INTRODUCTION

Learning Regulatory Cascades from Tahoe-100M

Gene regulatory networks provide a compact language for cellular behavior and response; which regulators appear to coordinate transcriptional programs, how those programs organize across cell types, and, when interventional evidence is available, how upstream changes propagate to downstream expression. In practice, most networks inferred from transcriptomes remain dominated by association. Single-cell RNA-seq amplifies this problem, strong latent state structure, cell cycle, stress, differentiation, together with technical effects, induces broad covariation, so many edges reflect shared context rather than direct regulation.

This is why “correlation is not causation” is not a slogan here; it is the identifiability problem.Two genes can rise and fall together because one regulates the other, because they share an upstream driver, or because both respond to an unmodeled shift in cellular state along a low-dimensional manifold. Motif priors and regulon-style approaches improve biological plausibility, but they do not, on their own, separate direct regulatory influence from downstream consequence. If we want directed edges that generalize, and not just descriptive modules, then we need information that is absent from observational snapshots. Perturbation datasets provide that information by injecting asymmetry into the data. With many interventions, especially when they span diverse targets, doses, and cellular backgrounds, the transcriptome becomes a readout of repeated “natural experiments,” where some changes are proximal and others reflect propagation through the network. Tahoe-100M was built to operate in that regime: ~100 million single-cell transcriptomes, spanning 379 compounds across 50 cell lines.

In this post we introduce TX1-CD (TX1 Causal Decoder), a structural-equation layer trained on top of Tahoe-X1(TX1) embeddings that decomposes each perturbation’s footprint into direct and cascading components, producing a directed GRN over 14,142 genes.

INFRASTRUCTURE

Eliminating Key Bottlenecks in Large-Scale Single-Cell Processing and Modeling with NVIDIA GPUs

Tahoe-100M pushes single-cell analysis into a regime where GPU acceleration isn’t optional. With ~ 100 million cells to process, even “standard” steps—normalization, covariate regression, neighbor graph construction, and large-scale gene–gene statistics—become throughput problems.

Our pipeline runs end-to-end on NVIDIA H200 GPUs. For single-cell preprocessing, we collaborated with scverse and NVIDIA to leverage rapids-singlecell (developed by scverse), which provides a drop-in replacement for Scanpy workflows on top of NVIDIA CUDA-X Data Science (DS). Under the hood, the core acceleration layer includes cuDF, cuML, cuGraph, cuVS, Dask-CUDA, cuBLAS, and RMM (RAPIDS Memory Manager); model training uses PyTorch DistributedDataParallel with the NCCL backend.

Pipeline Performance on 8× NVIDIA H200 (OCI Bare Metal)

On a CPU-only cluster, this pipeline becomes impractical at Tahoe-100M scale: covariate regression across tens of millions of cells and dozens of covariates is both memory-intensive and slow, and large gene–gene coexpression accumulation quickly turns into a multi-node, multi-hour (often multi-day) exercise once you include end-to-end iteration and I/O. NVIDIA’s GPU ecosystem—CUDA-X DS-enabled single-cell preprocessing, cuBLAS-accelerated linear algebra, and PyTorch DDP with NCCL for training—compresses that workload into a single-node run where the compute-heavy stages complete in minutes and the full pipeline lands in roughly an hour on bare metal.

APPROACH

TX1-CD is a causal decoder trained on top of TX1 single-cell embeddings, designed to separate what a drug perturbs directly from what changes downstream through regulatory propagation. Tahoe-100M cells are first embedded with TX1 to obtain 2,560-dimensional per-cell representations, and we compute a gene-gene coexpression matrix across all ~100 million cells using a CUDA-X DS-based pipeline, Figure 1A. To keep inference tractable we sparsify this into a candidate-edge mask, we retain the top 100 coexpression partners per gene, we add curated transcription factor-target edges from CollecTRI, DoRothEA, and TRRUST, and we obtain roughly 1.16 million candidate edges, 0.58% density, over 14,142 genes, chosen to balance biological coverage and tractability, we take the union of curated regulator sets, transcription factors, RNA-binding proteins, cancer hallmark genes, and we augment them with the top 10,000 highly variable genes identified by per-plate consensus.

Training is two-stage, we first fit a basal decoder on DMSO cells to predict expression from cell context alone, then we freeze it to anchor the unperturbed baseline. We then introduce the causal module, each drug is represented by a sparse direct-effect vector dp, constrained to known target genes and their local neighborhoods, and we learn a directed GRN W under the candidate-edge mask with a low-rank parameterization. Drug effects are propagated through W with a truncated Neumann-series approximation, capturing multi-step cascades, and the final prediction is the frozen basal output plus the propagated drug component. We regularize W for interpretability and stability using edge dropout, delayed L1 sparsity weighted by prior confidence so curated edges are penalized less than novel ones, and spectral-radius clamping, ρ≤0.95, yielding a directed GRN over 14,142 genes.

