93% of All Known Protein Complexes, Mapped. Rezo Still Has No Drug.

The Network Biology Thesis: Beyond the Single-Target Model

On November 17, 2022, Rezo Therapeutics emerged from stealth with $78 million in Series A financing, a single round that immediately placed it among the best-capitalized platform plays in the new wave of AI-native biotech, YesPress reported. The capital arrived before the company had a clinical candidate, a fact that underscores the conviction bet: the platform itself is the asset.

Rezo is building a disease-agnostic, AI-driven platform that maps entire disease networks to identify novel drug targets, compressing early discovery timelines and forcing the company to evolve into a full-stack biotech that can take candidates from concept to clinic. The conventional hunt starts with a mutated gene, finds the protein it encodes, and screens for molecules that block it. Rezo widens the frame.

"Proteins rarely act in isolation," the company states in its founding thesis. Its scientists use proximity labeling, mass spectrometry, CRISPR screens, and single-cell genomics to see how mutations rewire interactions across a cell. The output is not a list of suspicious proteins. It is a map of who touches whom, under which disease condition, and what changed.

That middle step — the map — is the wager. A mutation may not create an obvious pocket on one protein, but it may create or destroy an interaction with another. The interface can become a tumor-specific vulnerability, a resistance mechanism, or a biomarker. Rezo calls this Sequence to Systems to Drugs: genetics and proteomics identify disease-linked changes; interaction networks expose new dependencies and interfaces; structure and chemistry turn a network clue into a molecule.

Industry data shows Phase I success rates for AI-discovered compounds running between 80% and 90%, roughly double the historical industry average of 40–50%.

Rezo's platform integrates wet-lab proteomics and genetics with structural biology, chemistry, bioinformatics, and AI. The company began in 2021 around work from the Quantitative Biosciences Institute at the University of California, San Francisco. The founding group includes systems biologist Nevan Krogan, chemical biologist Kevan Shokat (inventor of drugs targeting KRAS and a Howard Hughes Medical Institute investigator), cancer-network researcher Sourav Bandyopadhyay, structural biologist Natalia Jura, and veteran biotech executives George Scangos and Norbert Bischofberger. Scangos took the board chair role; Krogan served as interim chief executive and president while the platform matured.

The company's initial focus is solid tumors, especially disease biology shaped by mutation-specific interactions and drug resistance. Rezo describes multiple therapeutic programs and says its AI platform cuts target discovery time in half. The public record does not yet name a drug candidate or a clinical trial. The chain of evidence — map, interface, molecule, validation — still has to hold in patients. But the thesis is specific: studying the relationship between proteins may reveal choices hidden by studying either protein alone. The lonely culprit made for a clean diagram. The network may make for a better drug.

From Data to Drug: The AI and Data Infrastructure

Rezo's Sequence to Systems to Drugs platform stitches together proteomics, genetics, structural biology, chemistry, bioinformatics, and machine learning into a single pipeline that maps molecular disease networks. The computational backbone leans on NVIDIA's BioNeMo context parallelism framework, which Rezo integrated to predict massive PPIs spanning up to 6,500 residues. That capability covers the vast majority of known protein complexes and unlocks rapid discovery of novel ones.

In benchmark runs, Boltz-2 predictions scaled to approximately 20,000 tokens using 256 NVIDIA H100 GPUs. The team folded a 3,605-residue TTC7A/PI4KA/FAM126A/EFR3A complex across four chains in under five minutes on four H100 GPUs without additional training, producing five structural samples at roughly 54 seconds each while maintaining all long-range inter-subunit contacts within the model context window.

The framework partitions the global (N × N) matrix into a grid of blocks. For a 10,000-residue complex representing 100 million interactions, each GPU manages only a specific sub-block, localizing the memory footprint from O(N²) to O(N²/P) per device.

That architecture shatters the reductionist compromise that has defined computational biology for decades. Researchers previously had to deconstruct complex systems into isolated fragments (single proteins or small domains) to fit GPU memory. The most common workaround sliced sequences into overlapping segments, destroying long-range information and making allostery or signal transduction across entire complexes impossible to model.

Context parallelism takes a different approach. Unlike traditional data parallelism, which assigns each GPU a different protein to fold, CP splits a single massive sample across GPUs. The NVIDIA BioNeMo implementation builds on Torch distributed APIs for GPU-to-GPU communications, implementing a multidimensional sharding strategy for linear capacity scaling. Distributed primitives orchestrate local computation with asynchronous peer-to-peer transfers. Halo-exchange-based primitives partition atom features so subsequent window-batch attention requires no inter-GPU communication, a direct contrast to AlphaFold3's sequence local attention, which limits atom attention to windows of 32 by 128.

