Many researchers showed that it could become an important player in spatial transcriptomics as well. However, it seemed that Illumina never wanted to jump on this opportunity.

Until now.

In this blog post, I break down everything we know so far about the upcoming Illumina spatial transcriptomics technology:

• How Illumina’s technology, originally built for DNA sequencing, can be cleverly adapted for spatial transcriptomics.

• A detailed comparison between Illumina’s upcoming product and Visium HD, including a prediction on the actual cost.

• The practical implications of Illumina’s promises for research.

The Illumina sequencer

Sequencers determine which DNA or RNA molecules are in a biological sample. These samples often contain a mixture of cells from tissues or fluids, and sequencing helps identify which genes are active, detect mutations, or track pathogens like viruses.

They read the nucleotides (A, C, G, T) of a molecule one by one, but this operation is run for all molecules in parallel.

Sequencing

Imagine you have a molecule with the sequence GATTACA. An Illumina sequencer can read it because it has nucleotides floating around that are bound to a specific fluorescent color. As an example, A might be blue, C=red, G=green, and T=yellow. Then, these nucleotides would sequentially bind to their complementary ones in the molecule.

The first nucleotide is a G, so its complementary C would bind, emitting red light. The second is A, so a T would bind, emitting yellow, and so on. By recording the sequence of colors, the machine can deduce the sequence of nucleotides.

However, there may be billions of DNA molecules in a sample. So how do you read them all at once?

Patterned flow cells

At the heart of Illumina’s approach is its patterned flow cell technology. Think of these flow cells as a grid of microscopic wells, spaced just 0.6 micrometers apart. As a reference, a human cell has a size of 5–20 micrometers.

DNA is extracted from the sample and spread over the flow cell. When a DNA molecule ends up in a well, it starts replicating, forming a cluster of identical copies. The replication happens so fast that there is not time for another molecule to enter the well, so every well contains copies of only one molecule.

And since each cluster consists of many identical DNA molecules, the fluorescence during sequencing is much stronger and easier to detect. This makes the signal way stronger and easier to spot.

Repurposing the grid for Spatial Transcriptomics

A grid of spots, where every spot has a huge amount of the same barcode. Does it remind you of another technology?

This core Illumina technology can be adapted, almost perfectly, for spatial transcriptomics. In 2024, researchers from Max Delbrück Center published Open-ST, where they hacked the flow cell into a spatial transcriptomics platform.

Here’s how:

1. Artificial barcodes: Instead of sequencing sample DNA, you generate synthetic barcoded DNA molecules that land in the wells of the flow cell.

2. Barcode mapping: Run a preliminary sequencing to map where each barcode landed.

3. Tissue transfer: A thin tissue section is placed over the flow cell. RNA from the tissue binds to the barcodes below.

4. Reverse transcription and sequencing: The RNA is converted into DNA linked to spatial barcodes, then sequenced.

Spatially resolved data: Combining barcode identity with pre-mapped positions creates a high-resolution gene expression map. Other similar examples include Pixel-seq and Seq-scope.

How does it compare to other technologies?

Although Illumina sequencers have been repurposed for spatial transcriptomics since 2022, Illumina’s official product isn’t launching until 2026. Why such a delay?

My hypothesis is that Illumina is not just repurposing flow cells but leveraging its expertise to develop a product more akin to Visium HD. In fact, Illumina’s product doesn’t have gaps between spots, just like Visium HD. This was one of the reasons for 10x Genomics to sue Illumina for patent infringement.

So let’s compare them.

Sample area

Illumina uses a 50×15 mm (7.5 cm²) slide versus a slide of two 6.5×6.5 mm capture areas in Visium HD (0.4 cm² total).

While large samples are appealing for studying tissue heterogeneity, excessively large sample areas may not always be practical. Researchers will have to fit multiple samples into the same slide, at the risk of tissue damage or lower quality of the results.

