Consider the clinical trial that failed in 2022. Or 2019. Or last month. Somewhere in the discarded tissue samples and frozen blood draws sits a signal no one thought to look for—a biomarker that might have predicted which patients would respond, buried in a measurement modality the study budget couldn't accommodate. Biology, after all, speaks in dozens of overlapping languages: genomics, proteomics, metabolomics, spatial transcriptomics. Researchers typically afford to run three or four.
Nine out of ten trials fail. What haunts drug developers isn't just the money lost, but the data never collected in the first place.
That gap—between what was measured and what mattered—has become the obsession of a new generation of AI companies. Their pitch: don't just analyze existing datasets. Predict the experiments you never ran. Use foundation models to translate between biological readouts, inferring expensive spatial proteomics from cheap pathology slides, or gene expression from genomic sequence alone. A virtual laboratory, if it works. Faster, cheaper, scalable.
Whether it actually works at scale remains the trillion-dollar question.
Betting on Prediction Over Production
The market, at least, has made up its mind. Grand View Research pegs AI in drug discovery at $2.35 billion this year, accelerating to $13.77 billion by 2033—call it 24.8% compounded annual growth. Other forecasters run even hotter. Precedence Research projects $16.49 billion by 2034. BCC Research models $15.2 billion by 2030 at a 31.7% CAGR. The ranges scatter, but the vector stays consistent: up and to the right.
Some of that optimism reflects pharma's productivity crisis. IQVIA's 2025 trends report noted marginal improvements in Phase III success rates and steadier enrollment timelines in 2024, but R&D spending keeps climbing. BCG modeled AI's potential upside in early 2025: 30–50% reductions in preclinical timelines, 25–50% cost savings. Impressive, if you believe computational predictions can reliably stand in for wet-lab work. And if regulators will accept them when approval decisions hang in the balance.
So far, that's a hypothesis awaiting proof.
The Cross-Modal Gambit
The technical shift is straightforward enough in concept. Instead of building models that analyze one data type, build systems that bridge between them. Take histology. A routine H&E slide—the workhorse stain pathologists have used for over a century—costs pennies and gets collected in every cancer case. Spatial transcriptomics or high-plex proteomics? Thousands of dollars per sample, specialized tissue prep, weeks of turnaround. If a model can predict the expensive readout from the cheap one with acceptable fidelity, the economics flip overnight.
January 2026 brought HEX, published in Nature Medicine: a system predicting 40 protein biomarkers from H&E across more than 2,300 non-small-cell lung cancer patients and 5,019 pan-cancer cases. The team validated externally using 57-plex CODEX spatial proteomics and claimed utility in prognosis and immunotherapy response prediction. A month earlier, npj Digital Medicine described STPath, a foundation model pretrained on whole-slide images and spatial transcriptomics, forecasting expression across 38,984 genes and 17 organs without task-specific fine-tuning.
Genomics follows the same playbook. CZI Biohub's VariantFormer—1.2 billion parameters, released November 2025—predicts gene-level expression from personalized genomes, trained on what the team describes as the largest paired whole-genome sequencing and RNA-seq dataset assembled to date. Nature Genetics published Borzoi in early 2025, forecasting RNA-seq coverage from DNA sequence and holding its own on quantitative trait loci tasks.
Not purely academic exercises, these. Y Combinator's Winter 2026 batch includes Strand AI, a two-person San Francisco startup building cross-modal foundation models for exactly this problem. According to the company's February launch materials, Strand's first model predicts spatial proteomics from routine H&E and was "trained in under six weeks," claiming to beat state-of-the-art benchmarks. The company also licenses curated multimodal datasets—a dual revenue stream that underscores how tightly model performance couples to training data quality.
Strand's CEO Yue Dai previously worked on spatial biology at Enable Medicine, Tempus AI, and Microsoft Research. CTO Oded Falik also comes from Enable Medicine's spatial platform. Their pitch centers on a simple claim: biomarker signal often hides in unmeasured modalities, and cross-modal models can fill those gaps without running the assays. To signal community engagement, Strand ran CZI's VariantFormer model to generate RNA-seq predictions for over 500 previously unmeasured 1000 Genomes Project samples, optimizing inference to run 37 times faster on NVIDIA A100 GPUs versus H100s.
