Some VCs have told me to “pick a lane”, here’s why I think this is fundamentally wrong

The “pick a lane” rule in biotech, where companies focus on either diagnostics or therapeutics but not both, is a holdover from a 1995 industry structure where the two businesses had incompatible cost structures, regulatory pathways, and capital profiles. Modern AI agents, sequencing infrastructure, and matching engines built on multi-modal patient data have collapsed those barriers. The integration of clinical data, AI matching, and therapeutic discovery has become the value.

The integrated vision is the only way to build the future of medicine. AI compounds the value of integrated data in ways that were not possible even five years ago, and splitting the diagnostic and therapeutic functions across separate companies destroys exactly the loop that AI now thrives on. Most major pharma companies already understand this, which is why they are spending billions through M&A to assemble the same integrated capability that vertical specialists like NeuroAge are building natively from day one. The pattern is visible to anyone watching the M&A flow, but the structural advantage goes to the companies that built it natively rather than assembling it after the fact.

This article walks through how the two lanes ended up separated in the first place, why integration is the better model, what pharma is doing about it, and which companies are building the integrated approach.

How we got here

The diagnostic-therapeutic split is an industry structure, not a biological one. It came from a century of regulatory, reimbursement, and capital decisions that were never planned to produce the system we have now. The FDA built two separate review pathways, reimbursement runs through different fee schedules, sales channels split between lab directors and prescribing physicians, and two specialized investor pools grew up around drug companies and clinical labs.

A company trying to operate in both lanes in 1995 needed two of every functional team across sales, contracting, regulatory affairs, market access, and compliance. The integrated company carried roughly twice the operating cost of a pure-play company, with no obvious revenue offset until the data and feedback loops matured enough to compound, which could take years.

By the early 2000s, the two lanes were so deeply institutionalized that it was hard to remember they were the result of historical accidents. The biology that generates the signal you measure is the same biology that produces the target you drug. There is no scientific reason to separate these into different companies, only the historical accident of how the institutions surrounding biotech grew up.

Investors who concluded from this that complexity was bad were not making a category mistake. In 1995, complexity in biotech was a tax. Two functional teams added cost without adding strategic capability, because the data and AI infrastructure that would have made integration valuable did not exist at sufficient scale. The “pick a lane” rule encoded a real observation about that era.

The error is applying 1995 rules to 2026 conditions, where complexity has become a moat rather than a tax. When integrated structure produces strategic value through data feedback loops and AI matching engines and the operational cost is shrinking exponentially with agentic AI, the complexity is justified. Competitors trying to replicate the moat through partnerships across separate companies inherit the friction of the split structure without the integrated value.

Why integration is the better model

Consider what happens when a drug company runs a Phase 2 Alzheimer’s trial in the traditional model. They enroll 200 patients with moderate disease and an MMSE score of 18 to 24. With no characterized database to draw from, they screen 3 to 5 patients for every one who qualifies. Screen failure rates run 60 to 80 percent, and each screen costs $3,000 to $6,000 for amyloid PET alone, or $5,000 to $10,000 for the full workup. The all-in per-enrolled-patient cost runs $20,000 to $50,000 in Phase 2, climbing to $50,000 to $150,000 in Phase 3.

Even with this expensive screening, the population is heterogeneous because the inclusion criteria are clinical rather than biological. Some patients have amyloid pathology driving their cognitive decline, some have vascular contribution, some have mixed pathology, and some have inflammation as the primary driver. The trial averages all these patients together and gets a mixed result. Maybe 30 percent respond clearly and 70 percent do not, but the trial reports the average and the drug looks marginal.

In the integrated model, the company already has multi-modal biomarker data on hundreds of thousands of patients. They can identify the biological signature of the responder population before the trial and pre-qualify patients from the existing cohort. Screen failure rates collapse from 70 percent to 20 percent or below. They enroll 200 patients with the relevant signature, drawn from a database of consenting patients whose amyloid status, APOE genotype, hippocampal volume, cognitive trajectory, and inflammation markers are already known. Per-enrolled-patient cost drops from $20,000 to $50,000 down to $5,000 to $15,000, a 60 to 80 percent reduction. They see a clean signal at smaller sample size and reach Phase 3 with a known responder population and a companion biomarker already validated.

Phase 2 to approval probability today is around 10 percent. The integrated model can plausibly reach 30 to 50 percent. Development cost per approved drug, currently around $2.6 billion, can fall by 60 to 80 percent. Compounded over a portfolio, this is the difference between bringing 5 drugs to market and bringing 50.

The deeper advantage is the feedback loop. In the traditional model, diagnostic data and therapeutic outcomes live in different organizations and never inform each other systematically. In the integrated model, every patient outcome updates the matching engine and the predictive models, and each drug response updates target prioritization for the next program.