Resolution and sensitivity

Illumina claims 1 μm resolution, twice as sharp as Visium HD’s 2 μm

Once you go below single-cell resolution, a 1 μm difference doesn’t make a big difference. That’s because when RNA molecules move vertically from the tissue to the barcoded spots, they also tend to move sideways and can end up 10-20 micrometers away from where they started (a process called lateral diffusion).

However, the real gain is barcode density: Illumina increased it by 4x. That boosts sensitivity, given that current Visium HD experiments give an average of 200-400 transcripts per cell, compared to the 4,000-10,000 in a scRNA-seq experiment.

Indeed, Illumina claims to detect 2× more genes than competitors.

Cost and sensitivity

Here’s where things get interesting, and where Illumina’s marketing claims need a reality check.

Illumina also claims to lower the cost per tissue area by 4x, compared to other sequencing-based spatial transcriptomics technologies.

But, as Prof. Krisztian Homicsko noted, it’s unclear whether sequencing costs are included in the comparison. My belief is that the estimates do not include the cost of sequencing, but I would be happy to be contradicted.

Illumina people are reading this post, you know how to find me :)

Furthermore, Illumina’s product is 4x cheaper if we manage to cover the whole slide. But, as I mentioned before, you would need amazing Tetris skills to fit 5-10 tissue samples in a single slide.

To better understand the implications, I did some back-of-the-envelope estimations of the cost of an Illumina experiment with respect to the size of the tissue sample, taking an example cost for Visium HD from the Emory University Genomics Core facility website.

DISCLAIMER: All these results are estimates, as the pricing has not been released by Illumina yet.

For a small sample with an area of 0.8 cm² (the equivalent of 2 Visium HD capture areas):

• Visium HD: $9.5k ($6.8k slide + $2.7k sequencing)

• Illumina: $26k ($15.3k slide + $10.7k sequencing)

Illumina is 2.7x more expensive than Visium HD.

For a big sample with an area of 7.2 cm² (18 Visium HD capture areas, which is almost a full Illumina capture area):

• Visium HD: $85.1k ($61.2k slide + $23.9k sequencing)

• Illumina: $110.8k ($15.3k slide + $95.5k sequencing)

Illumina is still more expensive (1.3x)!

However, most of the cost goes to sequencing rather than the slide cost (86% for Illumina vs 28% for Visium HD), and more sequencing means higher sensitivity.

What this means for research

Thanks to their experience in designing flow cells, Illumina can pack in way more barcodes than the competition. This translates to better sensitivity and more genes detected.

The cost story is more nuanced than Illumina’s marketing suggests.

• For small samples, stick with Visium HD, and if you want to have high sensitivity, consider running scRNA-seq on an adjacent slide. The cost will still be lower than running Illumina.

• For big samples, or if you can somehow fit many samples into one slide, Illumina could be the better deal. Also because, in the case of Visium HD, you will have to pay for the people and time to run nine experiments.

That “4x cost reduction” only works if you can fill that entire large slide, which is easier said than done.

But there are also other practical considerations that could impact adoption:

• We’re still looking at promises, not proof. Differences in sample preparation, processing time, and throughput between platforms like Illumina and Visium HD could influence their efficiency and cost-effectiveness. Until independent researchers publish datasets using Illumina’s platform, we’re looking at promises, not proven performance.

• Experience with the details matters. Companies like 10x Genomics and Nanostring/Bruker invested heavily in experimental and computational improvements to enhance cell segmentation, one of the biggest challenges in spatial transcriptomics. If you can’t tell where a cell is, even the fanciest gene detection doesn’t help much.

• Is sensitivity the most important metric? Apparently, yes. Is it worth the cost? It depends. But this will be a topic for another blog post :)

Illumina’s entry won’t revolutionize spatial transcriptomics overnight, but it adds serious competition to a field that could use more options, and that’s good news for researchers everywhere.

In August 2020, I began my PhD with the idea of working on spatial domains, a concept still in its infancy. At the time, only one method had been published. Fast forward four years, and spatial omics has exploded in popularity, with over 87 methods for identifying spatial domains now available. That’s right. Eighty-seven.

