Introducing the CellularAI algorithm: Predicting single-cell responses to genetic perturbations

November 19, 2026

On November 19, 2026, the Illumina BioInsight AI Lab announced CellularAI, an algorithm designed to predict transcriptome-wide responses to genetic perturbations in unseen cell types. Illumina plans to release an open-source initial version of CellularAI, providing transparency into the algorithm’s approach and giving researchers an opportunity to explore with their own perturbation reference data. This release represents an early step in an active area of Illumina R&D, with advancements already in development.

The CellularAI algorithm adds a predictive layer to the cellular perturbation ecosystem being developed by Illumina. Perturbation responses can now be observed at scale through data assets like the Billion Cell Atlas or through public datasets, responses can be predicted using the CellularAI algorithm for cell types that have not yet been experimentally tested, and resulting hypotheses can be tested experimentally through direct CRISPR capture capabilities available through the Illumina Single Cell Preparation workflow.

In designing CellularAI, the BioInsight AI Lab developed the algorithm to account for the inherent noise in CRISPR technology and cellular biology while demonstrating significant gains in performance as underlying cellular perturbation data becomes richer and more comprehensive.  The methods section of this blog provides an overview of the operating principles underlying CellularAI, as well as performance data demonstrating 2X higher performance in comparison to legacy predictive standards (mean level of expression across measured cells).

The release of CellularAI follows the algorithm’s first-place finish in the Arc Institute’s Virtual Cell Challenge, an annual international competition focused on developing algorithms that predict how cells respond to genetic or chemical perturbations. Click here to read more about how Illumina participated in the Virtual Cell Challenge in a recent press release.

Pictured above are Atreyo Pal, Eyal Ben David, Kishore Jaganathan, and Aniketh Janardhan Reddy, the Illumina team behind the CellularAI Algorithm and winners of the Arc Institute Virtual Cell Challenge.

If we can measure biology, why predict it?

For over a century, scientists across the globe have faced a universal challenge: how can we model the incredible complexity of biology in a laboratory setting?

Consider human biology. Even a single tissue can contain a tremendous diversity of cell types, more closely resembling an interactive, deeply intertwined ecosystem than a collection of individual cells. Within these systems, new mechanisms, cellular identities, and regulatory functions are still being uncovered today. A stroll through almost any conference poster session will reveal just how much remains to be discovered. Each new finding adds to an ever-growing body of evidence for how cells, tissues, and whole organisms function.

Modeling this complexity has traditionally required simplification. 

As the French biochemist Jacques Monod famously explained, "What is true for E. coli is true for the elephant." In other words, biology operates according to fundamental principles that can often be studied in more laboratory-friendly systems. By understanding those principles, scientists can begin to uncover the mechanisms underlying seemingly more complex organisms, including humans.1

Translatability to human biology is inconsistent across models

While Monod was certainly correct in that fundamental principles of life are, indeed, shared between organisms, the genome era has demonstrated that findings in simpler or non-human models do not always translate completely to humans.  In fact, a recent study found that only 5% of medicines that appear effective in mouse models ultimately obtain regulatory approval for human applications.2

Given the roughly $1 billion R&D cost per new medication, the ~90% failure rate of drugs entering discovery pipelines, and the 10–15-year timeline for a candidate to reach the market, one thing is clear: researchers need the right human-relevant data applied as early as possible to decrease failure rate and improve candidate viability before tremendous costs are incurred (Figure 1).3

Figure 1: Success rate of drug candidates throughout the average drug development pipeline. 

Digitally representing cellular biology – a community-wide quest to beat the average

The concept of a virtual cell model, or the ability to accurately predict human cellular biology, is not new. Nonprofits, AI labs, biopharma companies, and academic institutions have all pursued models that could help predict biological behavior and, at the very least, point researchers in the right direction for subsequent wet-bench experiments. Of great importance for pharmaceutical and translational research is a specific type of biological prediction – the ability to predict what happens to cells in response to a perturbation such as the activation or inhibition of a given gene. 