The engineering payoff shows in the data: greater than 3x enrichment of CP-resolved, high-quality novel protein complexes compared to predictions using only high-confidence PPIs in the public domain. Collaborators Proxima and Earendil Labs extended the framework further. Proxima embedded CP within their all-atom generative foundation model Neo for inference on assemblies up to 4,000 tokens, while Earendil demonstrated high-fidelity structural predictions at scale with compressed discovery timelines for next-generation biotherapeutics.

But physical capacity alone doesn't guarantee biological accuracy. Current models struggle with high-fidelity folding at scale because they were trained on small fragments. Fine-tuning with larger crop sizes is essential to capture the emergent logic of long-range interactions. Data scarcity remains a primary bottleneck, addressed partly through contributions to the AlphaFold Protein Structure Database. NVIDIA accelerated computing software (cuEquivariance and TensorRT) populates that database with high-throughput predictions of massive homomeric and heteromeric complexes, laying groundwork for foundation models that represent larger biological systems.

Rezo's stack reflects its full-stack ambition: wet-lab proteomics and structural biology generate the data; distributed GPU infrastructure and custom model architectures process it; AI and ML layers extract targetable insights. The integration challenge isn't just technical — it's organizational. Computational biologists, structural biologists, chemists, and ML engineers must share a common language and a unified data layer. That's why the company's current search for a Director of Program Management sits at the intersection of these disciplines, translating computational predictions into wet-lab validation campaigns and clinical candidate selection.

The $78M Launch: Funding, Founders, and the Market Bet

The Series A was led by SR One, Andreessen Horowitz's a16z Bio + Health fund, and Norwest Venture Partners, with participation from SV Angel, Liquid 2 Ventures, Hawktail Management, Alexandria Venture Investments, Eli Lilly and Company, and the UCSF Foundation Investment Company. Nine investors in total.

The scientific founding team reads like a directory of UCSF's Quantitative Biosciences Institute. Krogan, director of QBI, built the protein-interaction mapping infrastructure that underpins Rezo's platform. Shokat brought expertise in chemical biology and kinase targeting. Jura and Bandyopadhyay contributed structural biology and systems-level proteomics. Bischofberger, former chief scientific officer at Gilead, and Scangos, former CEO of Biogen and Exelixis, supplied the drug development and commercial translation muscle.

In September 2023, the board appointed Nadir Mahmood, Ph.D., as chief executive. Mahmood arrived from Nkarta Therapeutics, where he had served as chief financial and business officer. His mandate was to convert the platform's output into a pipeline. By August 2025, the company turned again, this time to Derek Hicks, a 25-year pharmaceutical veteran whose resume spans Pfizer, Spark Therapeutics, and most recently Intellia Therapeutics as chief business officer. Hicks oversaw the Cerevel creation at Pfizer and the Zoetis spinout; at Intellia he negotiated the partnerships that funded CRISPR programs into the clinic.

Demand for AI skills inside biotechnology roles has risen 40 percent over three years. Investors are funding industrialized platforms that can ingest heterogeneous data (proteomics, genetics, structural biology, chemistry) and output clinical candidates at a cadence traditional discovery cannot match.

Rezo's bet is that the integration layer, the "Systems" in its platform name, is where value accrues. Single-modality AI companies compete on model performance; full-stack platforms compete on throughput from data to IND. The $78 million buys the compute, the mass-spec time, the cryo-EM access, and the medicinal chemistry headcount to run that loop internally. It also buys the organizational discipline to keep computational and experimental teams from drifting into separate fiefdoms, a failure mode that has stalled better-funded efforts. The next test isn't scientific. It's operational: can a platform built in an academic institute operate at the velocity and rigor a public company demands?

Scaling the Full-Stack: The Organizational Leap from Discovery to Clinic

The technology worked. The problem is what comes next: turning maps into molecules that clear an IND and enter human trials. That transition is where AI-native biotechs either scale or stall.

The job posting for the Director, Program Management makes the bridge explicit: "As the Rezo pipeline advances, build a high-performing PM team to manage cross functional efforts and direct our therapeutic programs through IND submission and into the clinic." The role sits inside R&D, not operations, and its scope spans Platform, Discovery, CMC, Clinical, Regulatory, and BD. That list is the dependency graph for an IND-enabling package.

