Newsletter · 2026-06-18 · Ashutosh Agarwal

AI Drug Discovery Weekly - Week of June 18, 2026: The Productivity Numbers Get Real, and a Synthesizability Reality Check

AI drug discovery newsletter for the week of June 11–18, 2026. Insilico, Latent Labs, and Takeda put hard productivity numbers behind their AI-design claims, while a dedicated episode on the synthesizability gap supplies the counterweight.

AI Drug Discovery Weekly

Week of June 18, 2026: The Productivity Numbers Get Real, and a Synthesizability Reality Check

TL;DR

This was a week where AI-drug-discovery executives stopped talking platforms and started putting hard productivity numbers on the table. Insilico Medicine's CEO laid out 30 development candidates in five years and 13 already in the clinic; Latent Labs claims 47% of its AI-designed antibodies clear drug-like hurdles with zero optimization; and Takeda's CSO described an "AI rescue" of a dead program now heading to the clinic. The counterweight came from a technical discussion of why so many AI-proposed molecules still can't actually be made, the synthesizability gap that separates a slide from a clinical asset. None of the chatter touched our three public coverage names directly (RXRX, SDGR, LLY), but the read-through is squarely positive for the validation-of-platform thesis and a reminder that the value is accruing fastest to private labs and large-cap partners, not the listed pure-plays.

What's new

Insilico put up the most concrete productivity scorecard of the week. On Realities Remixed (Jun 18), CEO Alex Zhavoronkov said Insilico has "nominated 30 developmental candidates since 2021... and now we've got 13 clinical and three phase twos and one phase two complete." His framing of the productivity delta was the quotable part: a medicinal chemist typically nominates "five, six, maybe seven in their entire... career," so 30 in five years is, in his words, "massive productivity." He traced the lineage to Insilico's 2019 generative-reinforcement-learning work that produced a first-in-kind molecule in 46 days.

Latent Labs quantified the antibody-design leap. On Free Radicals (Jun 16), founder Simon Kohl, a former AlphaFold II researcher at DeepMind, walked through three frontier models shipped in nine months: Latent X1 (de novo binders), Latent X2 (antibodies, launched in December), and Latent Y (an autonomous design agent). The headline metric: "47% of the designs we've tested, antibody designs, already had passed critical drug-like hurdles out of the gate, so without optimization." He framed this as compressing 18-month design timelines to roughly one month, with Latent Y collapsing "weeks of work by expert protein designers to an afternoon."

Takeda gave the clearest "it's working in a real pipeline" example. On The BioCentury Show (Jun 18), CSO Chris Arendt described a partnership with Nabla (out of George Church's lab) that used de novo epitope design to rescue an undevelopable, non-manufacturable myasthenia gravis large molecule the company "was going to close," now yielding multiple candidates "entering the clinic... very, very soon" in what he called "a rescue by AI." He also said de novo antibody design is now hitting "double digit nanomolar binders" against multipass membrane proteins with no structural data, and that preliminary safety assessments that once took a subject-matter expert a full week each can now be reviewed at 20 per week.

The platform vision got its big-name endorsement. On View From The Top (Jun 18), Demis Hassabis reiterated that AlphaFold is now used by roughly 3 million researchers and positioned Isomorphic Labs (the Alphabet spin-out) as stacking "several more AlphaFold-level breakthroughs" to take drug discovery "down from years to months, maybe even one day weeks."

A grounded small-molecule data point: on The Long Run with Luke Timmerman (Jun 16), Octent's Sri Kosuri described ML-guided iterative chemistry that built ~250,000 analogs en route to OCT-980, an oral corrector for rhodopsin misfolding now in Phase I-B with data expected Q3 2027, a reminder that AI still rides on top of large, real, wet-lab chemistry cycles.

The debate

The bull case this week was unusually data-rich: multiple independent operators (Insilico, Latent Labs, Takeda) putting real candidate counts, hit rates, and clinic-entry timelines behind the AI-design claim, not just TAM rhetoric.

The bear case got its own dedicated episode. On Data in Biotech (Jun 17), the discussion centered on the synthesizability gap: generative models, often built on language-model-style 2D/string representations, routinely propose molecules that "cannot reasonably be made" in the lab, so "all of the work that's gone into that generative AI and the extremely large amounts of computation" can lead to candidates that can never be validated. The deeper critique: abstract 2D representations don't capture the 3D reality that actually drives binding, which is why AI predictions "very frequently don't pan out when you test them in the real world." The proposed fix, constraining the search to building-block chemistry that is known to be synthesizable, directly trades away the novelty that makes AI attractive in the first place.

Net: the week strengthens the "AI compresses timelines and rescues hard targets" thesis, but the same week supplies the strongest reason to discount headline hit-rate stats. A 47% "drug-like out of the gate" number means little if a meaningful slice of those designs can't be made, dosed, and survive the clinic.

Stocks in play

Recursion (RXRX) and Schrödinger (SDGR), the listed AI-platform pure-plays, drew no podcast or news coverage this week, and both sit near the low end of their 52-week ranges (RXRX ~$3.23 vs a $7.18 high; SDGR ~$15.76 vs a $23.75 high). The disconnect is the story: private labs are generating the validating headlines while the public pure-plays are out of the conversation and de-rated.

Eli Lilly (LLY) had the only hard corporate news in the universe, but it was clinical, not AI: on Jun 15 the company reported a Phase 3 win for a pirtobrutinib-based three-drug regimen in relapsed/refractory CLL/SLL, cutting risk of progression or death by 45% across 639 patients (MT Newswires, 06/15/2026). Shares were down ~1.25% on the day, trading near the high end of a wide 52-week range.

Read-throughs

• Validation accrues to platforms and large-cap partners, not pure-plays, for now. Takeda's Nabla rescue and Insilico's clinic count show pharma and private labs capturing the early AI-design wins. For SDGR (a software-plus-pipeline model) this is thesis-supportive on the science but a competitive warning on monetization: the de novo antibody and protein-design frontier (Latent Labs, Nabla) is moving fast and largely outside the listed names.
• The synthesizability critique is a direct check on hit-rate marketing. When SDGR or RXRX next tout AI throughput or design-success metrics, the Data in Biotech framing is the right diligence lens: ask how many designs are actually synthesizable and how predictions hold up in the wet lab, not just in silico.
• Big pharma is internalizing AI R&D. Takeda describing 200 planned AI apps and halved discovery timelines signals that the largest buyers (the natural customers for SDGR's software and RXRX's partnerships) are building in-house, a medium-term demand risk for selling platforms to pharma.
• Isomorphic/AlphaFold momentum raises the bar. Hassabis's "years to weeks" framing sets the narrative ceiling the listed names are measured against; it lifts sentiment for the theme but sharpens the question of why the public pure-plays are de-rated while the science accelerates.

What changed vs last week

This is the inaugural issue of AI Drug Discovery Weekly, so there is no prior baseline to compare against. Starting next week this section will track week-over-week shifts in the debate, candidate/pipeline milestones, and price action across the coverage universe. Baseline set this week: RXRX $3.23, SDGR $15.76, LLY $1,098.13; debate centered on AI-design productivity claims vs. the synthesizability gap.