Yet for accurate prediction of perturbation consequences, the field has repeatedly encountered a deceptively simple obstacle: outperforming baselines. Historically, published models have struggled to predict biologically meaningful features more accurately than simply taking the average across examples.4 In other words, despite increasingly sophisticated algorithmic approaches, it has remained surprisingly difficult to demonstrate that a model can meaningfully improve prediction of biological responses beyond a simple mean.

As the genomic and AI eras converge, better predictions of biology will ultimately depend on better data. But models don’t just need more data. They need data that represent the diversity and biological context of the systems they aim to predict. For cellular biology, that means causal, mechanistic information measured across tremendous cellular diversity and biological contexts.5 Population genomics programs such as UK Biobank and PRECISE (Figure 2) operate on a similar principle: greater representation across people, genetic backgrounds, and molecular profiles provides a richer view of human biology.6 Across both populations and cells, what we can predict depends fundamentally on the biology represented in the data. 

Figure 2: Representation shapes what biological data can reveal. Large scale genomics programs such as PRECISE illustrate how population-specific representation can create a richer view of human biology. In Singapore, locally representative genomic data have helped characterize genetic variation in the context of the local population, creating a stronger foundation for locally relevant research and implementation. Or in other words, what we can predict depends fundamentally on the biology represented in the data.

Building the right dataset: directly observing the consequences of genetic perturbations across the greatest possible diversity of cellular contexts

Most large single-cell datasets are observational. They document the inherently noisy biology of cells as they are, including which genes tend to rise and fall together, but not which gene is driving which. Perturbation data is different. In a CRISPR interference (CRISPRi) experiment, for example, a disabled Cas9 fused to a repressor sits on a gene’s promoter and turns it down without cutting the DNA. Pair that with single-cell RNA sequencing, which reads out transcriptome-wide expression in every individual cell, and researchers can watch how silencing one gene has systemic influence on a variety of cellular outcomes. When applied across a wide diversity of cell types, the outcome is a rich dataset of directly-measured mechanistic biology, rather than correlation.7,8

The idea behind the Illumina Billion Cell Atlas was to take the concept of Perturb-Seq to an extreme – what if Illumina could build one of the world’s most diverse, comprehensive dataset to directly measure what happens when each gene changes, repeatedly, across many kinds of cells?

Since its announcement in January 2026, the BioInsight AI Lab has done just that. The team has sequenced in total over 500 million single cells using the Illumina Single Cell Prep assay spanning multiple organ systems and disease representations in just 9 months, with billions more currently being planned. To capture as many biological contexts as possible, a significant span of human organ systems (ie, cardiovascular, immune, neuro, renal, and more), cellular identities (ie, induced pluripotent stem cells, primary cells, organoids, immortalized cell lines, and more), and disease states (ie, oncology, metabolic, immune, reproductive biology, and more) are included in the continually growing Billion Cell Atlas (Figure 3). The data generated thus far reveal that different cell types respond very differently to the same genetic perturbations. Seeing this variation in response is crucial for building strong predictive models because it allows a model to understand what is shared, and what is cell-type specific.

Figure 3: Sampling of broad organ system and cell lineage coverage in the Illumina Billion Cell Atlas.  The atlas was designed to support a broad range of translational research applications, including drug discovery and response prediction. Its composition spans a diverse spectrum of cells, encompassing those widely used in research today as well as those positioned to enable emerging and future applications.

A dataset like this raises two important questions: how do we make it work at this scale, and how do we know it worked?

Hyperscale biology requires a hyperscale platform

Capturing hundreds of millions of cells in months, and scaling to billions of cells, is largely possible because of the scalability of the Illumina Single Cell Prep Assay, in combination with rapid analysis through DRAGEN. The BioInsight AI Lab overloaded cell capture reactions from the Illumina Single Cell Prep CRISPR workflow, yielding an impressive 1.7 million recovered cells per reaction. Despite the tremendous scale of cellular capture, exceptionally high-quality data were obtained, on-par with expected quality metrics and with significant cellular population diversity detected across experiments (Figure 4).

