Noetik’s bet: learn which patients a cancer drug can help

Eighteen months before the first model

Two slides took roughly two weeks to process. Ron Alfa describes opening a lab, hiring staff, buying instruments, sourcing human tumors, and building a processing pipeline before knowing whether the resulting data would support useful models. About eighteen months later, the team could attempt training. Dan Bear says single-cell transformers offered some precedent, but scarce spatial datasets left substantial custom modeling work.

That investment follows Noetik’s founding hypothesis: poor patient selection explains much of cancer-drug failure. Alfa estimates that roughly 90–95% of cancer drugs fail in the clinic and argues that drug development has improved faster than the ability to identify whom a treatment will help. The failure rate is a broad estimate; the proposed explanation is his causal thesis.

Alfa describes two uses for models trained on patient biology: - Discover targets from humans: work backward from patient biology toward targets and molecules, an approach he calls reverse translation. - Identify likely responders: analyze patients from Phase II or III trials to find biology associated with response and inform subsequent enrollment. Noetik was already running inference on trial biopsies, but Alfa says those results had not been shared. Rescuing an unsuccessful trial remained a goal.

The difficulty starts with how cancers are classified and studied. Alfa suggests that a familiar lung-cancer subtype might contain several therapeutically distinct groups. He also criticizes long-established, immortalized cell lines for abnormal chromosomes and expression patterns, and mouse experiments that place tumor cells outside their original tissue context. His concern is that these experimental systems can leave clinical teams without a reliable map to human responders.

Consider Alfa’s hypothetical early-phase trial of fifty patients. Investigators must explore dose, safety, and efficacy while still uncertain which cancer subtypes matter. If a drug works in a narrow subgroup, the trial may contain too few responders to reveal it. The molecule can be canceled before anyone understands whom it helps.

Noetik wants therapeutically relevant subtypes learned through self-supervision, allowing distinctions to emerge from genetic changes, immune activity, or other measured biology. Bear argues that conventional biomarkers—measurements such as a mutation, protein stain, or gene-expression signature—favor simplicity and may miss meaningful variation. Richer subtypes predicting response more strongly remains the hypothesis.

Design the dataset to distinguish biology from processing

Noetik generates its data end to end because experimental design determines what a model can learn. The team points to deliberately accumulated protein-structure data and curated ImageNet images as precedents for building coherent datasets. Scale matters, but observations must also be collected around the relationships future models need to learn.

An interviewer asks whether biology’s shared physical processes might allow generalization with less data. The response separates cancer coverage from general biology: Noetik reports thousands of patients across a few major cancer subtypes, while other diseases, healthy tissues, species, and spatial relationships among genes and proteins make up a much larger domain.

Images offer information density and collection economics: one slide can contain material from many patients. They also carry batch effects, variation introduced by staining, laboratory processing, or the day a slide runs. A patient embedding—a learned representation of the patient’s data—could organize samples by staining batch instead of biology relevant to immunotherapy response.

Noetik samples each specimen dozens of times and builds arrays containing hundreds of randomized patient samples. Each patient appears on multiple arrays processed in different batches. Repeating patients across batches makes technical variation testable: downstream comparisons can distinguish patterns that follow patient biology from those that follow laboratory processing.

Connect tissue structure to cells and molecules

The dataset starts with tissue, where cells interact, and includes a modality already used in clinical practice. Four layers connect visible structure to cellular and molecular information: - H&E imaging: hematoxylin and eosin dyes create the pink-purple tissue contrast pathologists use to inspect cellular structures and classify tumors. - Protein imaging: antibody-based stains mark proteins associated with T cells, B cells, and other cell types, producing multichannel images of the tumor’s cellular environment. - Spatial transcriptomics: measurements locate RNA within those cells, connecting gene expression to tissue position. - Tumor genotyping: DNA measurements identify genomic alterations alongside the imaging and RNA data.

H&E gives a human reader limited ability to distinguish detailed immune-cell subtypes. Protein markers add that cellular information, and RNA measurements connect it to molecular activity. The spatial RNA assay binds probes to RNA and cycles through detection and amplification; one assay can take weeks.

The engineering analogy is an image with approximately 20,000 molecular channels instead of three color channels. The speakers describe assays capable of detecting nearly all of roughly that many genes in one sample. The challenge is to extract differences between patients from this high-dimensional spatial information, then determine which differences matter for treatment response.

Virtual cells for bounded biological predictions

The discussion distinguishes two ambitions for a virtual cell: - Comprehensive biochemical simulation: predict the intracellular reactions following an input. One speaker argues that the necessary data modalities are not yet available. - Useful input-output prediction: estimate what genes or proteins a cell would express in a specified context, sufficiently well to guide drug-development experiments. Noetik describes its initial models in the second sense, using cellular context to connect cell biology to patient biology.