Tahoe-100M cells are embedded with TX1 and paired with a CUDA-X DS coexpression prior, TX1-CD is trained in two stages, a frozen basal decoder fit on DMSO cells, then a causal module that learns a sparse directed GRN and propagates drug direct effects through the network to model cascades.

KEY RESULT

The learned GRN captures biological structure that raw coexpression does not

Our expectation is that the learned directed GRN captures structure beyond generic coexpression. To test this assertion, we evaluated the learned weights of GRN against four independent external references, STRING physical interactions, EBI Complex Portal and hu.MAP2 protein complexes, and Reactome pathway co-membership, scoring GRN and raw coexpression matrix on the exact same ~1.16M candidate edges defined by our mask. In this controlled hypothesis space, coexpression provides limited ranking signal, whereas TX1-CD consistently prioritizes edges supported by these orthogonal functional benchmarks, improving AUROC from 0.42–0.58 to 0.74–0.83, Fig. 2A. The learned network also exhibits the expected directionality, known transcription factors are enriched among the top out-degree hubs, and annotated drug targets are enriched among the top in-degree nodes, Fig. 2B. Together, these results suggest TX1-CD is not merely reweighting correlation, it learns a directed regulatory structure that aligns with known biology and places regulators and targets into distinct topological roles.

(A) ROC curves comparing edge rankings from raw coexpression and learned weights GRN against four independent references, evaluated on the same ~1.16M candidate edges. (B) Hub enrichment showing over-representation of known transcription factors among top out-degree nodes and annotated drug targets among top in-degree nodes.

OUTLOOK

What Comes Next

As we continue to exponentially grow Tahoe’s datasets with new perturbations and cellular contexts, we will scale TX1-CD training accordingly, longer runs on larger, more diverse interventional corpora are the most direct way to sharpen the separation between direct effects and downstream cascades, and to expand the regulatory resolution the model can support. To make that scaling practical, we are moving more of the workflow onto next-generation NVIDIA GPUs, including B200, and pushing both preprocessing and model development into a fully GPU-native, multi-node regime. On the preprocessing side this means deeper adoption of CUDA-X Data Science libraries, as well as rapids-singlecell, and on the modeling side it means scaling training efficiently through distributed sharding, for example nvFSDP-style approaches, so that throughput increases without host-device movement or memory fragmentation becoming the limiting constraint.

With that foundation, the goal is not simply a larger network, it is a more useful one for mechanism. TX1-CD learns a directed topology with the expected roles, high out-degree hubs are enriched for transcription factors, 46% of the top 50, p = 10⁻²⁸, while high in-degree nodes are enriched for known drug targets, 20%, p = 10⁻¹⁰, and these hub identities emerge from perturbation-driven signal propagation rather than being hard-coded by priors. More importantly, for each compound the model decomposes the response into a proximal signature and a propagated cascade, which lets us trace putative mechanism as paths through regulators and bottlenecks, prioritize drug combinations when distinct compounds converge on shared downstream control nodes, and flag high-confidence edges missing from existing references as testable hypotheses. As the GRN expands, it can be directly overlaid with curated gene sets and regulatory modules (e.g., Gene Ontology), revealing regulatory connections around canonical pathways such as MAPK or JAK–STAT that are invisible in traditional pathway maps. These connections point to new nodes and targets for therapeutic intervention.

Finally, models like TX1 and TX1-CD are intended to be foundations, not endpoints. Our vision is that autonomous research agents can reason on top of them to simulate and stress-test mechanistic hypotheses about drug response, tracing candidate cascades, comparing them across contexts, and proposing focused follow-up perturbations that would most clearly validate or falsify the explanation.

Built with PyTorch · NVIDIA CUDA-X / BioNeMo Platform · rapids-singlecell (scverse) · CuPy (Preferred Networks) · NCCL · Dask-CUDA · RMM

TX1 Foundation Model · Tahoe-100M Atlas

This reframing suggests a powerful and underexplored way to evaluate drugs. Rather than asking only whether a compound kills a diseased cell or inhibits a specific target, we can ask a more fundamental question: Does this drug shift the cell’s transcriptomic identity back toward normal?

This is the principle of cell-state reversal. If you can measure what makes a diseased cell different from its healthy counterpart at the level of gene expression, and you can measure what a drug does to that same cell’s transcriptome, then you can score every drug by how effectively it closes the gap. A drug that reverses the disease signature (pushing expression back toward the healthy reference) is, in a precise and measurable sense, restoring the cell toward its normal state.