Whether that's a genuine technical advantage or clever marketing remains to be seen.
From Proofs of Concept to Production Claims
The pace of publication since mid-2025 suggests the field has moved past validation and into productization mode—or at least wants investors to think so. Beyond HEX and STPath, recent work includes Img2ST-Net (January 2026), PEaRL (October 2025), 3D HoloTea (November 2025), and HEIST (June 2025), all tackling variations on the histology-to-omics problem. On the genomic side, long-context DNA models like HybriDNA and GENERator appeared in February 2025, followed by diffusion and hierarchical tokenizer-free architectures—D3LM in March 2026, dnaHNet in February.
The naming conventions alone betray the field's youth.
Momentum isn't confined to startups. NVIDIA announced a major expansion of its BioNeMo platform on January 12, 2026, at the J.P. Morgan Healthcare Conference, adding new open models and libraries. Eli Lilly and NVIDIA launched a co-innovation lab. Thermo Fisher partnered on autonomous laboratory infrastructure. NVIDIA's messaging frames biology as reaching "its transformer moment," positioning BioNeMo as the connective tissue between wet-lab experiments and AI-driven prediction.
Larger TechBio players are integrating multimodal approaches as well. Recursion completed BioHive-2 in May 2024—what it called the largest NVIDIA-powered supercomputer in life sciences—and has sustained partnerships with Sanofi, Roche, and Genentech throughout 2025, blending phenomics data with real-world datasets from Tempus and HealthVerity. Isomorphic Labs, Alphabet's drug discovery subsidiary, secured a $600 million raise in March 2025 and expanded multi-billion-dollar collaborations with Lilly and Novartis dating back to early 2024.
Tempus AI, which went public in June 2024 and upsized $650 million in convertibles in June 2025, sits at the intersection of real-world genomics and AI-driven clinical applications. PathAI has built foundation models for digital pathology and presented biomarker discovery work at AACR 2024. DeepMind's AlphaFold 3, released in May 2024, extended structure prediction to proteins, nucleic acids, and ligands. EvolutionaryScale's ESM3 generatively designs proteins, simulating what the company describes as "500 million years" of evolution—a claim that sounds impressive until you remember evolution doesn't actually have a clock.
The Data Moat
Model performance hinges on training data, and curated multimodal cohorts have become a competitive differentiator. Enable Medicine and Akoya Biosciences announced a 100-million-cell, 8,500-plus-sample Pan-Cancer Atlas in April 2025, billing it as the largest commercially available single-cell spatial proteomics dataset. That kind of petabyte-scale resource is exactly what Strand licenses alongside its models.
10x Genomics has been steadily expanding single-cell and spatial transcriptomics accessibility. The company shipped Visium HD in 2024 for near-single-cell resolution on FFPE tissue and has continued platform updates through 2025 and into 2026. High-plex proteomics technologies—CODEX, CosMx, Xenium—are generating the validation datasets that academic groups cite in their cross-modal prediction papers. HEX, for instance, used 57-plex CODEX as external validation.
This creates a feedback loop. Better instruments generate more multimodal data, which trains better models, which in turn reduce the need to run expensive assays on every sample.
But the data itself remains tightly controlled. HIPAA's de-identification rules govern U.S. patient data use. GDPR treats genetic and health data as "special category" information requiring explicit legal bases under Article 9. NIH's Data Management and Sharing Policy, effective since January 2023, mandates FAIR-aligned plans for funded research, shaping how academic datasets become available for model training.
For a field that likes to talk about democratizing biology, there's a lot of gatekeeping around who gets access to the raw materials.
Regulatory Reality Check
The technical progress has outpaced regulatory clarity—a familiar pattern in AI, but one with higher stakes when human lives enter the equation.
FDA released draft guidance on AI model credibility for drugs and biologics on January 6, 2025, emphasizing risk-based frameworks and "context of use" for models supporting submissions. For medical devices, final guidance on Predetermined Change Control Plans came in early 2025, establishing how adaptive AI systems can handle model updates post-market.