Figure 4: Hyperscale capture of a mixed cell population using Illumina Single Cell Prep. Illumina Single Cell Prep kits with direct CRISPR capture oligos were overloaded with ~4 million cells per reaction, yielding 1.7 million total captured cells, high-quality data, and proper separation via UMAP plot

Of course, capturing diverse cells at scale only matters if CRISPR constructs actually deliver the intended perturbations. Such checks are performed as data are generated and consistently demonstrate that the Billion Cell Atlas generates high-confidence mechanistic insights.  Not only do perturbations produce the intended biological effects, but they do so at a level of efficiency that matches or exceeds published examples from across the industry (Figure 5A).9,10 In Figure 5B and Figure 5C, for example, BioInsight AI Lab researchers demonstrated that across 6 million total cells from two cell lines that are commonly-utilized in pharmaceutical research, the assay produced 2-1000 fold repression across CRISPR library-targeted genes. 

Figure 5: CRISPRi performance metrics. Performance of CRISPRi knockdown, both for pilot and coding genome-wide CRISPRi libraries, was compared across multiple cell lines and to previously published examples (A). Knockdown efficiency was also assessed transcriptome-wide (B, C), with a representative snapshot of results shown here and 2-1000 fold repression observed across samples. 

Sixty cell types teaching us about the sixty first: how the CellularAI algorithm works

When the Arc Institute Virtual Cell competition kicked off, the BioInsight AI Lab had accumulated data from about 200 million CRISPRi-screened cells, capturing both the perturbation that was made and the transcriptome-wide response that followed. This provided the foundation for a different approach to cellular prediction. While previous approaches largely focused on training neural networks on vast amounts of single-cell data to learn how cells behave, CellularAI made a different bet: the best predictor of how a new cell responds to losing a gene is what actually happened when that same gene was lost in other, biologically similar human cells.

The principles for how to actually accomplish this concept came about through an incredibly iterative process – the lab built several independent evaluation harnesses from public perturbation datasets, allowing for scoring of every algorithm in a manner aligned with the Arc Institute’s scoring guidelines. Multiple AI agents then iterated on hundreds of ideas proposed by BioInsight scientists to rapidly identify the best performing approaches.

To demonstrate the overarching philosophy of how the algorithm works, think about how forecasters estimate rainfall across a region. No single weather station settles a given question about an upcoming weather event. Instead, forecasters pool readings from many relevant stations, weight each by how reliable its instruments are, and account for the fact that stations can disagree for real reasons, like terrain and elevation, not just faulty sensors. The CellularAI algorithm applies this logic to cells. Each cell line available in the reference library is treated like a weather station: its own reading of the same question is considered from its own vantage point.

With these concepts in mind, several operating principles guided the version of the CellularAI algorithm developed for the Arc Institute Challenge. The central idea being that virtual cells are built from real cells. CellularAI doesn’t generate cells from scratch. It starts with real, unperturbed cells from the new cell type and shifts each gene’s expression up or down by the amount the model predicts for that shift. Genes with no supporting evidence are left exactly as they were. Because every virtual cell begins as a real one, it keeps the natural variation of real single-cell data: most genes are detected in only some cells, and expression varies widely from one cell to the next.

The next question, then, is determining how much each gene should move. In the version developed for the Arc Institute Challenge, four key principles helped determine how experimental evidence was weighted:

  • Every experiment is weighted by how efficiently a given CRISPR construct knocks down its intended target. As shown in Figure 5B and Figure 5C, CRISPRi guides don’t silence their targets equally well in every cell type. Experiments where the knockdown removed 60% or more of the target’s mRNA count fully, and weaker ones count proportionally less.

