Discoveries in Health Policy: Ryan Fukushima Essays on Precision Medicine

Ryan Fukushima (Linked In here) is COO of TEMPUS AI, which has a $13B market cap and $700M in 2024 revenue.

I've seen his occasoinal Linked In essays in the past year on precision medicine innovation & strategies. I just discovered his longer set of essays at Substack - https://ryanfukushima.substack.com/about

Here are a few titles followed by an AI summary.

About Ryan Fukushima here.
The open source movement in pathology AI here.
AI and Lab in the Loop medicine here.
Multi modal AI beats Single source for cancer predictions here.
Deploying clinical multi modal AI here.
Typhoons to Tumors - Precision medicine's future here.

####

AI CORNER

####

##
GOOGLE NOTEBOOK LM

Based on the provided articles and biographical statement, Ryan Fukushima, COO of Tempus AI, presents a sophisticated and forward-looking perspective on the intersection of precision medicine and artificial intelligence. His thinking, targeted at a scientifically literate audience, revolves around several core principles: the necessity of multimodal data integration, the power of AI to bridge the translational gap between preclinical research and clinical reality, the shift from deterministic to probabilistic models in patient care, and a pragmatic approach to deploying AI in complex healthcare ecosystems.

The Centrality of Multimodal Data Integration

A cornerstone of Fukushima's philosophy is that transformative insights in oncology arise from the intelligent integration of multiple data modalities, not from any single source. He argues that focusing on one data type, whether genomics or digital pathology, provides an incomplete picture of a patient's disease.

Complementary Biological Signals: Fukushima emphasizes that different data types capture distinct and complementary aspects of tumor biology. For instance, a study he highlights showed that combining transcriptomics (dendritic cell abundance), genomics (NTRK1 expression), radiomics (metabolic tumor volume), and clinical data (neutrophil counts) painted a complete prognostic picture that no single biomarker could provide. This complementarity is crucial, as different modalities can provide opposing information for the same patient, and integrating them leads to significantly more powerful predictive models. A model combining clinical, pathomic, and RNA data yielded a log-rank p-value of 3.51e-09 for predicting 1-year mortality, far outperforming single-modality approaches.
Early vs. Late Fusion: He champions "early fusion" methods in AI model development, where different data streams are integrated at the beginning of the analytical process. This allows the model to learn the "fine-grained interactions" between modalities—for example, how specific molecular pathways directly influence tissue morphology. Citing the SURVPATH model, he notes that this approach yields significant performance improvements (a 7.3% gain over leading models) compared to "late fusion" methods that analyze data separately and combine conclusions at the end.
The Data Problem: Fukushima sees the primary challenge in the field not as an algorithmic one, but as a "data problem". He stresses the importance of bridging real-world, multimodal data at scale, using "source data" that is as unprocessed as possible to avoid introducing artifacts that could doom an AI model from the start. Organizations like Tempus, he notes, are working to solve this by building expansive, de-identified multimodal datasets that can power the next generation of models.

Bridging the Translational Gap with AI

Fukushima identifies the "translational gap"—the failure of promising preclinical findings to translate into effective human therapies—as one of medicine's most persistent and costly challenges. He sees AI as an essential tool for bridging this gap.

Improving Preclinical Models: He highlights the fundamental transcriptional differences between mouse models and human tumors, which often render preclinical efficacy studies unpredictive. To address this, he points to advanced deep learning methods like Novartis's MOBER (Multi-Origin Batch Effect Remover). MOBER can simultaneously analyze data from cell lines, patient-derived xenografts (PDXs), and clinical tumors to identify which lab models have the greatest "transcriptional fidelity" to real cancers. This allows researchers to select better models, potentially saving years of research and millions of dollars.
Transforming and Harmonizing Data: Beyond selection, AI tools like MOBER can computationally "transform" the expression profiles of preclinical models to better resemble clinical tumors. Fukushima finds this capability particularly compelling, citing an example where biomarkers identified in MOBER-transformed cell line data showed dramatically improved clinical translation and survival stratification in patient cohorts (P = 6.2× 10^-29). This AI-driven harmonization allows for more relevant preclinical experiments that can better de-risk future clinical trials.

Shifting from Deterministic to Probabilistic Frameworks

Fukushima argues for a paradigm shift in clinical decision-making, moving away from the "one-size-fits-most" approach of traditional evidence-based medicine toward a more personalized, probabilistic framework.

