Newsletter · · Ashutosh Agarwal

AI Drug Discovery - Week of June 11, 2026: An Open-Source Protein Model Designs Validated Antibodies

AI drug discovery newsletter for the week of June 11, 2026. The platform debate moved on real insider tape: a CZ Biohub protein language model designed cryo-EM-validated nanomolar antibodies, NVIDIA called simulation assay-grade and named Schrodinger, and the first AI-formulated oral drug entered Phase 1.

AI Drug Discovery Weekly

Week of June 11, 2026: An Open-Source Protein Model Designs Validated Antibodies

This week the platform debate finally produced physical output rather than slideware. The clearest signal came from an insider conversation outside the listed coverage names: a CZ Biohub protein language model that designs single-chain antibodies with nanomolar binding affinity, cryo-EM validated. Layered on top, NVIDIA's biopharma BD lead called free-energy-perturbation simulation assay-grade and named Schrodinger as the live example, the first self-described AI-formulated oral drug entered Phase 1 with regulators leaning in, and a cancer foundation-model pitch reframed the commercial wedge as patient selection rather than molecule design. Here is what the tape said.

TL;DR

  • On No Priors, Mark Zuckerberg and CZ Biohub's Alex Rives walked through an open-source protein language model that designs nanomolar single-chain antibodies, cryo-EM validated, a direct read-through to the wet-lab-first antibody platforms (ABCL, ABSI).
  • NVIDIA's biopharma BD lead said FEP molecular simulation is now "basically as good and as predictive as running experimental assays," and name-checked Schrodinger's tox panel as the live example, the most pointed validation SDGR's physics-based pitch has gotten on a podcast in weeks.
  • CRDMO Quotient Sciences began a Phase 1 study of what it calls the first AI-formulated oral drug after MHRA clearance, while a Pistoia Alliance poll showed 42% of clinical-trial respondents already see early ROI even as trust and regulatory uncertainty remain the top adoption barrier.

What's new

1. Zuckerberg and CZ Biohub open-source a protein model that designs validated antibodies. On No Priors' Biohub: The Future of Biology is Open-Source (Jun 10), Head of Science Alex Rives (ex-Evolutionary Scale) described a protein language model trained on billions of sequences that hits state-of-the-art structure prediction "especially on protein-protein interactions and protein-antibody interactions, which is really critical for therapeutic design," and has been used to design single-chain antibodies with nanomolar binding affinities validated by cryo-EM. Rives framed antibody design as an "emergent property" of a general protein model. Why it matters: an open-source engine that replaces "hundreds of thousands or millions of antibodies" in a screen with a compute instance is a structural threat to the moat narrative under AbCellera and Absci. Mark Zuckerberg and Priscilla Chan positioned the release as a discovery engine free to anyone.

2. NVIDIA says simulation now rivals the wet lab, and points at Schrodinger. On Once a Scientist Ep. 95 (Jun 10), Stacie Calad-Thomson, NVIDIA's head of BioPharma Labs & Manufacturing BD, said FEP molecular simulations are "basically as good and as predictive as running experimental assays now," citing Schrodinger's tox panel running ~100 off-target tox predictions computationally. She described NVIDIA as an ecosystem enabler, hosting open-source models on BioNeMo, with pharma "licensing those models" and fine-tuning on proprietary data: "we've seen a lot of deals this year." Why it matters: this is third-party validation of SDGR's core claim from the most credible compute vendor in the space. She also flagged the real bottleneck as lab-instrument connectivity, not compute, a tell on where the next capex goes.

3. First "AI-formulated" oral drug enters Phase 1; regulators lean in. On The Drug Discovery World Podcast, DDW Highlights, 9 June 2026, a Pistoia Alliance poll found 50% of clinical-trial respondents cite trust and regulatory uncertainty as the top barrier to AI adoption, while 42% already see early ROI. CRDMO Quotient Sciences began a Phase 1 study of what it calls the first AI-formulated oral drug after MHRA clearance; MHRA, Danish and Swedish regulators signaled openness, with Pistoia's Dr. Becky Upton stressing "validated, auditable and explainable approaches, not black box models." Why it matters: the bottleneck for AI-bio multiples is regulatory acceptance, and the agencies are now publicly constructive.