The concept is general: it applies anywhere you have a well-defined disease transcriptome and a compendium of drug-induced expression changes. But cancer is a particularly compelling place to start. Tumors are driven by well-characterized genetic lesions, i.e., mutations in oncogenes like KRAS, BRAF, and PIK3CA, that rewire gene expression in ways that have been extensively profiled in patients. And cancer is far more than uncontrolled proliferation: it involves aberrant differentiation, metabolic reprogramming, disrupted cell-cell signaling, and immune evasion. A drug that partially normalizes any of these dimensions might never register as a hit in a classical viability screen, yet could be profoundly relevant in a patient.

What’s been missing is the data infrastructure to test this idea at scale. You need a dense compendium of drug-induced transcriptomic responses across many cell lines or other cancer models and patient-derived reference profiles to define the “target”, the normal state you’re trying to recover.

Tahoe 100M provides the first two at a scale that didn’t previously exist: 100 million single-cell transcriptomic measurements spanning 60,000 drug–cell line experiments. Here, we apply cell-state reversal to colorectal cancer as a proof of concept — and show that the results are remarkably consistent with known biology, while surfacing signals that go beyond what viability assays alone can reveal.

The Logic of Cell-State Reversal

The dominant paradigm in cancer drug screening is phenotypic viability: treat cells, measure death. This approach gave rise to most classical chemotherapies and, combined with the molecular revolution in oncology, continues to power targeted drug development. Viability screens are valuable; they tell you which cellular features confer drug sensitivity.

But they capture only one axis of a drug’s effect. In vitro, this limitation is especially acute: the simplified microenvironment lacks immune-mediated killing, stromal interactions, and the metabolic context of a real tumor. A compound that restores differentiation programs, rebalances metabolism, or re-engages normal signaling architecture won’t necessarily kill cells in a dish, but it might reshape the disease in a patient.

Cell-state reversal offers a complementary lens. Instead of asking “does this drug kill the cell?”, we ask: does this drug make the cell’s gene expression look less like a tumor and more like normal tissue? This reframing opens the aperture to the full spectrum of cancer hallmarks, not just proliferation and survival, but differentiation, metabolism, and intercellular communication, and may ultimately enhance clinical translatability.

The analytical framework is conceptually simple:

First, for each drug–cell line pair in Tahoe 100M, we compute a differential expression profile (using DESeq2, relative to plate-matched DMSO controls). This captures what the drug does to the cell’s transcriptome.

Second, from patient-derived data, we compute a tumor-versus-normal differential expression profile for the relevant cancer type, defining the transcriptomic “distance” between diseased and healthy tissue.

Third, we correlate the two profiles. A strong negative Pearson correlation means the drug is reversing the cancer expression pattern, pushing the transcriptome back toward normal.

Validation: The Framework Recovers Known Biology

We applied this framework to colorectal cancer (CRC), using patient-derived epithelial cell profiles from Joanito et al. (Nature Genetics, 2022) to define tumor–normal differences for the iCMS2 and iCMS3 transcriptional subtypes. Drug-induced profiles were computed across nine CRC cell lines in Tahoe 100M.

The first test is whether drugs with known activity in CRC show stronger cancer-state reversal than the background. We defined positive control pairs: BRAF inhibitors (dabrafenib, encorafenib, vemurafenib) in BRAF-V600E cell lines, the (K)RAS inhibitor RMC-6236 in KRAS-mutant cell lines, and the multi-kinase inhibitor regorafenib across all CRC lines. RMC-6236 is not yet approved for CRC, but has shown promising clinical results in KRAS-mutated epithelial cancers.

The result was unambiguous: positive control pairs were strongly enriched in negative correlation values (mean Pearson coefficient –0.10 vs. –0.04 for all other pairs; Mann–Whitney U p = 3.17 × 10^⁻²⁰; Figure 1). The framework recovers known drug–cancer relationships from first principles, using only transcriptomic data.

Figure 1: Distribution of Pearson correlation coefficients across all drug–cell line pairs, with positive controls (red stars) clustered in the left tail.

A closer look at RMC-6236 reinforces the signal. For the iCMS3 subtype, which is enriched for KRAS gain-of-function mutations, we observed greater cancer-state reversal in KRAS-mutant cell lines compared to wild-type lines, consistently across all three doses tested (Figure 2). The drug’s transcriptomic effect is genotype-specific, exactly as the biology predicts.

Figure 2: RMC-6236 reversal scores across CRC cell lines in iCMS3, stratified by KRAS mutation status and dose.

