes dying cells or eliminates a biologically meaningful subpopulation.
It extracts patterns.
It does not exercise biological judgment.
Automation shifts the locus of expertise upward.
The real skill is no longer executing the pipeline.
It is knowing:
What you changed
Why you changed it
What assumption you introduced
Where that assumption might break
How it shapes interpretation
Single-cell analysis makes these modeling layers visible.
But this principle extends to GWAS, imaging, bulk RNA-seq, and most machine learning workflows.
Every analysis is a chain of assumptions layered onto data.
AI can execute that chain.
It cannot yet replace disciplined, context-aware interpretation.
And increasingly, that human element, critical thinking grounded in biological reasoning, is the differentiator.