The posting demands someone who can "evolve and implement consistent best practices in program management across the organization with clear continuity between functions, especially in the IND-enabling space." That phrase, "clear continuity between functions," is the operational definition of full-stack biotech. The platform generates targets across oncology and protein-protein interaction space; the program manager decides which targets get resourced, in what order, with what data packages, and when the CMC team needs to start process development so the clinical team isn't waiting on drug supply.

The founders (Nevan Krogan, Kevan Shokat, Sourav Bandyopadhyay, Natalia Jura, Norbert Bischofberger, George Scangos) represent the computational and translational depth required to make network biology work. But depth doesn't equal throughput. The next phase isn't scientific; it's logistical. Rezo's company page shows one role added in the past seven days: the Director of Program Management. That single opening is a critical bottleneck. Moving candidates through the clinic depends on filling it.

The Competitive Ripples: How Integrated Platforms Are Reshaping Pharma

The first three weeks of January 2026 set the tone for the year. Eli Lilly signed a deal with Chai Discovery. GSK partnered with Noetik. Pfizer inked an agreement with Boltz.

The market numbers reflect the shift.

More than 200 AI-designed drugs are now in clinical development. An estimated 15 to 20 AI-originated drugs are expected to enter pivotal Phase III trials during 2026. Analysts project a 60% probability that the first AI-designed drug receives regulatory approval by 2027, less than a decade from the first serious generative chemistry papers to a marketed drug, an extraordinary compression of the typical 15-year discovery-to-approval timeline.

Software licensing deals now span 13 of the world's top 20 pharma companies.

The talent market is already pricing the shift. Industry analysis indicates a 40% increase in demand for AI skills within biotechnology roles over the past three years. The role of AI Program Manager (bridging executives, AI engineers, cybersecurity professionals, software developers, legal teams, and business stakeholders) has emerged as one of the highest-stakes and fastest-growing positions in the sector. Rezo's own hiring of a Director of Program Management illustrates the pressure: the company needs someone who can translate computational target maps into wet-lab workstreams and clinical milestones without losing signal at the interfaces.

Regulatory infrastructure is racing to catch up. The FDA issued its inaugural draft guidance on AI intended to support regulatory decisions about drug safety and efficacy in January 2025. Most current AI applications sit in the clinical research phase: trial design, endpoints, biomarker development, real-world evidence generation, and post-market surveillance.

The FDA's move did not arrive in a vacuum. The agency has cited more than 500 drug and biologic applications with AI components between 2016 and 2023, concentrated heavily in oncology and neurology. The pace accelerated sharply, from roughly 29 applications with AI/ML components in 2019 to more than 100 in 2021 alone. The May 2023 discussion paper that preceded the draft guidance drew over 800 public comments, which FDA said directly informed the January 2025 structure. The two centers most affected (CDER and CBER), along with CDRH and the Office of Combination Products, published a coordination paper in March 2024 (revised February 2025) titled "Artificial Intelligence and Medical Products: How CBER, CDER, CDRH, and OCP Are Working Together." Concurrent with the draft guidance, CDER and CBER collaborated with the European Medicines Agency to develop 10 Guiding Principles of Good AI Practice, signaling transatlantic convergence.

The credibility framework demands that sponsors define the context of use with precision, assess risk based on the impact of model failure and the level of human oversight, gather credibility evidence proportional to that risk, document development and validation processes, and implement life cycle management plans for ongoing monitoring. Documentation expectations run consistently across FDA AI documents: data provenance, model development choices, and validation results must be traceable enough that an outside reviewer (an FDA inspector or a future acquirer) could assess whether the model does what it claims. That bar is materially higher than the internal validation practices many discovery-stage tools were built against, particularly generative models never intended to support a regulatory claim.

A practical gap remains. The predetermined change control plan (PCCP) lets a sponsor pre-specify, at initial submission, the kinds of updates an AI model will need and the validation methodology for implementing them, so later modifications within that scope do not require a new submission. But PCCP was built under statutory authority added by the 2022 Food and Drug Omnibus Reform Act specifically for AI-enabled device software functions under CDRH. No equivalent, formally named PCCP pathway exists yet for drugs and biologics under CDER or CBER. The January 2025 draft guidance references the device-side PCCP only as related context.

