oFM

A cancer patient's chart is not a snapshot. A biopsy at diagnosis, a sequencing report months later, a radiographic progression the following spring, and four lines of therapy in between are all evidence about one evolving disease — yet most clinical prediction models flatten them into a single row of features measured at a single moment. Precision oncology needs a representation that carries how the tumor and its treatment history changed over time.

oFM — the oncology foundation model — from Tempus AI is built for exactly that. It decomposes each patient record into atomic facts — diagnoses, lab values, medications, procedures, DNA alterations, RNA expression abnormalities, IHC biomarkers — groups them into day-level episodes, and encodes each episode as text with a clinical language model. Whole-slide H&E images enter the same latent space as separate tokens. A transformer over the date-stamped sequence emits one 1,024-dimensional patient-state embedding, read at any clinically meaningful anchor.

The distinction the work presses on is between prognostic and predictive signal: a prognostic marker says a patient will do badly, a predictive one says which treatment would change that. Because oFM is trained to predict a patient's future state conditioned on the intervention at the anchor day — with that intervention withheld from the input context — its embeddings are evaluated on differential treatment benefit as well as on survival. It was posted as a preprint in August 2026.

#Key Features

• Day-level episodes as text: Every clinical and molecular event on a given day is serialized into one textual episode, so a KRAS G12V call, a hemoglobin value, and a line-of-therapy change share a single encoder instead of per-modality feature engineering.
• Magnitude-aware numerics: A Fourier Number Embedding pathway encodes each measured value at 32 frequencies and injects it at the matching numeric token, preserving magnitude that subword tokenization destroys.
• Pathology as a droppable modality: Slides are tokenized with Virchow2 and pooled by PRISM2, then projected into the trajectory space — and dropped 30% of the time in training, so the model stays usable for the majority of patients with no imaging.
• Frozen-embedding evaluation: Every downstream result comes from a linear or Cox probe on frozen embeddings, so gains are attributable to the representation rather than to task-specific fine-tuning.

#Technical Details

The episode encoder is a fine-tuned GatorTron-base-2k clinical BERT (hidden size 1,024, 2,048-token context) with the numeric pathway, at 358M parameters; a two-layer, 16-head trajectory transformer with rotary positional embeddings adds 25.2M, and the reconstruction, projection, predictor and survival heads a further 38.1M — 421M trained parameters in all, sitting on a frozen 5.1B-parameter PRISM2 pathology stack. Training runs in three stages: denoising autoencoding of single episodes, trajectory training with the episode encoder frozen, then joint fine-tuning. The trajectory objective sums masked episode reconstruction, joint-embedding future prediction with VICReg variance and covariance terms, and a Cox partial-likelihood survival loss.

The corpus is 1,672,203 de-identified patients from the Tempus Data Model, of whom 1,045,011 trained the episode encoder, 386,382 the trajectory encoder, and 92,567 the joint stage; 23.0% carry DNA sequencing, 16.1% both DNA and RNA, and 14.8% linked H&E slides. Against a baseline selected from 7,520 curated clinical and molecular features, frozen oFM embeddings reach mean AUC 0.774 for overall survival (baseline 0.563), 0.688 for progression-free survival (0.544), and 0.585 for treatment response (0.513), winning 95% of tumor-by-therapy strata for the first two. Across 11 comparative-treatment cohorts spanning breast, colorectal, lung, renal-cell, prostate, and pan-cancer settings, treatment-benefit ranking reaches a pooled t-statistic of 4.61 against the baseline's 1.38, with oFM ahead in 9 of 11.

#Applications

The intended use is as a frozen feature extractor for oncology questions that would each otherwise need their own labeled cohort: risk stratification before first-line therapy, enrichment and endpoint modeling for trial design, and biomarker discovery. The paper also describes a mechanism-discovery layer that makes downstream models inspectable — leave-one-out attribution over atomic facts, a sparse autoencoder over the embedding, causal steering by ablating latents, and retrieval over CIViC, OncoKB, DrugBank and related sources using MedGemma — assembled into a temporally ordered graph linking concepts to outcomes.

#Impact

oFM carries the patient-level multimodal modeling of Verily's EHR + genomics foundation model into oncology, and is the longitudinal counterpart to single-timepoint pathology-plus-molecular models like Threads. Its most consequential claim is methodological rather than architectural: representation learning over patient trajectories can recover differential treatment benefit, a signal curated-feature baselines extract only inconsistently. The limits are real. All evaluations are retrospective and drawn from one vendor's data platform, with the confounding and censoring that implies, and the authors state that external and prospective validation remain to be done. Neither the code, the weights, nor the patient corpus has been released, and the paper carries no availability statement, so independent replication is not currently possible.