The "Average Patient" Fallacy: He notes that evidence-based medicine is inherently deterministic, providing guidance that works "on average" but often fails to account for the unique biological complexity of individual patients. Most patients exist in a "clinical gray zone," where their individual characteristics may question the applicability of population-level findings.
Learning from Other Complex Systems: Drawing powerful analogies from weather forecasting and traffic simulation, Fukushima illustrates the superiority of probabilistic models in complex, dynamic systems. He points to DeepMind's GenCast, a weather model that generates an ensemble of possible scenarios rather than a single deterministic forecast. The ensemble mean often "blurs out" critical details (like the precise path of a typhoon), much like population averages in medicine obscure the reality of an individual patient's disease trajectory.
Quantifying Uncertainty: By embracing uncertainty, probabilistic AI models can quantify confidence in treatment recommendations, identify patients needing closer monitoring, and design adaptive treatment protocols. This approach is better suited for the inherent stochasticity of disease progression and treatment response.

Pragmatic Deployment of Clinical AI

While optimistic about AI's potential, Fukushima maintains a pragmatic and clear-eyed view of the challenges involved in deploying these tools in real-world clinical settings.

AI as an "Intelligence Amplifier": He refutes the idea that AI is meant to replace clinical judgment, instead framing it as an "intelligence amplifier" that helps clinicians see hidden patterns in existing data. He uses the analogy of Google Maps: it doesn't drive your car but provides real-time information to help you make better decisions. The goal is to empower doctors with smarter systems that can make truly personalized decisions.
The Open-Source Ecosystem: Fukushima is a strong advocate for the open-source movement in AI, particularly in digital pathology. He highlights foundation models like Paige's Virchow, Microsoft's Prov-GigaPath, and Bioptimus's H-Optimus-0, which are democratizing access to state-of-the-art tools and creating a global ecosystem where innovation can accelerate exponentially. This allows research teams to build on established foundations rather than reinventing the wheel.
Beyond Model Building: He stresses that building a sophisticated model is meaningless if it cannot be deployed to impact patients. The true challenge lies in navigating technical barriers, regulatory hurdles (like the FDA clearance process), and workflow integration. His strategic framework for AI leaders emphasizes building on open foundations, focusing on clinical integration and validation, prioritizing multimodal data, and scaling through partnerships.

In summary, Ryan Fukushima's thought leadership is defined by a vision where the integration of comprehensive, multimodal data through sophisticated yet interpretable AI is not merely an academic exercise, but the fundamental driver of a more precise, personalized, and effective standard of care in oncology.

CHAT GPT 5

Here’s a synthesis of Ryan Fukushima’s recent essays and bio—the through-line is data realism + pragmatic deployment: use multimodal evidence, design for translation into clinic, and bias toward systems that adapt over time rather than static “point” models.

1) Why multimodal beats unimodal (and how to do it right). Across posts, Fukushima argues that single-source models (e.g., WSI-only) rarely capture the causal texture of disease. He highlights work like SURVPATH to show early-fusion architectures (a joint feature space where histology tokens “converse” with pathway tokens) outperform late fusion and deliver interpretable links between morphology and biology (e.g., EMT pathways ↔ image features). The practical takeaway for biotech teams: stop expecting any one modality to “rule them all”; build data pipelines that integrate transcriptomics + imaging + clinical from day one and keep interpretability first-class for translational adoption.

2) Deployment mindset: AI as an “intelligence amplifier,” not an intervention. He pushes back on calls for RCT-style evidence of the tool itself, reframing clinical AI as decision-support (like Google Maps): it surfaces patterns clinicians would otherwise miss, updates with new information, and helps tailor choices to patients whose profiles only partially overlap trial cohorts. He illustrates how feature importance is context-dependent (e.g., CRP risk flips with platelet strata) and how the salience of biomarkers shifts over the cancer journey—hence the need for adaptive, longitudinal models and robust guardrails (validate against outcomes, ensure explainability, bias monitoring, UI that turns signals into actions).

