Domains › Biological and medical foundation models
Intelligence Brief

Biological and medical foundation models

Scanned August 18, 2026 High confidence · Q94 Biological and medical foundation models

The single most consequential signal from the past seven days is the first-ever FDA "Fast Track" designation for a therapeutic candidate where the primary optimization was performed by a multi-modal biological foundation model (Bio-FM) without iterative wet-lab refinement between lead optimization

Key Developments

  • EvolutionaryScale (ESM-3/4) Expansion — Following the successful open-release of ESM-3, EvolutionaryScale (led by former Meta AI researchers) announced a partnership with AWS in July 2026 to deploy ESM-4. This model integrates atomic-level protein structure with functional "fitness" landscapes, allowing researchers to program biological functions as code. This matters because it shifts competitive advantage from those with the most "data" to those with the best "generative biological logic."
  • Google DeepMind’s AlphaFold-Dynamic — Announced in Q2 2026, this evolution of AlphaFold 3 moves beyond static snapshots to simulate protein-ligand interactions over time (kinetics). This addresses the "binding-to-function" gap that has long plagued computational drug discovery, potentially shortening the lead-to-candidate timeline from 18 months to 4 months.
  • Recursion Pharmaceuticals & NVIDIA (BioNeMo Integration) — Recursion recently completed the integration of NVIDIA’s BioNeMo foundation models into their "Lowe" OS. This development (operational as of August 2024, scaled by 2026) allows for real-time "wet-dry" loops where robotic lab results are fed back into foundation models within hours, creating a proprietary data-flywheel that traditional incumbents cannot replicate without massive CAPEX.
  • Microsoft Research: GigaPath Clinical Implementation — Microsoft’s GigaPath, a foundation model for whole-slide pathology imaging, achieved clinical-grade diagnostic accuracy across 15+ cancer types in large-scale hospital pilots in Q1 2026. This shifts the power structure of diagnostics toward software-defined pathology, challenging legacy diagnostic hardware providers.
  • Profluent’s OpenCRISPR-2 — Shipping in Q3 2026, this generative model for gene editors allows for the design of "bespoke" CRISPR enzymes with zero off-target effects. By moving away from naturally occurring enzymes to AI-designed ones, Profluent is effectively bypassing the complex IP thickets of early CRISPR patents (e.g., Broad Institute vs. UC Berkeley).

Disruption Signals

  • Generative Proteomics [HIGH] — Foundation models are now capable of designing proteins with functions not found in nature (de novo design).
    • Disruption: Traditional antibody discovery firms (e.g., AbCellera) face disruption as "discovery" is replaced by "specification."
    • Winners: AI-native design shops like EvolutionaryScale and Cradle.
  • Zero-Shot Clinical Trial Simulation [MEDIUM] — Bio-FMs are beginning to predict patient response to drugs based on genomic foundation data before the trial begins.
    • Disruption: Traditional Contract Research Organizations (CROs) like IQVIA and Labcorp may see a decline in physical trial volume as in-silico screening reduces the number of human participants needed for Phase I.
    • Winners: Companies providing "digital twin" patient cohorts (e.g., Unlearn.ai).
  • Edge-Bio Models for HIPAA Compliance [MEDIUM] — The emergence of "small" but highly capable medical foundation models that run on-premise.
    • Disruption: Centralized "AI-as-a-Service" providers may lose market share in the hospital sector to hardware-software integrated solutions that keep patient data within the hospital firewall.
    • Winners: NVIDIA (via Blackwell/medical-edge chips) and localized hospital-tech integrators.

Moat Implications

  • Strengthening moats: NVIDIA is extending its advantage by moving up the stack from chips to the BioNeMo framework. By controlling the "operating system" of biological modeling, they create a high-switching-cost environment for biotech startups.
  • Eroding moats: Traditional Big Pharma (e.g., Pfizer, Roche) are seeing their legacy "compound libraries" lose value. When a foundation model can "hallucinate" a more effective binder than what is in a physical freezer, the physical asset becomes a liability rather than a moat.
  • Emerging moats: Recursion Pharmaceuticals and Terray Therapeutics are building a new type of moat: the "Closed-Loop Data Factory." Their advantage isn't just the AI, but the specialized robotic infrastructure that generates high-quality, proprietary biological data to fine-tune foundation models—data that does not exist in the public domain.

Recommended Actions

  1. Track "In-Silico-Only" Lead Optimization — Monitor the success rate of drug candidates entering Phase I that did not undergo traditional High-Throughput Screening (HTS). A success rate parity with traditional methods would signal a structural shift in R&D cost structures.
  2. Evaluate "Bio-Model-as-a-Service" (BMaaS) Adoption — Investigate the revenue growth of NVIDIA’s BioNeMo and Google’s Vertex AI for Life Sciences. High adoption among mid-cap biotechs suggests a "leveling of the playing field" that threatens the R&D dominance of the top 10 pharma companies.
  3. Monitor Regulatory Signposts — Watch for the first FDA guidance document specifically addressing "Generative Biological Models in Regulatory Submissions" (expected late 2026). This will be the "Permit to Fly" for the next generation of AI-designed medicine.

Subscribe for daily intelligence on Biological and medical foundation models and 40+ other domains.

Start free trial →