In my last post, I shared how I ditched my fancy plan of using Graph Neural Networks and instead developed a simpler method called CellCharter1. Once it was ready, we needed to see how it compared against other methods, and for that, we turned to existing benchmarks used in the field.

However, we realized that nearly all benchmarks suffered from two of the most common pitfalls in statistics and machine learning.

The Human Cortex Dataset

To understand these pitfalls, let’s first look at the dataset most commonly used for benchmarking. It involves 12 samples of the human cortex, a brain region with six distinct layers, each containing different combinations of cell types.

Researchers used the Visium platform from 10x Genomics to capture gene expression in each spot of the tissue, providing spatially resolved transcriptomics data2. This grid-like dataset links gene expression with physical locations on the tissue, allowing scientists to assign spots to the appropriate cortex layer.

When the authors behind stLearn3, the first spatial domain method, tried to group spots based on their individual gene expression alone, the results were quite messy. The groups would only loosely match with the cortex layers and would not be well-defined regions.

However, when they considered each spot's gene expression along with its neighbors, the core idea behind spatial domains, the groups matched the layers better. Not perfect, but good enough for everyone to say, "Yup, let’s use this to benchmark everything else from now on.”

The ability to obtain spatial domains matching these cortical layers became the foundation for benchmarking almost every spatial domain method that followed.

But here’s the catch—actually, two catches.

Pitfall #1: Ignoring Randomness

All spatial domain methods include an element of randomness. Train the same model twice with identical settings, and you might get slightly different results. This variability isn’t inherently bad, but when it’s ignored during model comparisons, it can mislead us.

So, if a method is evaluated on a single run, it might come out on top just because the universe decided to be in a good mood that day.

Especially because small differences in the predicted labels can dramatically affect the score. For example, if I take the true labels and shift them by only a few spots, the similarity score drops from 1 to 0.66. That’s a big difference for something so small.

To address this, we had to take a more robust approach: running each model multiple times (we chose 10) and averaging the results.

Is it a pain to run? Yes. Does it make you sleep better at night? Also yes.

Pitfall #2: Overfitting

The second is the quintessential statistical pitfall, the one you learn about in any basic machine learning class and that would be the name of my music band if I had one: training on the test set.

Imagine you’re building a model to predict whether an image is of a cat or a dog. If you evaluate the model using the same dataset it was trained on, the model could just memorize the images and labels. As a result, it would perform well on the training data but fail miserably when given new images it hasn’t seen before.

That’s called overfitting.

At first, you might think, “Hey, this doesn’t apply here. The model doesn’t use labels during training, so no worries, right?“ Wrong.

Overfitting of the (hyper)parameters

The issue appears during parameter selection. Each method has parameters that need tuning, and a common practice here is to find the “best” values by evaluating the model on all 12 cortex samples.

Take BayesSpace4, for example. It has several parameters, such as the number of principal components and highly variable genes (HVGs). What is usually done is to test the score for every combination of parameters. What could go wrong? Well, a lot.

If you test thousands of parameter combinations, eventually you’ll stumble onto a combination that looks amazing. However, given that many attempts, it’s probably a lucky guess. That lucky score won’t hold up if you give the model a new sample it has never seen. It’s classic overfitting.

To avoid this, we used a more cautious approach: split the data into training and test sets. We trained on 3 cortex samples, tuned the parameters there, and then tested on the other 9 samples to see if the model could handle new data.

Where do we go from here?

Our changes made the benchmarking process more reliable, but also more complex. We couldn’t evaluate all 87 methods. Testing them all would’ve taken more than a full PhD.

In the end, we focused on the five most commonly used methods, but this raises an important question: how can we efficiently evaluate so many approaches?

The solution is to create a standardized framework that tests multiple methods across diverse, annotated datasets. This would allow researchers to easily plug in their new methods and test them across various tissue types and annotations. After all, if we only test on cortex data, we’ll find the best method for that specific tissue, not the best method overall.

But there is good news! People are working on it:

• SpaceHack2.0 started as a hackathon in 2023 and has morphed into a massive project building a framework to test 50 methods on 30 datasets using 20 different metrics. They’re looking for volunteers, by the way! They are also organizing SpaceHack3.0, focusing on best practices for analyzing imaging-based spatial transcriptomics.