3) Lab-in-the-loop translation: fix the model-to-patient gap. Fukushima centers the preclinical translation problem and spotlights MOBER as a deep-learning bridge across cell lines, PDXs, and tumors (n≈932/434/11,159). MOBER aligns datasets by biological rather than batch origin, identifies which lab models actually resemble human tumors, and can transform preclinical expression profiles to better match clinical biology—improving survival stratification in TCGA (P≈6.2×10⁻²⁹). The point isn’t a single algorithm; it’s a playbook: pick the right models, transform when needed, and iterate against real clinical outcomes.

4) Systems thinking: from “point measurements” to path-level, probabilistic digital twins. Borrowing from traffic calibration and DeepMind’s GenCast weather work, he argues precision medicine should model journeys (trajectories across scales and time) rather than isolated markers. That implies path-level data integration, multi-scale calibration, and probabilistic forecasts that quantify uncertainty—exactly what clinicians need for rare but high-impact “tail events” (e.g., atypical responses, adverse events). Biological complexity and fragmented data remain hard problems; still, hybrid mechanistic-statistical modeling and improving compute suggest a tractable roadmap.

5) Digital pathology and open source: build on foundations, ship to clinic. He sees an inflection point as open models (e.g., Virchow/PRISM, PROV-GigaPath, UNI/CONCH, H-Optimus-0) push benchmark→ecosystem. But he’s blunt: great models mean nothing if they never reach patients. His “leader’s framework”: (i) build on open foundations to move fast; (ii) focus on clinical integration (validation, workflows); (iii) prioritize multimodal (pathology+genomics+clinical); (iv) scale through partnerships (regulatory-proven companies, hospital integrations). He underscores Paige’s FDA experience and Tempus’ data + deployment infrastructure as the route from proof-of-concept to clinical-grade.

6) “What elite teams know”: let the data choose the biomarker strategy. In oncology (e.g., NSCLC, variable immunotherapy response), he urges moving beyond simplistic single-gene heuristics toward complex multimodal biomarkers tuned for clinical applicability (sample availability, assayability) and predictive lift. With large, well-annotated de-identified datasets and tooling (e.g., curation, automated pipelines), organizations can industrialize biomarker discovery→validation and avoid chasing narratives that don’t survive contact with longitudinal outcomes.

7) “Multisource > single-source” is also a culture and process claim. He describes real rooms where pathology, genomics, and radiology “talk past each other” and argues the win condition is operational: converge teams and tools around shared multimodal abstractions, enforce interpretability, and stage roll-outs where benefits clearly exceed integration costs. That’s how you cross the research→clinic chasm.

8) Biography → strategic lens. Fukushima’s bio frames this as a builder’s agenda: Tempus (AI-enabled diagnostics + massive clinical/molecular library for real-time decisions and biopharma R&D), Pathos (AI to understand why patients respond/don’t, lowering trial failure), and >$1.4B raised across companies focused on “right treatment, right time.” His newsletter promises implementation frameworks, tech evaluations, and market/regulatory analysis—i.e., an operator’s view of where AI + precision medicine actually change care.

What this adds up to (for expert readers)

Epistemology: Prefer path-level, probabilistic modeling over point estimates and averages; treat uncertainty as a signal to manage, not a nuisance to ignore.
Architecture: Invest in early fusion multimodal models with transparent tokens/factors that map onto biology and clinic; make interpretability non-negotiable to speed validation and payer/regulator uptake.
Translation: Operationalize lab-in-the-loop (model selection + transformation + outcome feedback) to shrink the preclinical↔clinical gap; measure gains in calibration and tail-risk detection.
Strategy: Stand on open foundations, but win on deployment: workflow fit, regulatory path, and partnerships that get models into care settings—because impact = model × (integration & scale).

If your team already lives in WSI/NGS/registry land, Fukushima’s message is not “add one more model,” it’s re-platform around multimodal, adaptive, clinically-tethered systems—and judge success by what reaches patients, not by leaderboards.

SIDEBAR

I asked Google LM and Chat GPT 5 to answer my question, does Fukushima risk underestimating the problem of Overfitting, since he aims to apply ultra high dimension data to simple questions like "survival" or "best drug."

I place the Google and CHAT answers in a sidebar.

Then, I asked both Google and CHAT to compare the overfitting answers they generated, with each other.

https://bqwebpage.blogspot.com/2025/09/does-high-multiplicity-ai-data-plus.html

And here, comedian Lewis Black summarizes it all.

https://bqwebpage.blogspot.com/2025/09/lewis-black-on-overfitting-high.html