4. A pan-cancer "foundation model" pitch for trial enrichment. On BIO from the BAYOU Ep. 140 (Jun 10), Genialis CEO Dr. Rafael Rosengarten described an RNA-seq-based cancer foundation model that decomposes biology into modular "Lego brick" models, letting clinical predictors train on as few as 50 patients, aimed at "the most pernicious bottleneck in drug development," trial failure. Why it matters: this is the commercial wedge, using AI to pick the right patients rather than design the molecule, and it is the version of "AI in pharma" with the clearest near-term P&L.

The debate

Bull: The platform is finally producing physical output, not slideware. Validated nanomolar antibodies from a model, FEP simulation at assay-grade accuracy, a first AI-formulated drug in the clinic, and regulators actively writing the rulebook. If the cost of a credible drug candidate keeps falling, the public picks-and-shovels names (compute, physics-based software, biosimulation) compound while discovery economics improve industry-wide.

Bear: The week's best evidence came from a non-profit (CZ Biohub) and a chip vendor (NVIDIA), not from the listed AI-bio names whose stocks need the wins. Open-sourcing the antibody-design engine is precisely the commoditization risk that compresses the moat under ABCL and ABSI. And the Pistoia data is blunt: half the industry still doesn't trust the tools. Validated binders and AI formulations are not approved drugs or revenue. With RXRX at $3.15 and SDGR at $14.60, both near 52-week lows and burning cash, the market is pricing skepticism, not a platform re-rate.

Stocks in play

  • Recursion (RXRX), $3.15, +3.6% on the day, market cap ~$1.4B, near the $2.77 52-week low (high $7.18). Bull: scaled phenomics plus compute, optionality on pipeline readouts. Bear: cash burn against a sub-$1.50 EPS loss; the open-source-model news cuts against proprietary-data moats. Next catalyst: clinical pipeline readouts and any partnership refresh.
  • Schrodinger (SDGR), $14.60, +3.2%, market cap ~$1.1B, 52-week range $10.95–$26.45. Bull: NVIDIA publicly validated FEP-as-assay and cited SDGR's tox panel by name; software-plus-pipeline model. Bear: stock sits near the low; proprietary-program value still unproven; simulation parity claims help the category, not necessarily the licensor. Next catalyst: software bookings cadence and proprietary pipeline progress.
  • Eli Lilly (LLY), $1,160.95, +2.2%, market cap ~$1.09T. Bull: retatrutide TRIUMPH-1 showed ~28.3% weight loss at 80 weeks plus knee-OA and sleep-apnea benefit; Jefferies lifted its target to $1,350 (Buy); $1B+ AlzeCure Alzheimer's licensing deal; Ebglyss atopic-dermatitis approval. Bear: the tape was GLP-1 and licensing, not AI; GLP-1 employer-coverage cuts in 2027 (~10% of covering employers) are a demand overhang. Next catalyst: GLP-1 coverage policy and any TuneLab external-partner disclosure.
  • NVIDIA (NVDA), the week's connective tissue: BioNeMo hosting, "a lot of deals this year," and the FEP-parity claim make NVDA the toll-taker on every AI-bio workflow. Watch: model-licensing deal flow as a leading indicator for the category.

Read-throughs

  • Antibody-discovery platforms (ABCL, ABSI): the CZ Biohub open-source antibody model is the competitive read-through to watch, generative design as a free utility pressures differentiated-data moats.
  • Compute and infrastructure (NVDA): ecosystem-enabler positioning and rising model-licensing deal volume; lab-instrument connectivity flagged as the next bottleneck (read-through to lab-automation and informatics vendors).
  • Physics-based and biosimulation (SDGR): FEP-as-assay validation lifts the category narrative.
  • AI diagnostics and trial enrichment: the clearest near-term revenue path is patient selection and trial design, not de novo molecule generation, per the Genialis cancer foundation-model pitch.

What changed vs last week

This is the first edition of AI Drug Discovery Weekly, so there is no prior issue to reconcile against; treat this as the baseline. One framing to carry forward: this week the most consequential AI-bio progress came from a non-profit (CZ Biohub) and a compute vendor (NVIDIA), with the regulatory door opening via the first AI-formulated drug in the clinic. We will track next week whether the public names start converting that platform momentum into disclosed wins, or whether open-source commoditization keeps the value capture upstream of the equities.