Mechanism of Action Reveals a Hierarchy of Reversal

When we aggregate reversal scores by drug mechanism of action (MOA), a coherent picture emerges (Figure 3).

MEK, RAF, JAK/STAT, and PI3K/AKT inhibitors show the strongest reversal of the cancer expression signature. This tracks directly with CRC genetics: gain-of-function mutations in KRAS, BRAF, and NRAS are among the most frequent drivers of colorectal cancer, and activating mutations in PIK3CA are similarly common. Several drugs within these MOA classes are already part of the standard of care or under active clinical investigation.

EGFR/ERBB inhibitors show more modest but above-average reversal. This is consistent with the fact that while EGFR itself is less frequently mutated or overexpressed in CRC, EGFR-targeted agents (cetuximab, panitumumab) are key components of approved combination regimens — their clinical benefit may be more context-dependent than the on-target biology alone would suggest.

Chemotherapies tell an interesting story. As a class, they are expected to shift cells to a different state rather than reverting them to normal — consistent with their mechanism as broadly cytotoxic agents. But within chemotherapies, DNA synthesis and repair inhibitors produce greater reversal than microtubule inhibitors. This distinction mirrors clinical practice: DNA-targeting agents (5-FU, oxaliplatin) form the backbone of CRC standard-of-care regimens like FOLFOX and FOLFOXIRI, while microtubule inhibitors do not.

JAK/STAT inhibitors also score highly, reflecting the frequent activation of IL-6/JAK/STAT3 signaling in colorectal tumors. Clinical translation of JAK/STAT inhibition in CRC has been disappointing so far, but this MOA remains under active investigation, and the transcriptomic data suggests the biological rationale is sound.

Figure 3: Reversal scores aggregated by mechanism of action, with targeted therapies in green, chemotherapies in yellow, and unclear MOA in gray.

The Top 20: From Expected Hits to Unexpected Insights

Zooming in on the top 20 drugs prioritized for the iCMS2 subtype (Figure 4, showing the dose with the strongest reversal), the list reads like a who’s-who of CRC-relevant pharmacology — with a few surprises.

The expected hits are well represented: MEK inhibitors (trametinib, cobimetinib, binimetinib), BRAF-V600E inhibitors (encorafenib, vemurafenib), the KRAS inhibitor RMC-6236, and the PI3K/mTOR inhibitor bimiralisib. Topoisomerase-II inhibitors idarubicin and mitoxantrone appear, as does the JAK/STAT inhibitor tofacitinib, the MET inhibitor capmatinib, and the nuclear export inhibitor selinexor. For the case of MEK inhibitors: while they have historically not been particularly successful in advanced or metastatic CRC, a recent clinical trial sponsored by Recursion showed promising results for FAP, a genetic disorder characterized by disseminated early-stage adenomas.

One compound that merits particular attention is elimusertib, nominally an ATR inhibitor. In a previous analysis of Tahoe 100M, we reported that elimusertib’s transcriptomic signature more closely resembles that of MEK inhibitors — a finding that is independently reinforced here by its ranking among the top reversal compounds alongside bona fide MEK inhibitors.

But perhaps the most intriguing signal comes from sodium salicylate. A recent placebo-controlled trial (Martling et al., New England Journal of Medicine, 2025) demonstrated that aspirin (acetylsalicylic acid) significantly increased 3-year disease-free survival (HR ≈ 0.55) among stage I–III CRC patients harboring PI3K/AKT pathway-activating mutations. In Tahoe 100M, however, salicylate and salicylic acid produce stronger cancer signature reversal than aspirin itself. This is a mechanistically meaningful distinction: aspirin’s additional acetyl group is what enables cyclooxygenase inhibition, so a stronger signal from the non-acetylated form suggests a COX-independent mechanism driving the transcriptomic reversal. The kind of observation that is only possible when you have the data density to compare closely related chemical structures across the same biological models.

Figure 4: Heatmap of top 20 drugs for iCMS2, showing Pearson correlation (reversal score) across nine CRC cell lines. Positive control pairs are highlighted in red.

What This Means — and Where It Goes

Across colorectal cancer models, cell-state reversal consistently prioritizes compounds that align with known biology: drugs targeting the oncogenic pathways most frequently activated in CRC rise to the top, while broadly cytotoxic agents that shift rather than revert the cancer state rank lower. The framework doesn’t just recover textbook pharmacology — it draws distinctions within drug classes (DNA synthesis inhibitors vs. microtubule inhibitors, salicylate vs. aspirin) that are consistent with clinical evidence.