  • Noisy measurements count for less. Some genes are naturally harder to measure than others. Their readings bounce around from cell to cell even when nothing has been done to them. The CellularAI algorithm estimates that background noise for every gene in every cell line, using knockdowns that have essentially no effect as a baseline, and gives more weight to results that stand out clearly from the noise. A change in a stable gene is therefore trusted more than the same change in a gene that fluctuates a lot.

  • Caution is built in. The CellularAI algorithm combines the results from all the cell lines into a single consensus, using a statistical method (a random-effects meta-analysis) designed for sources that can disagree for real reasons, not just by chance. If a response shows up in 2 of 50 cell lines, the consensus stays small, and a gene moves strongly in a prediction only when many cell types agree. When evidence is weak, sparse, or inconsistent, the prediction defaults to "no change."

  • Relatives respond alike. A cell’s identity shapes how it responds to losing a gene, so cell types that resemble each other tend to respond similarly. The CellularAI algorithm measures that resemblance directly. It compares the new cell type’s unperturbed cells with each cell line in the reference library. The comparison uses genes that best distinguish one cell type from another. Cell lines that are most similar receive greater weight in the prediction. A cap prevents any single close match from dominating. The algorithm also makes a small, conservative adjustment for how strongly the new cell type expresses each gene, since a gene that is barely expressed has little room to fall.

The CellularAI algorithm therefore entered the Arc Institute Challenge with no large neural network trained on the target cell type. A meta-analysis sits at its core, and every predicted change can be traced back to the measured experiments it came from. Since the challenge, the algorithm has been simplified and streamlined for open-source release, retaining the components that contributed most meaningfully to predictive performance while making the approach more transparent and practical (Figure 6).

Figure 6: Overview of CellularAI algorithm architecture. CellularAI optimally functions and recapitulates the results seen in the Arc Institute challenge using Billion Cell Atlas data as reference materials, however users of the first version of the algorithm can use CellularAI to weight their own experimental evidence to make predictions around cellular response to genetic inhibitions. See Box 1 for details regarding each correction made by the open source algorithm.  

Overcoming a historic obstacle: outperforming basic statistical baselines

Because predictive models of perturbation biology historically struggle to predict transcriptomic outcomes more accurately than simply taking the average across examples, the BioInsight AI Lab spent a significant amount of time determining if the CellularAI algorithm adequately accounted for enough meaningful biology to make better predictions.4

To this end, the team tested the algorithm on three cell types where the real answer was already known: human embryonic stem cells (H1) from the 2025 Virtual Cell Challenge, the K562 leukemia cell line from the published Replogle et al. CRISPRi screen, and a HEK293 kidney cell line screen.9 For each test, that cell line’s own data was first removed from the reference library, so the algorithm had to predict the response from other cell types rather than look it up. In other words, the algorithm never saw the real perturbation data it was being asked to predict.

These tests also allowed the team to ask what actually makes predictions better. Simply averaging the responses seen across other cell types provides a surprisingly strong starting point, but the algorithm consistently improved on that baseline. In Figure 7, the BioInsight AI Lab intended to demonstrate two key factors underlying improved prediction: the value of increased cellular diversity, and the impact of iterative data normalization, weighting, and correction according to patterns seen in the underlying data.  Using this relatively straightforward approach via quality underlying data and algorithmic steps, CellularAI performed 2X as well as the mean. Each step (Box 1) performed by the algorithm added additional performance, and increasing the number of cell lines from 5 to 50 yielded significant gains.

Figure 7: Algorithmic performance increases with greater diversity and quantity of data.  The relative contribution of each step in data normalization, weighting, and correction were assessed alongside the influence of increased quantity and diversity of underlying CRISPRi-screened cell lines. Cell lines closest to assessed benchmarks were removed for the purposes of this assessment, alongside optimizations that were performed to maximize performance in the Arc Institute Virtual Cell Challenge.  Performance metrics are made in comparison to the full repertoire of underlying data and optimizations used by the BioInsight AI Lab for the Virtual Cell Challenge. 