Cross-modal prediction models occupy uncertain territory. Used purely for research or patient stratification in early discovery, they may sidestep Software as a Medical Device classification. But move them closer to clinical decision support—say, using virtual proteomics to guide treatment selection—and device pathways kick in. Models used in regulatory submissions to support enrichment strategies or endpoint surrogates will fall under FDA's AI credibility framework and real-world evidence guidelines. Early engagement with the agency's Model-Informed Drug Development program, which runs through fiscal year 2027, is probably the safest route.
Europe's AI Act adds another layer. The law entered force in August 2024, with general-purpose AI obligations effective as of August 2, 2025, and most high-risk system rules arriving August 2, 2026. Full implementation stretches to August 2027. For pharma and medtech buyers, that means model documentation, transparency, and governance will increasingly become procurement criteria.
Then there's the question of whether the predictions actually work when stakes are high. Most cross-modal validation remains retrospective: models trained on paired datasets, tested on held-out cohorts, validated against existing measurements. External prospective trials using H&E-derived virtual omics to guide patient selection in real time? Scarce. A February 2026 preprint raised concerns that attention mechanisms in single-cell foundation models may capture co-expression patterns more than unique causal regulation, suggesting interpretability remains unsolved.
Eli Lilly's CIO Diogo Rau declared in June 2025 that "non-AI organizations won't make it to 2050." Bold claim. But declarations like that tend to age unpredictably.
McKinsey estimated generative AI's annual impact across pharma and medical products at $60 billion to $110 billion—2.6% to 4.5% of sector revenue. That forecast, however, dates to June 2023 and carries the usual caveats about adoption timelines. Consulting firms have a mixed track record on these things.
The Market Underneath
The spatial omics market itself is expanding, serving as the enabling technology for these models. Grand View Research pegs spatial transcriptomics at $385.7 million in 2024, growing to $1.31 billion by 2033 at a 14.69% CAGR. MarketsandMarkets forecasts $554.5 million in 2024 reaching $995.7 million by 2029. GIA estimates spatial omics more broadly at $767.7 million in 2024, climbing to $2 billion by 2030.
The ranges diverge. The direction doesn't.
What matters is whether cross-modal models can move from validation papers to routine decision-making: trial design, patient selection, dose optimization. Strand's claim that it trained a state-of-the-art spatial proteomics predictor in six weeks is either a preview of industrialized model development or an overpromise that will meet the usual friction of biology's complexity. Time will tell. The company's focus on licensing curated datasets alongside models suggests they understand that training data quality, not just architecture, determines performance. That's a more sophisticated insight than most AI-for-X startups demonstrate.
Larger players are hedging by building infrastructure that supports multiple approaches. NVIDIA's BioNeMo expansion ties GPU provisioning to lab-in-the-loop workflows. Enable Medicine's atlas strategy monetizes curation as much as computation. Academic groups like CZI are open-sourcing models like VariantFormer to accelerate community iteration.
What's Left Unresolved
The questions that matter cluster around reliability and trust. Batch effects, domain shift, and interpretability gaps haven't disappeared just because transformers have arrived in genomics. GDPR and HIPAA constrain how training data flows across borders and institutions, complicating the petabyte-scale ambitions that foundation models typically require. Federated learning approaches offer a path forward but add operational overhead that slows development cycles.
Still, the incentives are aligned—perhaps more than usual in biotech, where misalignment is the rule. Pharma needs faster, cheaper ways to stratify patients and prioritize targets. Academic labs want to extract more insight from expensive cohorts. Regulators are signaling openness to model-informed submissions if the credibility framework holds.
Strand and its peers are betting that biology's data gaps represent not just a scientific bottleneck, but a commercial opportunity large enough to support an entire layer of the stack. One that sits between raw measurements and clinical decisions, filling in what was never collected in the first place.
Whether that bet pays off depends on questions that can't yet be answered from validation papers and benchmark leaderboards. Can virtual omics survive contact with prospective trials? Will regulators accept them as surrogates when approval decisions hang in the balance? And perhaps most importantly: are these models actually learning biology, or just sophisticated pattern-matching on training data that doesn't generalize when the context shifts?
The ghost in the failed trial might be real data that was never measured. Or it might be a phantom—signal that didn't exist in the first place, now conjured by a model trained to see patterns whether they're there or not.
We'll find out which, but not for a few years yet.