This is what becomes possible when you have the data density of Tahoe 100M: 100 million cells, 60,000 experiments, hundreds of compounds profiled across genetically characterized cancer models. Cell-state reversal is not a replacement for viability screens — it’s a complementary lens that captures dimensions of drug activity that viability alone cannot see. As large-scale perturbational transcriptomic resources continue to expand, this approach may become a foundational tool for how we discover, repurpose, and personalize the next generation of therapies. And the exciting applications will for sure go beyond cancer, in disease areas where killing the cell is not the objective; but rather, rewiring it back to normal is what we are aiming for.

Tahoe 100M changes the equation. This atlas offers millions of transcriptional profiles at single-cell resolution across thousands of compounds, capturing drug effects with unprecedented depth. When combined with CRISPR-based CROP-seq, drug and genetic perturbations can finally be compared head-to-head — creating a unified map for target deconvolution, pathway mapping, and phenotypic drug design.

We combined:

• Tahoe-100M drug atlas, a massive single-cell reference of ~400 compounds tested across 50 cancer models, each at three doses. This scale captures a spectrum of dose-dependent drug responses and provides a uniquely rich landscape of transcriptional perturbations.

• Myllia’s CROP-seq dataset, a focused CRISPR perturbation dataset in which 218 genes were targeted in A549 cells and transcriptomic fingerprints were recorded by single-cell RNA sequencing.

Both were converted into gene-set scores (MSigDB collections, scored with Vision), reducing dimensionality and pooling genes into functional units. Differential scores were computed by comparing drug treatment to matched DMSO (Tahoe-100M), or gene perturbations to non-targeting guides (Myllia’s CROP-seq). Pairwise similarities between drug-induced and gene-induced differential signatures produced a drug–gene similarity matrix (Figure 1).

As a first test, we applied the framework to two well-characterized inhibitors:

• Trametinib (MEK1/2 inhibitor)

• RMC-6236 (pan-RAS inhibitor)

Both landed where they should: the top-ranked genetic perturbations mapped to the RAS/RAF/MEK/ERK pathway, consistent with their known mechanisms of action (Figure 2).

Importantly, MEK1/2 and KRAS themselves were not included in the CRISPR knockout set. Yet, despite their absence, the framework still identified the correct pathway-level convergence through related nodes in the cascade – demonstrating that the integration can resolve mechanisms even when direct targets aren’t explicitly perturbed.

Then came the surprise. Elimusertib, long reported as an ATR inhibitor, did not resemble DNA damage pathway perturbations at all. Instead, its transcriptional profile clustered tightly with MEK inhibitors (Figure 3).

To test this, we treated two cancer cell lines with Elimusertib and measured ERK phosphorylation (pERK). The result: Elimusertib suppressed pERK, phenocopying Trametinib and Binimetinib.

This was a striking twist. A compound presumed to act via ATR was revealed to moonlight as a MAPK pathway suppressor.

And it raises a broader question: what other compounds hide secondary mechanisms, silently shaping efficacy or toxicity? Traditionally, testing every possible mechanism is infeasible; the space is too vast. This integrative approach makes it possible to scan hundreds of pathways at once, rapidly surfacing the most likely hidden effects.

To generalize further, we assessed whether drugs and CRISPR perturbations with functionally related targets tend to elicit similar transcriptional effects. Using the HumanNetV3 functional gene network, we calculated the shortest paths from CRISPR-perturbed genes to drugs producing a similar gene expression response.

The average observed path length was consistently much shorter than expected by chance (compared to random drug sampling), showing that transcriptional similarity reflects true functional proximity (Figure 4).

This work demonstrates a new way of seeing drug action:

1. Target deconvolution – drug effects can be mapped to specific pathways by reference to genetic perturbations.

2. Unmasking moonlighting compounds – Elimusertib shows how unexpected mechanisms surface rapidly.

3. Network convergence – transcriptional similarities aren’t random; they reflect underlying biological relationships.

4. Phenocopy design – the next frontier is designing compounds that deliberately recreate specific transcriptional states.

Taken together, these findings suggest that the integration of drug and genetic perturbations at scale is more than a technical advance. It’s a conceptual shift: from thinking of drugs as single-target agents, to thinking of them as programmable phenotypic levers capable of reproducing desired cellular states.

Tahoe‑100M (for the data, see Hugging Face) is the largest interventional single-cell dataset to date, comprising over 100 million transcriptomes that capture responses to 1,100 chemical perturbations across 50 cancer cell lines. The scale and richness of Tahoe-100M have already enabled new discoveries—for example, recent work used the dataset to identify compounds that enhance tumor visibility by upregulating MHC-I expression, a result that would have required extensive experimental screening in the past.