Putting it to the test: the Arc Institute Virtual Cell Challenge

The Arc Institute’s Virtual Cell Challenge gives contestants the real, unperturbed cells of new cell types and a list of 300 genes to knock down. Contestants must generate 400 virtual cells per knockdown, per cell type, which are then compared to real experiments they never see. It’s a demanding test. The cell types have never been perturbed, and most knockdowns change only a small fraction of the roughly 18,500 genes, so a tool must pick out the few genes that respond across about 300 different knockdowns.

The BioInsight AI Lab is grateful for the opportunity to participate in this community-wide effort and to the Arc Institute for creating a rigorous benchmark for progress in virtual cell modeling. Challenges like this give the field an opportunity to test different ideas against the same unseen experimental data, learn from one another, and collectively push the science forward. While final results from the Arc Institute Virtual Cell challenge were not yet available at the time of blog publication, we'll update this post with CellularAI algorithm performance metrics shortly following the challenge's conclusion.

Observe, predict, test

The CellularAI algorithm is part of a connected ecosystem of Illumina-enabled single-cell research solutions:

  • Observe: Perturbation responses can be directly observed at scale through data assets like the Illumina Billion Cell Atlas, a growing dataset currently available through a consortium partnership model.
  • Predict: Cellular responses to perturbations can be predicted in cell types that have haven’t been yet experimentally tested through the new CellularAI algorithm.
  • Test: Predictions and hypotheses can be addressed or validated at the bench using direct CRISPR capture in the Illumina Single Cell Prep workflow.

Each step feeds the next. New experiments expand a reference library, a richer and more diverse library gives the CellularAI algorithm more biological evidence to draw from, and new predictions can guide the next experiments.

Open-source, built for the community

A first, streamlined version of the CellularAI algorithm is being released as open-source to provide transparency into the algorithm’s methodology and enable researchers to explore the approach with their own compatible data. The release does not include the Billion Cell Atlas nor was it trained on the Billion Cell Atlas. Rather, CellularAI provides the analytical framework that can operate on a reference library, such as the Billion Cell Atlas, to make predictions. Current partners in the Billion Cell Atlas consortia will be able to utilize CellularAI to its fullest potential, on the largest possible reference dataset available.11

This open-source release is intended as a window into an active area of research at Illumina, rather than an endpoint. The BioInsight AI Lab is developing more advanced approaches that build on lessons from CellularAI, the Virtual Cell Challenge, and the expanding diversity of perturbation data generated by Illumina. A full manuscript providing greater detail into the methods underlying this first version of the CellularAI algorithm will be released in early 2027.

Collaborate with BioInsight

As the BioInsight AI Lab improves and develops new advanced annotation tools, the foundational importance of high-quality training data has been exceptionally clear. Every model that the lab has produced performs best on data that resemble its training data. Therefore, the diversification and expansion of training data are crucially important to further advance capabilities and accuracy.  

To accelerate development in this critical field, BioInsight welcomes opportunities for collaboration, both for CellularAI and other models in development.  Contact BioInsight-Info@Illumina.com to inquire regarding improvement of cohort-specific performance, rare disease research, improved ancestry representation, model evaluation, research into future therapies, or other relevant topics. 

BioInsight also welcomes additional partners interested in accessing the world’s largest genome-wide perturbation dataset. The Billion Cell Atlas Is built to accelerate drug discovery and to provide significant diverse data for AI model development. Those interested in the Billion Cell Atlas or further perturbation biology partnerships are encouraged to contact BusinessDevelopment@Illumina.com to start a conversation.

