David Roberts spent years in Carolyn Bertozzi's Stanford lab learning how to coax secrets from proteins. The work was painstaking—weeks to characterize a single candidate, months to prepare documentation rigorous enough for regulatory scrutiny. By early 2025, that timeline had become untenable.
Not because the science got harder. Because it got too easy.
When DeepMind released AlphaFold 3 in May 2024, extending its predictions beyond isolated proteins to entire molecular complexes, the computational biology field entered what one venture capitalist privately called "the promise years." Pharma partnerships proliferated. Isomorphic Labs added Eli Lilly and Novartis to its roster, then Johnson & Johnson in February 2025. EvolutionaryScale pulled in $142 million that June, showcasing its ESM3 model by generating a novel fluorescent protein from scratch. Generate Biomedicines pushed an AI-engineered therapeutic into Phase 3 trials by January 2026.
The revolution delivered. Then it created a new crisis.
Drug developers could suddenly generate hundreds of promising protein candidates in silico—computationally, before ever touching a lab bench. What they couldn't do, at least not quickly, was validate them. The bottleneck shifted from design to characterization, from algorithm to empirical proof. Instrument makers scrambled to boost throughput. The analysis software? That lagged behind, still demanding weeks of manual interpretation to transform raw data into anything actionable.
Perhaps no one expected the computers to get quite this good, quite this fast.
When Speed Becomes the Problem
The numbers sketch the contours of the challenge. AlphaFold 3 didn't merely refine previous predictions—it broadened scope to ligand binding, nucleic acid interactions, entire molecular assemblies. This accelerated the design pipeline in ways that caught even optimists off guard. Isomorphic's February 2025 expansion with Novartis signaled that AI-to-chemistry workflows had matured past proof-of-concept. Novo Nordisk unveiled results in March 2024 from an AI-plus-high-throughput experimentation system that delivered the first calcitonin-based compound with ten-fold amylin selectivity. These weren't distant possibilities. They were molecules advancing toward patients.
But every AI-designed candidate still requires empirical validation. Always has, likely always will.
Mass spectrometry identifies post-translational modifications that affect function. Surface plasmon resonance measures binding kinetics. Glycoproteomics maps sugar chains that determine stability and immunogenicity. The instruments exist, and they're improving rapidly. Thermo Fisher launched its Tandem Direct Injection workflow in June 2024, promising four-fold throughput gains with TMTpro 32plex multiplexing. Bruker unveiled the timsTOF Ultra 2 at ASMS 2025, targeting single-cell sensitivity with its Athena Ion Processor.
Hardware was catching up. Software wasn't.
A 2025 review in Analytical Chemistry noted the convergence of high-throughput proteomics with machine learning, painting an optimistic picture. Industry insiders knew differently. Most labs still relied on bespoke scripts, manual interpretation, and the kind of PhD-level expertise that doesn't scale. Even pharma giants with deep pockets admitted privately that characterization lagged their computational capacity. Data came fast. Making sense of it? That remained stubbornly slow.
The Consolidation Play
Market forces responded with predictable consolidation. Thermo Fisher acquired Olink for $3.1 billion in July 2024, integrating high-plex biomarker platforms into its mass spectrometry ecosystem. Illumina closed a deal worth up to $425 million for SomaLogic and Sengenics proteomics assets from Standard BioTools on January 30, 2026.
The message was unsubtle: the future belonged to companies offering end-to-end solutions, not isolated instruments.
Market projections reflected the urgency, though estimates varied wildly depending on methodology. Grand View Research pegged the proteomics market at $27.8 billion in 2024, forecasting growth toward $58.16 billion by 2030 at a 12.9% compound annual growth rate. The Business Research Company placed the 2025 figure at $39.37 billion. Precedence Research projected the market reaching $162.62 billion by 2035. Which estimate proves accurate matters less than the trajectory—decisively upward. Services, including analytical work, were growing even faster at 13.4% annually, according to MarketsandMarkets.
Instrument vendors sensed opportunity and moved accordingly. Carterra shipped its first Ultra HT-SPR platform in January 2025, explicitly positioning the technology for AI-driven discovery pipelines needing high-throughput epitope mapping. Sartorius launched the Octet R8e in May 2025, promising "unprecedented sensitivity" for label-free interaction analysis—a claim that drew some skepticism from veterans who'd heard similar promises before. Nicoya introduced its Alto Automation Suite in January 2025 to ease SPR integration into biologics workflows.
Each release emphasized speed and throughput, attempting to match computational design's blistering pace.
But faster instruments only magnified the analysis problem. A research team running 32-plex multiplexed proteomics might generate terabytes of data in days. Interpreting it—identifying genuine signals, quantifying modifications, preparing documentation suitable for regulatory submission—still took weeks. Longer if the protein was novel, without established reference standards or precedent.
Someone needed to fix the software layer. The question was who.
Three Scientists Walk Into Y Combinator
This is where 10x Science enters the narrative, though whether they solve the problem or merely add noise remains an open question.
The startup emerged from Y Combinator's Winter 2026 batch, founded by three scientists who understood the bottleneck intimately. Roberts, the CEO, completed postdoctoral work in Bertozzi's Stanford lab—Bertozzi being the 2022 Nobel laureate in chemistry known for bioorthogonal chemistry and glycobiology. Andrew Reiter, the COO, spent time at the Broad Institute's proteomics platform under Steven Carr before pursuing a Stanford PhD co-advised by Bertozzi and Or Gozani. Vishnu Tejus, the CTO, brought two prior Y Combinator companies and AI infrastructure experience from deployments at University of Washington, UCSF, and Stanford.