With datasets of this scale, AI and deep learning offer new possibilities for experimental biology. In the near term, machine learning models trained on interventional single-cell data can help researchers prioritize experiments, predict cellular responses to untested conditions, and reveal subtle patterns that might otherwise go unnoticed. Looking further ahead, these approaches hold promise for advancing our understanding of complex biological systems and for guiding the development of novel therapeutic strategies.

Why We Built scDataset

Despite the promise of large-scale interventional data, the practical reality of working with these datasets has been shaped by data storage and access constraints. The standard format for storing and sharing single-cell data is AnnData, which allows rich annotation and efficient storage of large, sparse matrices. However, most analysis tools and workflows for AnnData were designed for interactive exploration or modestly sized datasets—not for training deep learning models directly on tens or hundreds of millions of cells.

Traditional approaches often require loading the entire dataset into memory or converting to dense formats, quickly becoming infeasible at Tahoe-100M scale. As a result, the step from data collection to large-scale model training has remained a major bottleneck.

In our recent paper, we introduce scDataset, a PyTorch IterableDataset designed for efficient, scalable training on large single-cell omics datasets stored in the AnnData format. Unlike existing solutions, scDataset enables fast, memory-efficient, and shuffled data loading directly from disk, making it possible to train deep learning models on datasets as large as Tahoe-100M using standard hardware.

scDataset solves the AnnData bottleneck by combining block sampling and batched fetching, balancing randomness and I/O efficiency. In our benchmarks on the Tahoe-100M dataset, scDataset achieved up to a 48× speed-up over AnnLoader and substantial improvements over other popular data loading solutions (See below for a comparison with other loaders).

Getting Started with scDataset

To help the community get up and running, we've prepared a quickstart notebook that demonstrates how to use scDataset to efficiently load the Tahoe-100M dataset and train a simple linear classifier in PyTorch. 

The notebook walks through:

- Loading Tahoe-100M .h5ad files

- Performing a stratified train/test split

- Wrapping the data with scDataset for efficient loading

- Training and evaluating a linear classifier for cell line recognition

You'll also find installation instructions, API documentation, and additional examples in the scDataset GitHub repository. If you have questions or feedback, feel free to reach out or open an issue on GitHub.

The Virtual Cell Challenge

The Arc Virtual Cell Challenge is the perfect venue to test the speed and scalability of scDataset in real-world applications. This open competition challenges the community to build predictive models of cellular responses using massive single-cell datasets. With scDataset, participants can focus on developing innovative models and analyses, confident that data loading will not be a limiting factor.

Whether you're new to large-scale single-cell analysis or aiming to accelerate your deep learning workflows, scDataset provides a practical and flexible foundation. We look forward to seeing what the community builds with scDataset and Tahoe-100M!

Successfully overcoming this evasion mechanism could revolutionize cancer treatment by significantly enhancing the effectiveness of immune checkpoint inhibitors (ICIs), which depend on robust tumor antigen presentation. Addressing this fundamental challenge—identifying compounds that effectively upregulate MHC-I expression—can dramatically enhance therapeutic strategies.

Figure 1. Schematic illustrating MHC-I upregulation and effects on immune cell engagement.

We took a crack at that by digging into our Tahoe-100M, the world's largest drug-perturbed, single-cell transcriptomic dataset.

With an unprecedented scale, Tahoe-100M maps how various drug molecules impact different pathways, including immune-related pathways in diverse contexts (aka cancer models). We systematically explored an extensive array of these immune-related pathways to uncover drugs capable of modulating critical signaling mechanisms, including IFN-α and MHC-I. We designed a computational pipeline to evaluate hundreds of compounds simultaneously, allowing us to rapidly prioritize compounds based on their impact on these key immune pathways (Figure 2).

Figure 2. Computational pipeline illustrating method of prioritization of compounds for a given pathway of interest.

We found quite a few compounds that can effectively upregulate MHC-I. Some of these compounds were previously known immunomodulators (such as MEKi inhibitors). This provides valuable validation. But more importantly, it identified numerous new candidates (e.g., Elimusertib, an ATR kinase inhibitor, and TAK-901 , an Aurora kinase inhibitor) with previously unknown potential to stimulate immune recognition via MHC-I upregulation (Figures 3 and 4).

Figure 3. Tahoe-100M data showing compounds, grouped by MOA class, ranked by their effect on MHC-I geneset score upregulation.

Figure 4. Scatterplot of compounds by their perturbation effect on IFN-a and MHC-I geneset scores. Compounds are colored based on their MOA.

To get to this point, we only leveraged the data that is already in Tahoe-100M and the analysis pipeline we built on top of it; the question is whether orthogonal experimental assays would validate our findings. In other words, do independent experiments confirm that the newly identified compounds have the ability to upregulate MHC-I? The answer is “Indeed, they do”!

