OCTO

Most drugs fail not because their molecular target was modeled badly, but because it was the wrong target, or the right target given to the wrong patients. What a target screen actually needs is an answer to "if this gene were turned up in this patient's tumor, what would happen to the immune cells around it?" — an experiment that cannot be run in a person. OCTO treats that as a simulation problem: train one generative model on enough spatially aligned tumor measurements that perturbing its inputs stands in for the experiment.

OCTO — the Oncology Counterfactual Therapeutics Oracle, also described by Noetik as a world model for cancer biology — was introduced in a company technical report in July 2024. It is a transformer whose layers run over a single sequence of multimodal tokens drawn from four spatially aligned measurements of the same tumor sample: multiplex protein staining, spatial gene expression, DNA sequencing, and H&E. That representation is decoded into a 16-channel multiplex fluorescence protein image, one decoder per channel. The model is queried by visual prompting — hand it a mostly blank image and it fills in the rest.

OCTO is the parent of Noetik's model family rather than a version of any member of it: OCTO-VC narrows to 1000-plex spatial transcriptomes and predicts a center cell's expression from its neighbors, while TARIO-2 runs from H&E alone to a predicted transcriptome. The model is proprietary — no weights, code or training data released — and its evidence base is Noetik's own technical reports plus two SITC 2024 abstracts, not peer-reviewed papers.

#Key Features

• Structured multimodal masking: Where standard masked image modeling hides ~75% of patches, OCTO trains with nearly every token masked across all 16 protein channels and the other modalities, so the only way to reconstruct anything is to combine information across them.
• Aligned biological tokens: Multiplex fluorescence is split into per-protein channels and patchified; spatial gene expression is tokenized one token per gene at each cell-containing location, keeping each modality's structure intact in a shared sequence.
• Counterfactual simulation: Real data is perturbed in one modality — raising a gene panel, knocking out a gene — and fed back through the fixed checkpoint; the difference between the two predicted images is the simulated effect.
• Emergent cell-type inference: Prompted with nuclear staining alone, the model separates tumor regions from immune nuclei, a task it was never explicitly trained on.
• Sparse visual prompts: Because training masks so aggressively, a prompt can be as thin as a single altered gene or protein channel.

#Technical Details

OCTO was pretrained on 20B tokens across 128 GPUs and holds up to 16,000 tokens in its context window at inference. Its corpus is Noetik's own, generated in-house from thousands of patient tumors spanning hundreds of millions of cells, with each patient represented by dozens of spatially randomized samples to keep the model off slide-level artifacts. The first atlas covers over 1,000 non-small cell lung cancer cases with paired protein staining, H&E, spatial transcriptomics and whole exome sequencing from the same specimens; roughly 10% carry long-term treatment follow-up. Parameter count is not disclosed.

Reported validation is qualitative agreement with known biology. The model reconstructs full 16-channel protein images with only 1-2% of patches revealed; masking non-randomly shows it has learned co-expression — CD8 with CD3, CD163 with CD68, PanCK with E-cadherin — and counterfactually raising interferon-gamma-responsive genes produces dose-dependent increases in predicted tumor HLA protein. The masking scheme earns its ablation: an autoencoder trained with standard masking cannot do the nuclei-to-cell-type task at all.

#Applications

The intended workflow is in-silico target screening. For each candidate target, the model simulates what would happen to cohorts of patient tumors if it were perturbed; targets whose simulated effect is therapeutically desirable, such as increased immune engagement with the tumor, are ranked and advanced to in vivo testing, and those results feed back as further training signal in a loop Noetik compares to RLHF. Two SITC 2024 posters reported the first applications: tertiary lymphoid structure identification, and candidate targets promoting anti-tumor immune response in lung tumors.

#Impact

OCTO's substantive bet is on training data rather than architecture: it learns from human tumor resections profiled across four modalities on the same tissue, not from cell lines, and the technical report argues that the masking strategy — not scale alone — is what turns that data into a simulator. It anchors a family that has since specialized in both directions, toward single-cell expression and toward H&E-only inference. The limitations are equally plain: every result is developer-reported and shown as qualitative agreement with known biology rather than benchmarked against competing methods, counterfactual effects are model beliefs about association rather than demonstrated causation, and weights, code and the training corpus are all closed.