The team has deliberately done little integration of electronic health records. It wants models to learn genes, proteins, cells, and tissue directly through self-supervision. Its hypothesis is that clinical text reflects what doctors knew to document, while much of the therapeutically predictive biology remains unrecorded or unrecognized.

The decision depends on what is known about a drug: - Before treatment data exist: simulations could identify patient groups in which the drug’s target matters. Redirecting attention from lung to ovarian cancer is an illustrative possibility. - After some patients have responded: investigators can ask whether responders concentrate in a cluster of the learned patient representation. That concentration generates a hypothesis for the next cohort.

A concrete simulation asks what a T cell would express inside a particular patient’s tumor environment. Another asks what happens when a target gene or protein is reduced: does the predicted response suggest stronger antitumor immune activity or less tumor growth? These readouts are candidate correlates of clinical benefit, which still need to be connected to treatment outcomes.

Train on rich measurements; infer from H&E

Noetik’s deployment strategy is to train on multimodal measurements and require only H&E at inference. The pathology image can come from a newly processed specimen or an archived trial, allowing retrospective analysis without repeating the expensive molecular assays. H&E’s widespread availability in oncology makes this asymmetry useful.

The models reportedly predict spatial gene expression as well as patient representations. Investigators can compare predicted expression in responder and nonresponder groups, including whether the responder group expresses the drug’s target. Predicted molecular differences help interpret patient clusters; they remain inferred values rather than direct measurements of the archived sample.

The companion article describes TARIO-2, an autoregressive transformer that predicts a spatial map of approximately 19,000 genes from H&E. That number describes output breadth; the written account gives no prediction accuracy, validation cohort, or clinical outcome measurement.

Noetik reports applying the H&E approach across collaborations involving different drug mechanisms, including an announced collaboration with Agenus. Using the same input to guide selection among marketed treatments is a possible extension.

The team identifies three differentiators: data scale, custom multimodal architectures, and interpretable predictions. It reports more than 100 million spatially resolved cells with transcriptomics paired with H&E and protein measurements. Models trained on 40% or 10% of the dataset reportedly performed substantially worse, especially on cancer types absent from training. No numerical performance comparison accompanies that scaling claim.

PerturbMap tests interventions in tissue

Noetik also collects experimental perturbation data. Its mouse platform, PerturbMap, tests genetic interventions in living tissue and provides an experimental check on simulated predictions about human tumors.

The procedure preserves each intervention’s identity: 1. Disable an individual gene in a cancer-cell population through a gene knockout. 2. Add a protein barcode: a combination of expressed protein tags identifies the knockout. 3. Introduce the population into mice alongside roughly a hundred other populations carrying different knockouts. 4. Measure the resulting tissue spatially: antibody labeling detects the barcode proteins, linking observed tumor biology to the intervention.

Mouse lungs can then contain hundreds of tumors with different biological profiles. Genetic perturbations can produce tumors lacking immune-cell infiltration, allowing researchers to investigate causal relationships behind profiles also seen in humans. Treating the mice adds a drug intervention. Fifty knockouts combined with fifty drugs across a panel of mice is a proposed expanded design, not a completed screen.

The computational connection is a human-trained model applied directly to mouse H&E, producing predictions expressed in human-gene terms. Matching structures between separate human and mouse latent spaces is another approach under investigation, but the concrete method described uses the human model on the mouse image.

Validation starts with known biology: - Immune-cold tumors: reducing tumor cells’ ability to present antigens produces tumors with immune cells absent from their vicinity. The team reports that human-trained models predict this phenotype from mouse tissue. - Shared pathway effects: a signaling pathway is a sequence of protein interactions that changes cellular behavior. Different gene knockouts in the same pathway reportedly yield similar model-inferred phenotypes, consistent with that relationship.

Recovering known controls builds confidence; new biology remains uncertain. The examples offer biological consistency checks, without a quantitative account of cross-species accuracy or evidence that a predicted response translates into patient benefit.

What a token means—and why tissue area matters

The earlier virtual-cell model uses masked autoencoding: divide data into chunks, hide some, and reconstruct them from the visible chunks. The recording transcribes that model’s name as OctoVC and a subsequent transformer as TARIO; those spellings are unverified.

A token’s meaning depends on the modality: - Spatial RNA: one gene’s expression at one location. - Protein imaging: an image patch for one protein at one location. - Histology: an image patch, as in a vision transformer.

Heavy masking makes local interpolation insufficient. In the speaker’s hypothetical example, hiding only about 10% of an image lets a model succeed by continuing nearby edges. Hiding much more forces it to learn broader structure. Noetik reports a similar effect in tissue data, where spatial correlations between proteins support reconstruction. An interviewer recalls approximately 99% masking, but the reply does not establish an exact operating rate.