• Also, there’s a new conference coming up: Modern Benchmarking: Advancing Computational Methods in Molecular Biology in March 2025. If you’re into benchmarking, it’s the place to be.

Bottom line: We need to do better at measuring success in spatial omics. It’s not just about playing with numbers. This stuff could actually help with real-world applications, like, you know, saving lives.

No pressure or anything.

In spatial biology, technologies generate overwhelming amounts of data, and it's easy to assume that the most sophisticated algorithms are the most effective. Yet, as I discovered during my PhD, there's often great value in taking a step back, starting simple, and letting the data guide your way.

During my Master's thesis, I became fascinated with Graph Neural Networks but didn't get the chance to apply them. As I entered my PhD, I wanted to find the perfect use case for them, and I thought I'd found it.

Once you have a hammer, everything looks like a nail.

However, my aspiration for complex methods unexpectedly led to a simpler, yet powerful solution to the problem.

The problem: spatial domain identification

Imagine you're looking at a city from above. You can see different neighborhoods - residential areas, commercial zones, industrial parks. Each of these areas has its own unique blend of buildings.

This is similar to what we see in biological tissues. Different areas of a tissue have their own unique compositions of cells. We call these areas "spatial domains", “spatial clusters”, “cell neighborhoods” or "cell niches". I talked about this metaphor in a previous post.

Understanding these domains is crucial because they can tell us a lot about how tissues function in health and disease. For example, the same immune cells can organize into different domains and influence the outcome of cancer therapies1.

The Graph Neural Network Approach

When I started my PhD, I was excited about using Graph Neural Networks (GNNs) to identify spatial domains. GNNs seemed perfect for the task, given their ability to process complex spatial relationships.

To understand how GNNs work, let’s return to our city metaphor. Imagine each building (representing a cell) shares information like energy usage and occupancy with its immediate neighbors. The neighbors then share this information with their neighbors, creating a ripple effect throughout the city.

In computational terms, each building updates its state based on the states of its neighboring buildings in a recursive manner. The core idea is that cells (buildings) surrounded by similar neighbors will have similar states, and thus belong to the same spatial domain (neighborhood).

A Pivot to Simplicity: Neighborhood Aggregation

I was all set to apply GNNs to find spatial domains. However, as any good researcher knows, before developing a complex model, it's important to create simple baselines.

This practice offers several advantages:

• Deeper understanding: Implementing baselines forces you to grapple with your data and problem, leading to better-informed decisions for more advanced models.

• Sanity check: If your complex model underperforms compared to a simple baseline, it may indicate issues with your implementation, data preprocessing, or model design.

• Incremental implementation: Baselines allow you to set up your data processing and evaluation pipeline step-by-step, without getting stuck in the complexities of advanced methods (yet).

With these principles in mind, I developed a method called neighborhood aggregation, a simpler, more scalable alternative to GNNs. Here’s how it works, using our city metaphor:

1. Buildings’s data: For each building, we gather data like energy usage and occupancy.

2. 1-hop neighbors’ data: We then move to the immediate neighbors (1-hop neighbors) and average their data. This gives us a set of average energy and average occupancy that we attach to the original building’s data.

   

3. 2-hop neighbors’ data: We repeat this process for neighbors of neighbors (2-hop neighbors), again averaging their data and concatenating it with the previous values.

   

4. N-hop neighbors’ data: We continue until a maximum distance is reached, which is usually around 2 or 3 jumps.

5. Domain identification: the concatenation of these N+1 sets of energy and occupancy data are used to groups buildings with similar values (also known as clustering).

Unlike GNNs, neighborhood aggregation is not recursive, making it much more scalable. It can incorporate information from all neighbor layers at once, capturing essential spatial relationships without repeatedly updating each cell's state.

This simple algorithm ended up being the main component of CellCharter, the method I developed throughout my whole PhD2.