International regulators are moving in parallel. The U.K.'s MHRA post-market surveillance regulations entered force June 16, 2025, requiring manufacturers to actively track safety and performance, report serious incidents on shorter timelines, and maintain periodic safety update reports. The E.U. AI Act treats most AI medical devices as "high-risk" systems, mandating strict compliance with risk management, data governance, technical documentation, transparency, human oversight, and post-market monitoring. Health Canada encourages reporting of demographic distribution in training and testing datasets, a noted limitation in FDA submissions where only 23.7 percent reported dataset demographics.

Rezo's initial focus on solid tumors and protein-protein interactions places it squarely in the oncology therapeutic area where FDA's Oncology Center of Excellence launched its AI Program in 2023, responding to growing reviewer interest and sponsor engagement. The OCE's earlier Integrated Project Team (2017) built a data platform integrating approximately 30 internal FDA systems and external HHS websites, infrastructure that now supports AI-enabled oncology review. As Rezo advances candidates from its network-mapping platform toward IND-enabling studies, the credibility framework will shape how its computational evidence is packaged, documented, and defended. That search for the Director, Program Management reflects the operational reality: someone must bridge the computational teams building credibility evidence and the wet-lab teams generating the safety and efficacy data that regulators ultimately judge. The first IND filing will test whether Rezo's full-stack model can translate network-biology insight into a regulatory package that survives FDA's risk-based scrutiny.

On the 2024 CORUM collection of known human protein complexes, Rezo reported that a serial Boltz-1x model could score less than 30 percent. Most of the map was technically present and practically unreachable.

Share of CORUM complexes within reach

Serial

<30%

Fold-CP

93%

The response, developed with NVIDIA's BioNeMo team, was to change how the workload ran. Fold-CP distributes the model's context across accelerators, while cuEquivariance speeds the geometry-heavy operations used in molecular modeling. Rezo says the combined approach produced high-quality predictions for complexes as large as 6,500 residues and brought 93 percent of CORUM within range, YesPress's data shows. That is not proof of a drug. It is proof that the company can ask its model about a much less convenient class of biology.

The unglamorous economics of asking bigger questions

Bigger questions produce bigger cloud bills. Rezo's structural workloads may need thousands of CPU instances or hundreds of specialized accelerators, but not forever. Keeping that capacity idle would be absurd. Finding it on demand across regions, surviving interruptions and moving terabytes of sequence data without maintaining a large infrastructure team is the real engineering problem.

500TPUs scaled dynamically across four regions

**90%+**reported storage savings for one workflow

**67%**reported compute savings through orchestration

With Union.ai orchestrating workloads on Google Cloud, Rezo uses discounted spot instances with retries and an on-demand fallback. For multiple-sequence alignments, it provisions local solid-state drives, downloads about 1.5 terabytes from cloud storage, runs the batch and releases the hardware. The company says that change took one monthly bill from roughly $10,000 to hundreds of dollars. Other workloads dynamically expand from zero to 500 TPUs across four regions.

The copyable idea is not “use more AI.” It is to match every stage of a scientific workflow to the cheapest suitable hardware; treat interruption as a design assumption; cache work that should not be repeated; and delete idle capacity. A small biotech can rent the shape of a supercomputer for an afternoon without becoming a data-center operator.

A useful boundary

This pattern works best for large, bursty, fault-tolerant batch jobs. It is less compelling for small or steady workloads, latency-sensitive experiments, pipelines that cannot resume safely, or regulated data that cannot move across regions. Cheap prediction also does not replace wet-lab validation. A beautifully folded hallucination is still a hallucination.

Derek Hicks arrived in 2025 with a mechanical engineer's degree and a dealmaker's résumé. The proteins may network naturally; the partnerships require a human.

A new operator, and the oldest biotech test

In August 2025, Rezo appointed Derek Hicks chief executive. Hicks had spent more than 25 years across Pfizer, Spark Therapeutics and Intellia Therapeutics, much of it in business development. Krogan, the founding CEO, moved to President. The change is legible: the academic platform had been built far enough to need an operator experienced in partnerships, portfolio choices and the long handoff from discovery into development.

Rezo describes multiple therapeutic programs and an initial focus on solid tumors, especially disease biology shaped by mutation-specific interactions and drug resistance. Its users today are its own biologists, chemists and computational scientists. Potential biopharma collaborators are the external customer. Patients come last, and only if the chain of evidence holds.

That is the old test beneath the new machinery. Network maps can improve target selection. AI can make enormous complexes inspectable. Elastic cloud infrastructure can buy more predictions for the same dollar. None of those conveniences can tell us whether a molecule will be safe and effective in people. They can only improve the quality and quantity of the bets entering that test.