Why AI changes the math

The current generation of AI models requires data scale that no single drug program can generate. Foundation models for biology need millions of patient observations across multiple modalities to learn the patterns that predict disease progression and treatment response. Multi-modal AI systems need integrated datasets where the same patients have data across all modalities, not separate datasets that have to be linked retrospectively. Causal inference at scale needs longitudinal data with consistent measurement methodology over time. Only the integrated company can generate this kind of dataset.

The capability gap between AI-enabled and non-AI-enabled drug discovery is widening. Recursion identifies drug hits in weeks for thousands of dollars where traditional approaches take years and millions per program. Insitro nominates targets from cellular AI models that human researchers would not have considered. Tempus runs matching engines at oncology scale that no individual oncologist could replicate. These integrated companies are doing things that pure-play diagnostic or pure-play therapeutic companies cannot do at all.

AI agents also collapse the doubled-overhead problem that made integrated companies expensive in earlier eras. The 1995 hybrid company needed two sales forces, two regulatory affairs teams, two market access groups, two contracting teams, and two compliance infrastructures because each function required specialized human expertise that did not transfer across the diagnostic-therapeutic divide. AI agents now handle large portions of regulatory submission preparation, contracting analysis, market access modeling, payer negotiation support, and compliance monitoring. A single integrated team augmented by AI agents can do what previously required two specialized human teams, often at a fraction of the cost.

The commoditization of diagnostic wet lab infrastructure compounds this advantage. Whole genome sequencing, whole transcriptome sequencing, MRI imaging, and standard laboratory assays are now available as commodity services from third-party providers competing on price and turnaround time. The integrated company captures value at the data integration layer, the AI matching layer, the longitudinal cohort layer, and the therapeutic discovery layer, where in-house preclinical wet lab capacity validates targets identified from the data. The diagnostic supply chain becomes orchestrated commodity infrastructure rather than a capital-intensive internal operation.

This is why pharma is racing right now. Five years ago, the pharma view was that AI would augment existing drug discovery workflows. Today, the view is that AI-enabled discovery is its own paradigm and that the companies with the best integrated data infrastructure will win. The TuneLab platform from Lilly, the Tempus partnerships across pharma, the BMS-Insitro collaboration on ALS, and the wave of similar deals announced in 2025 and 2026 all reflect the same recognition.

The tech parallel

The dominant tech companies share an architectural pattern. They integrate data feedback loops with decisions inside a single company. Google’s PageRank algorithm could be copied within months, but the click-stream data accumulating across billions of queries built a moat that twenty years of competitors have not overcome. Tesla collects real-world driving data from millions of vehicles continuously, training autopilot models that deploy back to the fleet in a loop Waymo cannot match. Netflix uses viewing data to inform which originals get greenlit, against competitors who license content and run analytics on separate platforms. Splitting the data from the decisions, in any of these cases, would have broken the loop and let unified competitors win.

Pharma has been operating in the split alternative for forty years, with diagnostic data sitting in lab corps and EHR vendors, therapeutic decisions sitting in drug companies, and patient outcomes scattered across hospital systems, payer claims, and pharmacy records that none of the original parties can access in real time. There is no Google equivalent in pharma because the structure that lets data and decisions inform each other has not existed at scale. The integrated vertical specialist model closes this loop, and once closed, the data compounds the way Google’s did. Telling founders to pick a lane between data infrastructure and therapeutic discovery is structurally the same advice as telling Larry Page in 1998 to pick between owning the search index and owning the ranking algorithm.

How pharma is converging on this model

Pharma already understands that the integrated model wins. The largest drug companies have spent the past decade trying to acquire or partner their way into integrated diagnostic and data capabilities, often paying multi-billion-dollar premiums to do it.

Roche acquired Flatiron Health for $1.9 billion in February 2018, buying the leading oncology real-world evidence platform after already holding a 12.6 percent equity stake. Roche then completed its acquisition of Foundation Medicine, the leading cancer genomic profiling company, for an additional $2.4 billion in June 2018, building a fully integrated stack of cancer drugs, real-world cancer data, and tumor genomic profiling. The combined assembly is essentially a vertical specialist for oncology, built through M&A rather than constructed natively as a single integrated company.

Eli Lilly has taken a different path with TuneLab, a platform that gives biotech partners access to machine learning models trained on Lilly’s proprietary preclinical data. Bristol Myers Squibb has built integrated capability through partnership, including a major collaboration with Insitro on novel ALS targets. Similar partnerships exist between Pfizer and Tempus, GSK and Tempus, and Novartis with multiple AI drug discovery platforms.

Pharma is converging on the structure that VCs are still telling biotech founders to abandon. The companies pursuing integrated diagnostic-therapeutic models from day one are building exactly what pharma is paying billions to assemble through M&A. The native integrated builders should have a structural advantage over the M&A-assembled stacks, which are glued together across acquired entities with separate cultures, systems, and incentives.

The vertical specialists building this natively

A new generation of companies is constructing the integrated diagnostic-therapeutic model from scratch, vertical by vertical. Each one combines longitudinal patient data, clinical biomarker assays, AI-enabled matching, and therapeutic discovery in a single integrated entity.

