Newsletter · 2026-06-25 · Ashutosh Agarwal

AI Drug Discovery - Week of June 25, 2026: Physics Beats Language at BIO 2026, AlphaFold's Jumper Joins Anthropic

AI drug discovery newsletter for the week of June 19 to 25, 2026. BIO 2026 elevated physics-based virtual screening (Sandbox AQ with NVIDIA BioNeMo) over language models, AlphaFold architect John Jumper left DeepMind for Anthropic, and Eli Lilly dominated the deal tape while pure-play techbio names stayed quiet.

AI Drug Discovery Weekly

Week of June 25, 2026: Physics Beats Language at BIO 2026, AlphaFold's Jumper Joins Anthropic

TL;DR

• The week's center of gravity was BIO 2026 in San Diego, where Sandbox AQ broke news of a physics-based virtual-screening partnership with NVIDIA BioNeMo, and where the louder argument was that drug discovery needs to move beyond language models entirely.
• John Jumper, the Nobel laureate behind AlphaFold, left DeepMind for Anthropic, and gave a rare interview reframing AlphaFold as a narrow, engineered predictor rather than a general biology AGI.
• Eli Lilly (LLY) dominated the deal tape: a $1.9B Abbisko research pact, the closed $7.8B Centessa acquisition, stalking-horse bids for bankrupt Sangamo's genomic platforms, a BioArctic neuro deal, and a controversial single-patient compassionate-use program for its next-gen obesity drug Reditrutide.
• The pure-play AI techbio names (RXRX, SDGR) were quiet, no company-specific podcast or news flow this week, but both ground higher on a healthcare small-cap bid.

What's New

Physics-based screening goes mainstream at BIO. Sandbox AQ CEO Jack Hittery announced live from the conference floor a tie-up letting drug companies "use a combination of BioNemo, which is a really strong platform from NVIDIA for the biopharma market, and AQ State" to screen GPCR-targeting ligands, the same receptor class as Ozempic and Mounjaro, compressing timelines "from years to decades" down to "weeks and months" (Squawk on the Street, Jun 23). Hittery noted "more than 30 percent of all new drugs that are going in front of the FDA are focused on this target," and that the tool predicts not just binding but mechanism, agonist vs. antagonist.

AlphaFold's architect joins Anthropic. On Machine Learning Street Talk (Jun 22), John Jumper confirmed his move and pushed back on the idea that AlphaFold was a generic foundation model: "I don't really love the bitter lesson as people try and apply it. In fact, AlphaFold 2 is the opposite of that." He framed AlphaFold 3's expansion to small molecules as the bridge to drug design, "now we can say, this is exactly where that drug sticks", the foundation Isomorphic Labs is building drug programs on.

A rare validated AI-target benchmark. Verge Genomics CEO Alice Zhang disclosed that in its ALS partnership, 83% of AI-derived targets validated in Lilly wet-lab experiments, against the 20% bar Lilly itself called success, with two targets optioned into Lilly's internal pipeline (This Week in Startups, Jun 22). She put hard numbers on the techbio deal model: "$25 to 42 million up front... milestones that total up to anywhere between 700 to $800 million each," across signed Lilly and AstraZeneca Alexion deals.

Lilly's AI strategy, from the inside. SVP Mahi Ryassam said "the record amount of innovations that I have seen in the last 12 months, I have not seen in the previous 10 years," and described a three-filter framework: patient-impact ROI first, honesty about where "AI is not the best tool," and "re-imagining workflows" rather than just automating them (Transform NOW, Jun 23). Notably, he rejected the "trust AI" framing: "we're actually not telling scientists to trust AI... AI is a tool just like any other tool."

The Debate

Physics models vs. language models. This was the live tension of the week. Hittery was blunt: "we really have to go beyond language models if we want to tackle the biggest problems in medicine... these are physics-based models" (Squawk on the Street, Jun 23). Jumper's complementary-but-distinct view is that the winning systems may be narrow specialist predictors, not general models, AlphaFold "predict[s] nature-level science with the press of a button in a very narrow category of nature-level science" (MLST, Jun 22). The investment read-through: skepticism that simply scaling LLMs into biology is the path.

Does scaling even work in biology? Zhang directly challenged the text-AI scaling narrative: "The biggest gains we've seen have actually come from scaling data and modalities... it's not about making the AI bigger, it's about feeding it the right pair of data" (TWiS, Jun 22). Her LIDAR-vs-camera analogy, proprietary brain tissue as "the molecular ground truth of disease", reframes the moat as data modality, not compute.

Access equity in obesity. On The Readout Loud (Jun 25), STAT's Lizzie Lawrence detailed how Lilly and the FDA quietly granted a single "well-connected" 79-year-old patient access to Reditrutide via compassionate use, drawing "the interest of top health officials," which is "not typical." Obesity expert Dr. Angela Fitch's pushback: "there's 40 million people this could apply to," and her own outreach to Lilly went unanswered.

Stocks in Play

Eli Lilly (LLY) was the week's deal machine. It struck a $1.9B research collaboration with Abbisko Therapeutics (Jun 24), closed the $7.8B Centessa acquisition at $38/share plus a CVR worth up to $9 (Jun 24), and lined up stalking-horse bids for bankrupt Sangamo Therapeutics' genomic platforms and prion program (Jun 23), a tell that Lilly is buying gene-editing capability cheaply. It also signed a BioArctic BrainTransporter neuro pact ($30M upfront, up to $770M milestones, Jun 22). Sell-side moved with it: Leerink lifted its target to $1,232 (Outperform) and Berenberg to $1,135 (Hold). On the access front, Lilly detailed its Medicare GLP-1 Bridge program ($50/month for eligible Part D patients, effective July 1).

Schrödinger (SDGR) and Recursion (RXRX) had no company-specific podcast or news flow this week. Both bounced (SDGR +5.3%, RXRX +3.4%) on the broad rotation into healthcare small-caps after the tech-led selloff. RXRX, at $3.34, sits just above its 52-week low ($2.77) with a market cap below the broader AI-techbio enthusiasm on display at BIO, a widening gap between platform narrative and public-market valuation worth watching.

Read-throughs

• NVIDIA (NVDA) keeps extending BioNeMo's reach into the biopharma stack, the Sandbox AQ deal is one more proof point that GPU-native, physics-based discovery is becoming infrastructure, not experiment.
• Alphabet (GOOGL)/Isomorphic Labs gains conceptually from AlphaFold 3's small-molecule capability moving toward real drug design, even as Jumper's exit to Anthropic raises a quiet talent-retention question.
• Anthropic is now a player in structured scientific AI, its hire of Jumper signals ambitions beyond general chat models into engineered biology prediction.
• Lilly as ecosystem buyer: the Abbisko, Centessa, Sangamo, BioArctic, and AbSci ($100M round) moves show LLY consolidating both AI-discovery capability and GLP-1 adjacencies, a steady demand sink for techbio platforms that can feed validated targets in.

What changed vs last week

The baseline this week: physics-based screening (Sandbox AQ with NVIDIA) and narrow-specialist AlphaFold descendants are the dominant technical narratives; Lilly is the most active strategic buyer; and the pure-play public techbio names (RXRX, SDGR) remain quiet and de-rated relative to the private-market and conference enthusiasm.