Rezo's most interesting claim, then, is modest by biotech standards. It is not that the company has solved drug discovery. It is that studying the relationship between proteins may reveal choices hidden by studying either protein alone. The lonely culprit made for a clean diagram. The network may make for a better drug.

Follow the network

Rezo website LinkedIn updates Company news Meet the team Higher-order complexes [Google Cloud interview](https://www.youtube.com/watch? v=378S2ajmGno)

Lessons for Frontier Tech: What Full-Stack Platforms Teach Us

The version-control system that software engineers take for granted (commit history, branching, rollback) barely exists in biology. A 2022 Nature Communications paper introduced CellRepo, a cloud-based version control system that links digital engineering records to living cell lines through chromosomal barcodes. The authors argue that wider adoption of version-control principles could substantially improve research quality. The gap is not theoretical: engineering a synthetic biology agent produces terabytes of data, yet only a small percentage becomes available when results are published. That gap fuels a reproducibility crisis, high-profile retractions, and weakening public trust. The lesson for any high-stakes domain: if you cannot trace the exact lineage of an artifact (code, cell line, satellite subsystem), you cannot audit it, reproduce it, or regulate it.

The platform embodies the same architectural pattern. Its stack mirrors the four-layer model that Intuition Labs documented across modern biotech: infrastructure (cloud/HPC), data management (FAIR-compliant, heterogeneous omics), applications (ELN, LIMS, clinical systems), and analytics/AI. Benchling's 2023 survey of 300 R&D and IT professionals found that 53 percent of scientists use more than five scientific applications daily and 84 percent rely on custom-built software. Organization-wide interoperability sits at 28 percent; reusability at 30 percent. The "spaghetti" dilemma — stitching point solutions without a unifying platform — is not unique to biotech. Any domain fusing physical experimentation with heavy computation faces it: robotics, advanced materials, quantum hardware. The fix is not more tools; it is a deliberate platform layer that makes data consumable across the organization.

Talent is the bottleneck. Intuition Labs reports that finding staff who understand both the domain science and data systems is hard. Rezo's own hiring reflects this: a Director, Program Management role sits explicitly at the intersection of computational and wet-lab teams. The same hybrid profile — domain expert plus systems thinker — appears in every frontier sector. Companies that treat this as a recruiting problem instead of an organizational design problem will keep losing cycles to translation friction.

Regulatory compliance must be designed in, not bolted on. The FDA's 21 CFR Part 11 places responsibility for electronic-record integrity on the regulated entity, not the vendor. Selecting a compliant platform does not discharge that obligation. Intuition Labs notes that quality procedures and electronic-system controls must be assessed for their intended use rather than assumed out-of-the-box. Teams that instrument their pipelines for auditability from day one move faster at submission time.

AI is an assistive layer, not magic. Recursion reported capturing "80% of the value with 40% of the wet lab work" through AI integration, but Intuition Labs cautions that generative-AI outputs must be carefully validated experimentally. The pattern holds across domains: the model proposes, the physical test disposes. Over-reliance on simulation without a closed experimental loop produces brittle systems — whether the output is a drug candidate, a flight controller, or a fusion plasma configuration.

The platform-to-product transition is a strategic fork. Bowtied Biotech observes that platform companies either become product companies or fall into fee-for-service models. Rezo's $78M Series A and its disease-agnostic network-mapping platform put it on the first path, but the organizational leap from discovery to clinic — hiring program management, building clinical data systems, engaging regulators — is where many platform ventures stall.

Build digital foundations early, but match investment to data volume and regulatory milestones — not funding round labels. Intuition Labs emphasizes that Series C is a financing event, not a technical milestone. Over-engineering the stack before the workflow exists wastes resources; under-investing creates technical debt that compounds when the first IND-enabling study starts. The discipline is to prioritize systems by assay volume, trial phase, and manufacturing plan, then defer everything else.

The same capital flood is hitting every frontier domain. The winners will not be the ones with the biggest models — they will be the ones whose version control, data contracts, and regulatory instrumentation let them turn model output into physical reality, repeatedly and audibly, at scale.

The Director of Program Management they're hiring now will decide whether those maps become medicines.

Working in frontier tech? Zero G Talent tracks the openings: see every open Rezo Therapeutics role, browse frontier tech jobs, openings at Leap, and the people building the field.