Tempus

Tempus is the most advanced vertical specialist in the integrated model and the proof point that the model can reach public-market scale. The company was founded in 2015 in Chicago by Eric Lefkofsky after his wife’s breast cancer diagnosis, motivated by the recognition that her treating physicians had no systematic way to learn from the experiences of patients with similar profiles.

The company built four integrated layers. The first is clinical and molecular oncology data, aggregated from a network covering more than 65 percent of US academic medical centers and more than 55 percent of US oncologists. The second is genomic and molecular profiling, with Tempus running its own diagnostic assays and integrating results with clinical context. The third is AI-enabled clinical decision support that surfaces relevant outcomes and trial matches at the point of care. The fourth is therapeutic discovery work conducted in partnership with major pharma companies.

Tempus completed its IPO in June 2024 at a $6.1 billion valuation and reported approximately $1.27 billion in 2025 revenue, up 83 percent year over year. 95 percent of top pharma companies by revenue are customers. Tempus has expanded beyond oncology into cardiology, neuropsychology, radiology, and pharmacogenomics for depression, validating that once the integrated infrastructure is built in one therapeutic area, extending it to adjacent areas is faster and cheaper than building from scratch.

NeuroAge Therapeutics

NeuroAge Therapeutics, the company I founded in 2022, applies the integrated vertical specialist model to brain aging and neurodegenerative disease. The platform integrates multi-modal data on each patient, including brain MRI volumetrics, whole genome RNA sequencing, whole genome DNA sequencing, and longitudinal cognitive testing through the NeuroGames assessment. Each client generates a personalized brain aging trajectory rather than a population-average estimate, with the multi-modal data enabling identification of the specific biological drivers of their cognitive change.

The supply chain reflects where value accumulates. NeuroAge’s proprietary assets are the RNA biomarker panel and the AI algorithms, while sequencing runs through third-party partners and MRI through commodity imaging facilities. The company captures value at the data integration layer, the AI analysis layer, the longitudinal cohort that grows with each patient, and the consumer-direct patient relationship. This is structurally similar to how successful software companies operate on top of cloud infrastructure rather than building their own data centers.

The therapeutic discovery side runs differently. NeuroAge operates an in-house preclinical wet lab focused on validating drug targets identified from analysis of the consumer platform data. Consumer patients generate the multi-modal data that the AI analyzes for biological signatures of accelerated brain aging. The signatures point to specific molecular targets that drive aging in identified patient subpopulations, and the wet lab validates these targets through cellular and animal model work. The same data identifies which patients are most likely to respond to which therapeutic candidates, enabling trial enrichment for partner pharma sponsors and eventually for NeuroAge’s own clinical programs. The diagnostic platform and the therapeutic discovery platform are two ends of the same value chain, and the integration is what allows targets to be discovered, validated, and ultimately tested in the patients most likely to benefit.

The brain aging vertical has structural advantages for the integrated model. Brain aging is a slow process with measurable biomarkers years to decades before clinical symptoms emerge, which means the data layer can support both prevention research and earlier therapeutic intervention than the traditional Alzheimer’s drug development paradigm allows. APOE4 carriers, the population at highest genetic risk for late-onset Alzheimer’s, represent a clear high-value target for the integrated approach. A 2019 paper that I co-authored out of MIT showed that being biologically five or more years younger than chronological age confers approximately sixfold lower Alzheimer’s risk even in APOE4 carriers, a finding that only emerges with longitudinal multi-modal data on aging trajectories.

Cardiometabolic disease, autoimmune disease, mental health, and rare disease all need their own integrated diagnostic and therapeutic vertical specialists.

What this means for the next decade

The integrated diagnostic and therapeutic vertical specialist model has been validated by Tempus at public-market scale in oncology. The next wave of companies will extend the model into brain aging, cardiometabolic disease, autoimmune disease, mental health, and rare disease.

Each successful vertical specialist makes the model more recognizable to investors, more familiar to pharma partners, and more credible to patients considering whether to share their privacy protected and anonymized data and participate in trials. Tempus reaching public-market scale provides the comparable that crossover funds and tech-adjacent capital need to underwrite the next generation of vertical specialists.

For patients, the integrated model means faster, cheaper drug development with higher success rates. The same R&D budget can advance more candidates through clinical development, including for rare diseases that historically failed to recruit enough patients to run a trial at all. Earlier market entry means longer effective patent life, which gives sponsors room to price drugs lower while still recovering their development costs. Better patient matching means fewer patients enrolled in trials of drugs that will not work for them, and more patients reached by the drugs that will.

For pharma, the integrated model is no longer optional. The companies that build or acquire integrated capability will compete effectively for the next decade. The companies that try to maintain the old separation will be at a structural disadvantage that compounds over time.

The “pick a lane” advice is a holdover from an industry structure that no longer reflects the underlying biology or the capital available to support its integration. The next decade of medicine will be defined by the vertical specialists who built the integrated model natively rather than the assemblies that pharma is now paying billions to put together after the fact.