When we treated cells from CFPAC-1 cancer cell line with some of the new compounds we had identified, we saw marked increases in total MHC-I protein expression, as confirmed by western blotting (Figure 5). Flow cytometry analyses further validated that most selected compounds significantly enhanced MHC-I surface expression, underscoring their therapeutic potential to make tumors more visible and vulnerable to the immune system (Figure 6).

Figure 5. Experimental validation of total protein expression change after compound treatment. Western blot of CFPAC-1 cells after drug treatment and blotting for MHC-I shows upregulation for all compounds after drug treatment.

Figure 6. Experimental validation of cell surface MHC-I levels after compound treatment. Flow cytometry of CFPAC-1 cells with fluorescently labeled anti-MHCI antibody after drug treatment shows upregulation for some compounds after drug treatment. Normalized to vehicle treated (DMSO) cells.

These findings illustrate the unique capacity of the Tahoe-100M dataset to systematically address targeted biological questions—such as identifying drugs that specifically enhance MHC-I expression—at an unprecedented scale.

We have been saying consistently that gigascale, perturbative, single-cell maps like Tahoe-100M are a new modality of data for driving discovery. This example is a demonstration of that. Tahoe-100M allows cancer biologists to rapidly query their favorite biological pathways. It facilitates the discovery of novel therapeutic strategies and in this case, opens a new path for the development of precision-driven combinations aimed at significantly improving responses to ICIs in cancer treatment.

Tahoe-100M Manuscript

Tahoe-100M Dataset Download

Today we’re launching The Tahoe Blog. In the same open-source spirit that led us to release Tahoe-100M—our gigascale, perturbative single-cell dataset—we’re now sharing many of the discoveries it has enabled.

Tahoe-100M isn’t just a large-scale dataset that enables the next generation of AI models in biology. It’s a new modality of data for discovery. Instead of designing a new experiment for every question, you can ask hundreds of questions—across thousands of perturbations and cancer models—with high-content, single-cell data already in hand.

This changes the game. You can answer questions you weren’t even asking. Discoveries emerge before the hypothesis. It’s part of a larger shift: from hypothesis-driven to hypothesis-generating science, from narrow to panoramic views of biology.

Think of it like the transition from Recurrent Neural Networks (RNNs) to transformers in AI. On the surface, transformers might look like scaled-up RNNs—but their ability to operate across all contexts at once unlocked entirely new capabilities.

Tahoe-100M does something similar for biology. Measuring how a handful of cancer models respond to a drug or two is common. But scaling that across thousands of models and drugs—with the resolution of single-cell data—reveals relationships and mechanisms you’d otherwise miss. That’s what makes this data more than the sum of its parts.

In this blog, we’ll highlight discoveries that surprised even us—insights hiding in plain sight, enabled only by this new scale and structure of data. We’re sharing them in hopes they’ll spark new questions, inspire new therapeutic directions, and ultimately help patients.

We begin with one such finding: We discovered previously unknown mechanisms associated with Saquinavir; mechanisms that could have predicted its cardiovascular toxicity, which were later detected during its clinical usage. The drug, which had been developed by Roche, was later discontinued in the US due to these toxicities. 

Read our first blog post and stay tuned for new ones once every couple of weeks. You could also subscribe to get notified when we publish new ones.

Enjoy!

P.S. Reach out if you would like to contribute as a guest.

Drugs are polypharmacological. During discovery and development, we often treat them as chemical agents that are optimized to interact with only the desired target. But in reality, they do a lot more; they interact with other parts of the biological system that is the patient.

They bind to other proteins and perturb other, unintended landmark genesets.

Many of these perturbations remain unknown for decades, even after the drug is approved for use in patients. Some of them are desirable; they are a lucky contributor to the drug action. And some are deleterious off-target effects.

Overall, these perturbations add up to become the overall phenotype of the drug and how effective it is. What if, instead of designing a drug and hoping that the polypharmacology works out, we were able to deconvolve and account for the effect of that drug on many key landmark genesets throughout discovery and development?

Gigaset, perturbative, single-cell maps like Tahoe-100M make this a reality. They provide a high-precision reference of how thousands of compounds perturb landmark genesets at single cell resolution. To deconvolve the polypharmacology of a given drug, all you need to do is to compare the post-perturbation transcriptional profile of that drug with these giant maps and see what other perturbations it mimics.

That is exactly what we did for a poorly characterized, but consequential drug: Saquinavir, a once approved HIV protease inhibitor, which was later discontinued.