Their pitch was refreshingly direct: automate the analysis pipeline that currently demands manual intervention at every step. Raw mass spectrometry data goes in; characterization reports suitable for IND filings come out—in minutes rather than weeks. The company claims this automation could save development teams over $150,000 monthly in labor costs, though those figures come from 10x Science's own materials and await independent validation.
Bertozzi's February LinkedIn endorsement of her former lab members lent academic credibility. Y Combinator's public post emphasized the team's proteomics pedigree and market timing—AI flooding discovery pipelines just as characterization became rate-limiting. Whether 10x Science can deliver on the promise remains uncertain. But the timing is notable.
The company launched as regulatory frameworks began formalizing AI's role in drug development, for better or worse.
The Regulatory Wildcards
The FDA proposed its "Considerations for the Use of AI to Support Drug and Biological Product Development" framework on January 6, 2025. The draft guidance encouraged early engagement and noted that more than 500 submissions with AI components had been filed since 2016. The comment period closed April 7, 2025; final guidance hasn't materialized. The document signals where regulation is headed, not where it currently stands—a distinction companies building AI infrastructure must navigate with care.
Europe moved more definitively, as it often does. The EU AI Act entered force August 1, 2024, with prohibited practices and AI literacy requirements effective February 2, 2025. Most obligations apply starting August 2, 2026—months away as of this writing. High-risk systems embedded in medical contexts face the strictest requirements, potentially including protein characterization platforms if they inform clinical decisions. The European Medicines Agency finalized its reflection paper on AI in the medicinal product lifecycle on September 30, 2024, complementing ICH Q14 guidance on analytical procedure development.
Compliance isn't trivial. McKinsey estimated in a January 2026 synthesis that organizations typically spend five dollars on change management for every dollar invested in technology—a sobering ratio for startups promising to transform entrenched workflows. The FDA also launched "Elsa," an internal generative AI tool to aid reviews, in June 2025. Regulators are simultaneously adopting and scrutinizing the technology, which creates... interesting dynamics.
For companies like 10x Science, regulatory evolution presents both opportunity and risk. Automated analytical platforms could streamline FDA interactions if they produce audit-ready documentation. But any AI system touching clinical data must demonstrate reliability, reproducibility, and transparency—requirements that add engineering complexity and validation burden. Not to mention cost.
The Twelve-Month Test
The next year will test whether protein characterization infrastructure can scale to match AI-driven discovery. Generate Biomedicines has multiple AI-engineered therapeutics in human testing. Novo Nordisk showcased that calcitonin-based workflow in March 2024. Absci validated AI-designed antibodies using high-throughput SPR without iterative optimization, according to a Carterra case study. These aren't pilot projects anymore. They're commercial programs with real money behind them.
Yet each success story depends on rigorous biophysical characterization. A 2025 blog post from Oxford Protein Informatics Group noted persistent heterogeneity in affinity datasets, complicating machine learning model generalization. The tools exist to measure binding kinetics, glycosylation patterns, aggregation propensity. What's missing is connective tissue—software that translates instrument output into standardized, regulatory-compliant reports without requiring manual PhD-level interpretation.
And frankly, without exhausting whoever has to do the interpreting.
Deloitte's 2026 life sciences outlook reported that executives expect AI-enabled platforms to be a key growth driver this year. BCG's New Drug Modalities 2025 report highlighted how the shift toward biologics and GLP-1 agonists increases protein analytics demand. A Nature Communications paper from August 2025 described an autonomous enzyme engineering platform completing four design-build-test-learn cycles in four weeks with fewer than 500 variants per round—demonstrating what becomes possible when throughput constraints lift.
The market opportunity is clear. The technical challenge? Substantial. And the competitive landscape is just taking shape, which means it's still anyone's game.
GlycoEra raised $130 million in May 2025 to develop extracellular protein degraders, a modality requiring precise glycoprotein characterization. Industry conferences like the Proteomic-Based Drug Discovery Summit 2026 are dedicating entire sessions to AI platform integration. Bio-IT World 2026 launched an investor and partnering program focused on AI tools for biotech—a sign that capital is starting to flow toward infrastructure plays, not just drug candidates.
Established instrument makers hold advantages in customer relationships and regulatory familiarity. But they also carry legacy software architectures and organizational inertia. Startups like 10x Science enter unencumbered by existing codebases, able to design AI-native systems from scratch.
Whether that agility translates to market share depends on execution. And on whether early customers validate the productivity claims, which remain unproven outside controlled demos.
The Wrong Problem First
What's certain is that AI solved the wrong problem first.
It taught computers to design proteins before teaching them to analyze the ones we make in the lab. The infrastructure is catching up now, unevenly and with substantial capital at stake. Whoever builds the analysis layer that actually works—that handles novel proteins, produces reliable documentation, and scales economically—will capture a meaningful share of a market rushing toward $50 billion and beyond.
The bottleneck shifted. The race is on to unclog it.
And somewhere in a Stanford lab, presumably, a postdoc is still spending weeks characterizing a single protein the old-fashioned way. At least until someone makes the software work.