Results and Implications

When we applied CellCharter to a spatial proteomics dataset of mouse spleens, the results were exciting. Our simple method identified domains that closely matched the known anatomical regions of the spleen, as annotated by an expert pathologist, and additional, more refined, domains.

This was a significant finding because it showed that even a straightforward approach could yield biologically meaningful insights. The method's simplicity also made it highly scalable, an important feature as the volume of spatial omics data continues to grow.

How to Apply This Approach to Your Research

This experience confirmed once again the principle of implementing simple baselines before shiny cutting-edge methods. These are the main takeaways:

1. Start simple: Beginning with a basic approach can provide valuable insights and sometimes lead to unexpected breakthroughs.

2. Balance novelty and utility: While it's exciting to work with cutting-edge methods, sometimes a straightforward approach that solves a real problem is more valuable.

3. Stay flexible: Be prepared to pivot your research direction based on your findings. Some of the most interesting discoveries come from unexpected results.

It’s easy to get caught up in the excitement of new techniques, but sometimes the simplest approaches can lead to the most impactful results.

While we usually think of these cells as defenders against cancer, tumors can manipulate them to support their growth. Spatial omics technologies are helping us understand how this happens, and these discoveries are not just academic. They are laying the groundwork for more precise and effective cancer therapies.

In the end, the future of cancer treatment may lie in targeting the interactions between these apparently benevolent cells.

Background: the fibroblast-macrophage connection

Scientists have long suspected that two types of normal cells might collaborate to support the tumor: fibroblasts and macrophages. What is their role in normal conditions?

• Fibroblasts help repair wounds by producing collagen and other fibers that form the scaffold of tissues, known as extracellular matrix.

• Macrophages are the body's janitors, capable of digesting pathogens, cell debris, and other foreign materials through a process called phagocytosis.

Tumors that have both fibroblasts and macrophages generally lead to worse survival rates compared to tumors with just one of these cell types.

Before spatial omics, these observations would come from counting the number of fibroblasts and macrophages in tumors and observing the progression of cancer, without understanding their spatial interactions or dynamics.

Without interaction information, most therapies were developed to target only a single cell type rather than trying to disrupt the interaction between multiple ones.

The Sibrotuzumab Case Study

Since fibroblasts and macrophages seemed to be helping tumors grow, it made sense for early clinical trials to target these cells individually. For example, Sibrotuzumab is a drug that targets FAP, a protein present in tumor fibroblasts but rare in normal tissues. In the case of a wound, the FAP protein can break the extracellular matrix and allow the arrival of immune cells to the site, but cancer cells can exploit FAP+ fibroblasts to spread more easily or develop more vasculature to feed the tumor.

However, phase 1 and 2 clinical trials of Sibrotuzumab in humans failed to show a reduction in tumor growth12. The failure of Sibrotuzumab in clinical trials raised questions, and researchers have proposed a few explanations:

• FAP is still active: even though Sibrotuzumab binds to FAP, FAP may still be able to degrade the extracellular matrix. How is that possible? Sibrotuzumab may bind to a region of the FAP protein that is not involved in remodeling the extracellular matrix.

• No killing ability: antibodies can induce an immune response towards the cells they bind to. However, Sibrotuzumab was not designed to provoke this response.

• Cell plasticity: fibroblasts are highly adaptive. It’s possible that targeting FAP induced fibroblasts into deactivating FAP but activating other proteins with similar functions.

• Marker vs cause: the tumor growth linked to FAP fibroblasts isn’t directly caused by FAP itself. Instead, while FAP could be used to identify these dangerous fibroblasts, other properties of the cells may be the ones responsible for the tumor growth, such as interactions with different cell types.

What do you think could be other reasons behind Sibrotuzumab's failure? Share your thoughts in the comments, I’d love to discuss them!

Advancements from Spatial Omics

When I first started my PhD, I hadn’t yet heard about the potential interaction between fibroblasts and macrophages. However, from 2022 on, when spatial omics started to pick up steam, these two cell types kept appearing together in study after study345.