About the BioInsight AI Lab

The BioInsight AI Lab develops algorithms intended to predict complex biology, especially for difficult areas of the genome, noncoding regions, and complex phenotypic effects. Previous work includes the launch and widespread research use of models such as SpliceAI2, PrimateAI-3D, Promoter-AI, FlexRV, and more.

Acknowledgements

The CellularAI algorithm was built under the leadership of Kyle Farh. Kishore Jaganathan, Eyal Ben David, Aniketh Janardhan Reddy, and Atreyo Pal are key contributors, utilizing data generated by more than 50 scientists in the BioInsight wet lab running perturbation experiments at scale. We are grateful to Pavel Avdeyev, Brady Gilg, and the many involved for all that has been accomplished.

The blog post was authored by Laurel Mastro, Kishore Jaganathan, Eyal Ben David, and Robert Yamulla. 

References

  1. Ball P. What distinguishes the elephant from E. coli: Causal spreading and the biological principles of metazoan complexity. J Biosci. 2023;48:14. PMID: 37194562.
  2. Ineichen BV, Furrer E, Grüninger SL, Zürrer WE, Macleod MR (2024) Analysis of animal-to-human translation shows that only 5% of animal-tested therapeutic interventions obtain regulatory approval for human applications. PLOS Biology 22(6): e3002667. https://doi.org/10.1371/journal.pbio.3002667
  3. Sun D, Gao W, Hu H, Zhou S. Why 90% of clinical drug development fails and how to improve it? Acta Pharm Sin B. 2022 Jul;12(7):3049-3062. doi: 10.1016/j.apsb.2022.02.002. Epub 2022 Feb 11. PMID: 35865092; PMCID: PMC9293739.
  4. Csendes, G., Sanz, G., Szalay, K.Z. et al. Benchmarking foundation cell models for post-perturbation RNA-seq prediction. BMC Genomics 26, 393 (2025). https://doi.org/10.1186/s12864-025-11600-2
  5. Dibaeinia P, Babu S, Knudson M, ElSheikh A, Wen Y, Liu H, Perera J, Khan AA. Virtual Cells Need Context, Not Just Scale. bioRxiv [Preprint]. 2026 Feb 9:2026.02.04.703804. doi: 10.64898/2026.02.04.703804. PMID: 41727131; PMCID: PMC12919078.
  6. The Power and Necessity of Genomic Diversity in Research.” STAT, Sponsored Insight, 26 Sept. 2022, www.statnews.com/sponsor/2022/09/26/the-power-and-necessity-of-genomic-diversity-in-research/
  7. Dixit A, Parnas O, Li B, Chen J, Fulco CP, Jerby-Arnon L, Marjanovic ND, Dionne D, Burks T, Raychowdhury R, Adamson B, Norman TM, Lander ES, Weissman JS, Friedman N, Regev A. Perturb-Seq: Dissecting Molecular Circuits with Scalable Single-Cell RNA Profiling of Pooled Genetic Screens. Cell. 2016 Dec 15;167(7):1853-1866.e17. doi: 10.1016/j.cell.2016.11.038. PMID: 27984732; PMCID: PMC5181115.
  8. Pratapa, A., Jalihal, A.P., Law, J.N. et al. Benchmarking algorithms for gene regulatory network inference from single-cell transcriptomic data. Nat Methods 17, 147–154 (2020). https://doi.org/10.1038/s41592-019-0690-6
  9. Replogle JM, et al. Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq. Cell. 2022 Jul 7;185(14):2559-2575.e28. doi: 10.1016/j.cell.2022.05.013. Epub 2022 Jun 9. PMID: 35688146; PMCID: PMC9380471
  10. Huang, A. C. et al. X-Atlas/Orion: Genome-wide Perturb-seq datasets via a scalable fix-cryopreserve platform for training dose-dependent biological foundation models. bioRxiv (2025). https://doi.org/10.1101/2025.06.11.659105
  11. https://www.illumina.com/areas-of-interest/genomics-in-drug-development/cell-atlas.html