By comparing transcriptional responses of Saquinavir (among many other drugs) with unknown mechanisms to well-characterized reference drugs, we created a detailed drug-drug similarity map. This approach leverages the rich transcriptional data from our Mosaic platform as a lens to quickly pinpoint drug similarities, elucidating their cellular effects (Fig 1a).

(If you want to know more about how we created these similarity maps, take a look at the postscript in the bottom of this post)

Transcriptional similarity in Mosaic data allows for drug::drug mechanism of action mapping. A) Conceptual heatmap illustrating clustering of transcriptional profiles from known and unknown drugs using Mosaic data as a way of inferring MOAs. B) Ranked plot of Mosaic transcriptional similarity scores between Saquinavir and a library of compounds. Norepinephrine and Vilanterol rank highly, indicating strong transcriptional resemblance and possible mechanistic overlap.

And that’s where surprises started to emerge from the data: Saquinavir, rather unexpectedly, mimics classic adrenoceptor agonists like Norepinephrine and Vilanterol. Classically, adrenoceptor-agnonists like Norepinephrine activate adrenergic signaling, leading to hallmark effects in cells such as increased intracellular levels of cAMP and phosphorylation of CREB (pCREB). Traditionally, Norepinephrine is clinically utilized to manage acute hypotension, highlighting its significant cardiovascular activity. Saquinavir’s unexpected resemblance to these agonists (Fig 1b) suggested potential unintended effects on cardiovascular function.

We discovered this in our Tahoe-100M similarity maps; does it stand up when tested in traditional experimental assays? The answer is yes! We treated a representative cell line with Saquinavir and, separately, with adrenoceptor agonists. We confirmed that after treatment with drugs, Saquinavir showed the same levels of upregulation of pCREB as traditional adrenoceptor-agonists Vilanterol and Norepinephrine (Fig 2a,b).

Saquinavir induces a transcriptional program similar to adrenoceptor agonists. A) Western blot showing protein expression of phospho-CREB, CREB, and GAPDH following 24hr treatment B) Cartoon summary of mechanism of action (MoA)

Things got even more exciting when we dug into the clinical history of Saquinavir: Saquinavir has been associated clinically with QT prolongation and cardiac arrhythmias, prompting an FDA safety review in 2016 (Fig 3). While QT prolongation typically results from direct ion channel inhibition, our data-driven discovery uncovered an additional mechanistic similarity between Saquinavir and adrenoceptor agonists, potentially contributing to broader cardiovascular risks through adrenergic pathway activation. This provides novel mechanistic context to documented clinical safety concerns and expands our understanding of the cardiovascular liabilities associated with Saquinavir.

FDA Drug Safety Communication on Saquinavir highlighting unknown mechanisms underlying cardiovascular adverse effects. Source.

This is remarkable! Access to Tahoe-100M allowed us to uncover off-target effects that were discovered after three decades of widespread clinical usage. Imagine what else we are missing in the pipeline of drugs that are being designed today and making their way through the clinic in the next decade.

What makes this even more remarkable is two important observations:

• Despite stark differences in chemical structure, Norepinephrine, Vilanterol, and Saquinavir converge functionally through a shared transcriptional mechanism of action (Fig 4). Such revelations highlight the transformative power of transcriptional similarity in reshaping our understanding of drug pharmacology and safety.

• Interestingly, Saquinavir was specifically designed as an antiviral agent to inhibit HIV protease, targeting the viral lifecycle rather than human biology. Because its development prioritized antiviral efficacy over broader human cellular responses, the unexpected adrenoceptor agonism and cardiovascular risks remained undetected.

This underscores the power of Tahoe-100M (and datasets with similar content and scale) as a new modality of data to guide discovery and design of drugs by rapidly uncovering drug effects across diverse biological pathways—revealing critical mechanistic and safety insights as well as therapeutic opportunities that traditional screens would miss.

Chemical structures of Norepinephrine, Vilanterol, and Saquinavir.

Tahoe-100M exemplifies how new modalities and scales of pharmacogenomic data advance therapeutic innovation and inform patient safety. We hope this example inspires broader applications of our data, continually driving progress toward safer and more effective treatments. Until next time.

Tahoe-100M Manuscript

Tahoe-100M Dataset Download

Postscript

We mentioned that we created drug-drug similarity maps from Tahoe-100M. Here is how we did it:

Workflow for generating transcriptional comparisons to infer drug mechanism of action. Single-cell gene expression data are first converted into pseudo-bulk profiles from AnnData objects. Gene set enrichment is performed using Vision scoring, followed by differential gene set analysis to quantify drug-specific effects. These differential scores are then used to compute drug–drug similarity, enabling comparison of drugs with unknown MoA to those with known mechanisms.