Three billion dollars buys a lot of confidence. That's what Isomorphic Labs wagered late last year in partnerships with Eli Lilly and Novartis—a bet that artificial intelligence could fundamentally reshape how new medicines get made. Generate:Biomedicines followed with its own multi-target deals. McKinsey, never shy about attaching eye-popping figures to emerging trends, estimates generative AI could unlock somewhere between $60 and $110 billion annually across the pharmaceutical industry.
But here's the thing nobody mentions in those glossy press releases: the molecules are piling up faster than anyone can figure out if they actually work.
AI systems are now churning out thousands of novel protein therapeutics—antibodies, bispecifics, antibody-drug conjugates, the whole alphabet soup of modern biologics. They're arriving at a pace that would have seemed fantastical a decade ago. The bottleneck, it turns out, isn't in the algorithm. It's in the mass spectrometer.
Every AI-designed biologic, no matter how elegant its computational pedigree, must survive months of rigorous protein characterization before it gets anywhere near a patient. Mass spectrometry workflows. Glycan analysis. Aggregation studies. The painstaking mapping of post-translational modifications. This is where promising candidates go to wait, and wait, and sometimes stall out entirely.
Which might explain why three scientists from Stanford's Bertozzi Lab and the Broad Institute just emerged from Y Combinator's Winter 2026 batch with a startup called 10x Science—and a pitch that the industry's next breakthrough won't come from better drug design algorithms, but from faster, smarter ways to prove those designs aren't going to kill anyone.
Numbers That Tell a Story
The proteomics market is booming, if you believe the projections. MarketsandMarkets sees growth from $36.3 billion in 2025 to $65.8 billion by 2030—a 12.6% compound annual growth rate. The Business Research Company is even more optimistic, forecasting $73.3 billion by decade's end.
But market size was never the problem. Throughput is.
AI drug discovery platforms keep accelerating. IQVIA's 2025 Global Trends report shows clinical program duration dropped from 10.1 years in 2022 to 9.3 years in 2024. Roughly $10 billion in AI and machine learning deals got signed last year alone. BCG research suggests that AI-first biopharma companies can slash early discovery timelines, with initial Phase 1 success rates for AI-discovered molecules appearing higher than traditional approaches. (The sample size, admittedly, remains small enough that nobody should be declaring victory yet.)
Meanwhile, characterization workflows are stuck somewhere in 2015.
A 2025 study with the memorable name PROPHET-Ab ran 10 assays across 246 antibodies to improve machine learning predictions of developability—things like solubility, stability, viscosity, aggregation, polyreactivity. That's the scale needed for reliable AI models: hundreds of candidates, dozens of assays, standardized datasets that can actually train an algorithm properly. Most labs can't generate that data quickly enough. Some can't generate it at all.
The Industry Scrambles
Instrument makers, sensing both opportunity and existential threat, are racing to close the gap. Thermo Fisher spent $3.1 billion to acquire Olink last July, adding proximity extension proteomics to its already formidable portfolio. At this year's ASMS 2025 conference, the company unveiled the Orbitrap Astral Zoom and Orbitrap Excedion Pro—instruments targeting higher throughput and the complex biomolecule analysis that keeps lab directors up at night. Bruker countered with the timsTOF Ultra 2 for single-cell proteomics and the timsOmni for proteoform sequencing. Waters integrated SEC-MALS with liquid chromatography through something called HPLC CONNECT, which sounds less revolutionary than it apparently is.
The software layer is evolving too, though not always gracefully. DIA-BERT, published in Nature Communications in 2025, uses transformer pre-training on 276 million precursors to boost peptide identifications versus traditional algorithms. A February 2026 preprint describes DIA-CLIP, a cross-modal pre-trained model claiming 45% more identifications in benchmarks through "zero-shot" inference—a term that sounds impressive until you try explaining it to a CFO. Thermo Fisher's Ardia platform attempts to integrate AI-enhanced peptide identification across its instrument fleet.
Integrating these tools into actual biopharma workflows, though? That's where things get messy.
Antibody-drug conjugates present particular headaches. A 2025 AAPS Journal review highlighted persistent analytical difficulties in characterizing drug-antibody ratios and positional isomers—the kind of technical challenges that don't yield to brute computational force. Multiple separation techniques are often required: HIC, RP, AIEX, iCIEF, SEC, MS. Method selection depends on linker chemistry, and there's no one-size-fits-all protocol. You could staff an entire department just figuring out which test to run first.
Regulators Create Space (Sort Of)
To their credit, regulators are trying not to be the industry's villain in this story. The FDA finalized ICH Q14 guidance in 2024, encouraging lifecycle, science-based analytical procedure development. The approach supports flexible post-approval changes and explicitly accommodates AI-assisted multivariate methods—bureaucratic language that actually matters when you're trying to get a drug approved.