All these studies noticed that fibroblasts and macrophages are not only present in the same tumor but are also localized in the same areas.

Not an obvious conclusion, as we saw in the previous blog post.

And that’s not all.

As we will see in the rest of the blog post, thanks to spatial omics, we now have new ideas about what is happening between these cells.

Advancement one: what’s happening inside the tumor?

A recent study6 discovered the dynamics between fibroblasts and macrophages in these tumors. The authors gave immunotherapy to liver cancer patients and classified the patients into two groups: responders and non-responders.

In non-responders, fibroblasts and macrophages form a barrier around the tumor. This barrier acts like a shield, preventing immune cells, specifically T cells, from entering the tumor core and attacking the cancer cells.

This could explain why Sibrotuzumab failed. Even if Sibrotuzumab managed to disrupt the barrier, it’s still not guaranteed that T cells can enter the tumor core, simply because some patients may not have a strong immune response in the first place.

Combining Sibrotuzumab with immunotherapy could help the non-responders break the barrier and bring T cells to the tumor. Still, no one has tested this hypothesis as far as I know.

Advancement two: how are they communicating?

Some studies started to identify the molecules involved in the interaction between fibroblasts and macrophages. For example, by looking at which of the 20,000 genes are active whenever these two types of cells are near each other.

This opens up many possibilities for new cancer therapies. Since they are based on combinations of cell types, they have higher specificity, meaning fewer side effects.

Advancement three: easier to detect

During my master’s thesis, I explored a computational approach called Graph Neural Networks but never had the chance to use it. When I transitioned to my PhD, I decided to work on finding cell communities from spatial omics mainly because they were the perfect use case for Graph Neural Networks.

Spoiler alert: I didn’t use Graph Neural Networks in my PhD either. But this is a story for future blog posts.

Anyway, the more time I spent developing a method to find cell communities, called CellCharter, the more I realized its clinical applicability. Structures like the fibroblast-macrophage barrier are indeed communities and tools like mine could easily detect them.

This means that it could be used to find new potential cancer targets and to assign the right treatment to patients based on the presence or absence of certain interactions.

Today, we stand at the beginning of another potential revolution – spatial omics. Despite the uncertainty, in this post and the next, I am going to show how this technology could mature into an established pillar of biology and medicine.

To explain where spatial omics stands today - on the cusp of medical breakthrough and public anticipation - I will first tell you about the rise and fall of the hype surrounding interferon therapy.

A look back at interferon therapy

It was 1957. At the National Institute for Medical Research, Alick Isaacs, a British virologist, teamed up with Jean Lindenmann, a Swiss post-doc, to study the body's defenses against viral infections. During their studies, they observed that virus-infected cells produced a substance that interfered with the infection of nearby cells by the same or other viruses. Naming this substance was straightforward, given its unique function. They called it interferon, alluding to its capacity to interfere with viral replication. The discovery was groundbreaking.

Initial studies showed promising results, showcasing its powerful antiviral and anticancer properties. The media played a substantial role in elevating the profile of interferon. In the late 1970s and early 1980s, news outlets began to report on the miracle drug that might cure all cancers and combat all viruses. This led to significant investments and competition among biotech firms and pharmaceutical companies to produce and patent interferon therapies.

Over time, as more research was conducted, a deeper understanding of interferon therapy emerged. It became clear that while interferons had therapeutic value in specific contexts, they were not the panacea that early hype suggested. Its application was not universal, and its side effects were not suitable for all patients1.

In essence, the case of interferon therapy underscores the complexities of translating scientific discovery into medical practice. Hype can accelerate interest and funding but can also set unrealistic expectations. As with many medical innovations, the true value and applicability of a therapy often become clearer with time, rigorous research, and real-world clinical experience.

What about spatial omics?

Fast forward to the present, and a technology called spatial omics is capturing a lot of attention in the research community.

Without going into the technical details, spatial omics studies cells in their spatial context within the tissue. As I described in my previous post, it's like studying a city from a detailed map of a city (the tissue sample), whereas before it was only possible to do so from a list of buildings (the list of cells in the tissue).