The newer transformer uses autoregressive next-token prediction for spatial transcriptomics, an attempt to extend language-model-style scaling to this modality. The consequential reported result is conditional: larger models helped at longer contexts, where longer context means seeing more tissue area at once.

Two explanations remain plausible: - Broader spatial relationships: more tissue may reveal complicated nonlinear patterns in patient biology. - Better coverage of sparse signals: more context may capture low-expression genes that are nevertheless predictive. Experiments holding the amount of context constant while varying tissue area reportedly favored larger regions. That suggests spatial coverage contributes beyond context quantity alone, although no effect size or evaluation metric is given.

License the model; adapt it to pharmaceutical data

Noetik describes a GSK agreement announced in January: $50 million including an upfront payment and milestones, plus a separate annual model-licensing fee. The $50 million includes contingent payments; it is not cash already received. The licensed asset is its virtual-cell foundation model.

The agreement covers models already trained on lung and colon cancer. GSK can use them for internal simulations and therapeutic discovery, then fine-tune them on its own data. This gives accumulated trial pathology specimens a role in adapting a foundation model. The broader commercial hypothesis is demand for model access across a pharmaceutical pipeline, beyond collaborations tied to individual drug programs.

Large pharmaceutical data holdings can still be scattered and siloed. Noetik’s investment is in consistently processed measurements at enough scale to support meaningful architecture comparisons. The dataset must reveal whether an algorithmic idea improves performance, as well as contain enough examples to train it.

That requirement explains the long interval before the first model. The team estimates eighteen months to two years of collection before working trained models, with initial virtual-cell training around 2024. Tumors had to be prepared into arrays, assayed, aligned across modalities, and post-processed. Existing single-cell transformers supplied no established route through the spatial modeling problem.

Work backward from the prediction you need

The advice to biotech builders is to choose the ML problem before committing to a dataset. A participant recalls years of platform iteration at Recursion, learning which controls and experimental-design choices mattered. Reusing those practices at Noetik reduced some trial and error, though their value in a new setting still needed testing. An interesting dataset is not necessarily suitable for the intended prediction.

Measurement technology changes which problems are feasible. One participant recalls a week of manual work to examine roughly two genes in one sample while finishing a PhD in 2016. Five or six years later, machines could assay thousands of genes at once, although the assays remained expensive. Dataset design needs to account for both current capabilities and their improvement.

The financial uncertainty remains: a small pilot may fall below a useful scale, yet collecting enough biological data is expensive. The team argues that some domains require generating a dataset before knowing whether an approach works. An interviewer offers a useful qualification: individual protein structures can have immediate value, even without confidence in the models a future database might enable.

Noetik uses two experimental comparisons to investigate its requirements: - Vary data quantity: train on 10%, 40%, and 100% of the available data. - Withhold whole cancer types: exclude pancreatic or breast cancer, for example, to test generalization beyond training indications. The team speculates that a few hundred patients across a range of indications might support broad cancer generalization because some cancer and immune-cell biology is shared. That remains an expectation, not a demonstrated coverage threshold.

Patient-level predictions are the goal

Keytruda illustrates the size of the search. A speaker describes Merck exploring the drug across hundreds, possibly more than a thousand, trials in different populations. Patient embeddings could give such trials a narrower, biologically informed starting point, reducing uncertainty about where to test a mechanism.

This goal determines Noetik’s preferred abstraction. A fully mechanistic cell model assembled upward into tissues and patients could require many more modalities than gene expression. The team starts instead with variation among actual patients and asks which variation predicts drug response. Its bet is that interacting cells in functional tissue may support useful patient-level predictions without first resolving every subcellular mechanism.

A speaker draws an analogy to neuroscience: neural networks built from abstract linear and nonlinear units became, in that speaker’s account, highly predictive of neuronal responses to real-world stimuli. The lesson proposed for tissue modeling is to choose an abstraction around the prediction needed. It is a motivation for the strategy, rather than evidence that the same outcome will follow.

The closing invitation welcomes ML practitioners, including those without a biology background who are willing to learn the necessary biology. It preserves a distinction made during the discussion: understanding cancer biology and curing cancer are different achievements. Protein folding, binding prediction, literature-reading agents, and workflow automation address different parts of therapeutic development; success in one part does not complete the process.

Noetik’s chosen task is to turn treatments that help some patients into successful trials with better-defined cohorts. The team argues that reading existing literature cannot supply biology that has never been measured. New measurements, further ML research, and successful drug development remain necessary work—and places for engineers to contribute.