In September 2025, the agency issued final guidance on biosimilar development that increases emphasis on robust comparative analytical assessment. ICH M10, implemented in 2022-2023, clarified bioanalytical method validation expectations for biologics. And ICH Q5E comparability remains foundational—any manufacturing process change requires deep analytical head-to-head characterization before and after. No shortcuts.
The FDA is also engaging directly with AI and machine learning in drug development through discussion papers and risk-based credibility frameworks. For platforms like 10x Science, this regulatory evolution matters more than any venture capital check. The agency appears willing to embrace advanced analytics, provided—and this is the catch—the underlying data quality and validation are sound.
Nobody's getting a free pass just because their deck has the letters "AI" in 72-point font.
The Stanford Bet
This is where 10x Science enters the picture, though "enters" might be generous for a company barely out of Y Combinator. The team describes its offering as an "AI-native platform for next-generation protein characterization," designed to remove the bottleneck between AI drug design and biopharma development. Whether that's visionary or wishful thinking depends on execution details the company hasn't fully disclosed yet.
The founding team brings relevant depth, at least on paper. CEO David Roberts is a postdoctoral researcher in the Bertozzi Lab at Stanford—one of those elite academic lineages that still carries weight in biotech—and a Damon Runyon Fellow with expertise in top-down mass spectrometry and glycobiology. COO Andrew Reiter developed proteomics methods at the Broad Institute and is pursuing a Stanford PhD co-advised by Carolyn Bertozzi and Or Gozani. CTO Vishnu R. Tejus is a two-time YC founder and former founding engineer at Nooks, bringing the software DNA that pure-play biotech startups often lack.
The Bertozzi Lab connection isn't just résumé polish. Glycobiology and post-translational modifications remain among the field's thorniest analytical challenges—critical quality attributes for therapeutic antibodies that influence efficacy, safety, and immunogenicity in ways that aren't always predictable. Multiple 2025 and 2026 reviews discuss high-throughput glycan analysis methods and therapeutic antibody glycan characterization, reflecting ongoing technical hurdles that haven't yielded to conventional approaches.
What Success Actually Looks Like
Nature Methods named spatial proteomics Method of the Year 2024, which tells you something about where the field's attention is focused. 10x Genomics launched Xenium Protein in August 2025, enabling same-cell RNA and protein analysis in situ. Quantum-Si is pushing single-molecule protein sequencing with its Platinum Pro platform. The field is maturing from pure discovery mode into translational contexts that pharmaceutical companies might actually pay for.
But—and this is a significant qualifier—spatial and single-cell methods complement, rather than replace, the bulk characterization workflows required for CMC submissions and regulatory comparability studies. Every biologic still needs conventional analytics at scale. The question isn't whether those old-school assays remain necessary. They do. The question is whether those workflows can be automated, standardized, and AI-enhanced enough to match the pace of upstream discovery.
Benchling launched Bioprocess this year for design-to-execution-to-insights in process development, with instrument integrations for high-throughput workflows. CROs like BioPharmaSpec offer ICH Q6B-aligned characterization services. The ecosystem is building toward integrated, data-rich platforms, though whether any single player can own the entire stack remains an open question. Pharma companies have long memories about vendor lock-in.
The Real Test
The industry is converging on a few uncomfortable realities.
First, AI-generated candidates will continue flooding development pipelines. The Isomorphic and Generate deals suggest sustained investment in upstream discovery, and there's too much momentum—and too much capital—to reverse course now. Second, regulators are comfortable with advanced analytical approaches if properly validated, which sounds permissive until you parse the "if properly validated" part. Third, characterization remains stubbornly physical and time-intensive, resistant to the kind of Moore's Law exponential improvements that transformed computing.
10x Science's bet is that AI can solve this. Not by replacing mass spectrometers or chromatography columns—those aren't going anywhere—but by orchestrating them more intelligently. Selecting optimal assays. Predicting quality attributes before running expensive tests. Integrating disparate datasets that currently live in different software silos. Automating method development under ICH Q14 principles in ways that human scientists, no matter how talented, simply can't match for speed.
Whether that vision materializes depends entirely on execution, and on details the company hasn't shared publicly yet. The proteomics market is large and growing, yes, but it's also crowded with established players who've spent decades building relationships with pharma QC labs. Startups need differentiated technology and a clear path to adoption. Being right about the problem doesn't guarantee success with the solution.
For biotech founders watching AI-designed molecules pile up in their characterization queues, though, the problem is real enough—and urgent enough—that they'll try almost anything.
The next breakthrough in drug development might not come from better algorithms or more sophisticated protein folding predictions. It might come from something far more prosaic: faster labs. Less glamorous, perhaps, but no less essential. Sometimes the bottleneck is exactly where you'd least expect it.