Spoiler alert: it’s too early to say. Even though the pace at which technology is evolving seems to increase exponentially, spatial omics only started to emerge as a field in 2016. We are right now during the equivalent of the period between 1957 and the 1970s for interferons.

Early studies that used spatial omics in cancer often reached conclusions such as detecting high heterogeneity of the tumor cells within the same patient (called intra-tumor heterogeneity), which is one of the main causes of therapy failures. However, intra-tumor heterogeneity has been a well-known concept long before spatial omics2, and is detectable using simpler, cheaper technologies.

Yet, there are certain aspects that only spatial omics can capture.

In our study on CellCharter3, a computational method to find recurring communities of cells from spatial omics data, we observed something intriguing. Some lung cancer patients had similar proportions of non-tumor cell types, with a high proportion of cells called neutrophils and fibroblasts.

Through the lens of non-spatial technology, one might think that these patients are very similar, in the sense that their tumor cells are placed in very similar microenvironments.

But once we looked at how those neutrophils and fibroblasts are spatially arranged, a totally different picture came up. These two cell types were intermixed in Patient 12 but physically separated in Patient 9. In these two patients, the cellular communities around the tumor, and the interactions between cells in those communities, are radically different.

Cell interactions are becoming increasingly relevant for cancer therapies, especially immunotherapy, so these new observations may be the key to understanding why a therapy works or fails for a certain patient.

We still don’t have data on whether these differences impact the outcomes of therapies. As I said, we are still in the infancy period of this technology, and having clinical information for these patients is rare. It will take years to prove, or disprove, these observations, but we are finally reaching the equivalent of the 1970s and 1980s for interferon. Hopefully, the research will lead to different and more promising conclusions.

In our bodies, cells do not function in isolation but they are parts of a complex, evolving ecosystem. When I was a kid, the animated series “Once Upon a Time… Life” was popular among children of my age. It represented entities within our bodies – like cells, antibodies, bacteria, and viruses – as interdependent members of a society. Cells would interact by competing and cooperating.

Scientifically, capturing these interactions from biological experiments on a large scale has always been difficult. Now, a new technology called “spatial transcriptomics” allows us to do that through an important aspect: cells need to be close to each other to interact. If two cells are far apart, it is unlikely that molecules from one will reach the other.

But how does spatial transcriptomics measure the proximity of cells to figure out how they interact?

To understand this, in this post we will explore the evolution of this technology. People in the field have used many metaphors to explain the transition from non-spatial to spatial technologies. Over time, I realized that one specific metaphor is particularly accurate, but I have never seen it described explicitly.

The society of cells

Imagine being the mayor of a city. A city is composed by various types of buildings, such as residential buildings, commercial buildings, banks... Every building is inhabited by different types of people. For example, commercial buildings may have chefs, waiters, and cashiers, while banks might be filled with office workers and managers.

Similarly, an organ is composed of different types of cells, such as tumor cells, immune cells, neurons... Every type of cell is characterized by the activity of different genes. The activity of certain genes results in the production of associated proteins, which, just like people, are the molecules that carry out tasks (structure, storage, metabolism) and that are used to communicate between cells.

To summarize, in this metaphor:

• organ = city

• cell = building

• gene/protein = person

In this society, your goal as the mayor is to understand the city's health, predict its evolution, and find ways to improve it.

Bulk RNA sequencing: the city-wide overview

Since 15-20 years ago, as a mayor, the only information you could access was a general view of the city. You received a broad, averaged-out picture of the city's characteristics, like surface area, number of inhabitants, average income, unemployment rate, and crime rate.

This mirrors the technology called bulk RNA sequencing. Here, you get a measurement of activity for each gene, but it’s an average over all the cells in the sample, much like aggregated statistics of a city can hide the unique features of every building.

This can already provide a lot of information, but it is certainly not enough.

Single-cell RNA sequencing: the building list

Realizing that this broad view isn't enough to make effective policies, you decide to inspect individual buildings. You send out teams to visit each building, count the number of people, understand what they are doing, and how they interact within that building. This way, you get detailed information about each building's function and the activities within it.

You can see whether a city has a healthy proportion of buildings, with the right balance of residential buildings, and commercial ones. If 80% of the buildings are Starbucks, probably the city doesn’t have a promising future.

This is equivalent to single-cell RNA sequencing, developed 10-15 years ago, where the gene activity is measured for each cell individually. By analyzing individual cells, you can identify specific cell types, their functions, and their gene expression profiles. You understand the diversity and complexity of people and their jobs (proteins) within each building (cell).

Spatial transcriptomics: building the map

While inspecting individual buildings has given you a lot of detailed information, you now realize that the context and interactions between buildings are crucial. Two cities composed by the same proportion of residential and commercial buildings can look very different. One city could be more US-style, divided into residential areas, without a single shop nearby, and commercial areas, where no people live. The other one could be composed of a European-style mix of residential and commercial buildings.

You decide to hire a team of cartographers to map the entire city, detailing not only the function of each building but also their exact locations and interactions with each other.

This is equivalent to spatial transcriptomics, developed around 5-8 years ago, which allows you to build a map of the city. You can understand not only the function of individual cells but also how they are arranged. You can now explore the connections between buildings (i.e., cell-cell communication) and identify neighborhoods (i.e., spatial domains).

A snapshot in time

An important limitation of spatial transcriptomics is its equivalence to capturing a city's map at a single point in time. The process can only be done once, because the cells in a tissue sample die once extracted from a patient, preventing any observation of how the cells might have evolved. This means that we have spatial information, but we cannot observe its temporal evolution. Spatiotemporal data are still a distant dream, though a few pioneering efforts are bringing us closer to this reality, which I will explore in future posts.

Despite this limitation, the unexplored areas of biology now accessible through spatial transcriptomics are surely going to keep us busy until the next technological breakthrough.

I have also been guilty of using many metaphors to describe spatial transcriptomics. If you want to have another point of view, last year I made a YouTube video describing spatial transcriptomics as a fruit pie.

There are only two hard things in Computer Science: cache invalidation and naming things. - Phil Karlton

Choosing a name for this blog took longer than I had anticipated. I knew it would focus on topics related to Computational Biology and Cancer. It would also include my other academic interests like the analysis of spatial omics data. Don’t worry if you don’t know what spatial omics is, you’ll have plenty of chances to hear about it in future posts.

Subscribe now

The challenge was finding a name that would connect the biological world of cells, tissues, and experimental protocols to the computational world of models, metrics, and analyses.

In computational biology, batch effects appear when comparing two biological samples that have been processed differently. Even the smallest difference in the experiment can cause changes in the data that are hard to remove because they often overlap with true biological variation. Let’s see an example.

Degradation or downregulation?

Imagine two samples for which we want to do sequence the RNA: one from a healthy individual and another from patient with a certain disease. Due to logistics, the samples from the healthy individual are processed immediately after collection, while the other one is left at room temperature for a few hours before processing.

Over time, RNA can degrade, especially when exposed to room temperature. This degradation can decrease the abundance of certain RNA molecules. Not all RNAs may degrade at the same rate because some may me less stable than others. As a result, when researchers sequenced the RNA of these samples, the data might indicate decreased levels for specific transcripts in the disease sample compared to the healthy one.

Now, suppose this particular disease inherently causes the inactivation of certain genes. Then, it will be complicated to disentangle whether a gene is expressed at lower levels because of true biological variation or because of the batch effect from RNA degradation.

Why Batch Effect?

Batch effects are effects that every computational biologist will encounter at some point in their career. They are a reminder to young scientists like me, pulling us back from an idealized perception of how science and experiments work.

Batch effects remind us that data, especially biological data, is messy, complex, and inherently noisy. They represent the dance between nature and technology, between the raw truth of biology and the structured logic of computation.

That’s why I chose “the Batch Effect”.

Through Batch Effect, I want to navigate the complexities of cancer research, explore the promising field of spatial omics, and discuss the computational strategies that are pushing the boundaries of what we know about biology.