Brian Reed: Eliza — before we get anywhere near a framework or a big picture, I want to drop you inside a specific room. Insilico Medicine, a drug called INS018_055, and a generative AI system proposing molecular structures that no human had ever synthesized. Just react to that image for a second.
Eliza Ward: I mean — a machine designing a molecule from scratch, not from a library, just... inventing it. That's the part that stops me every time.
Brian Reed: Right, and here's what makes it concrete rather than abstract. That process — Insilico's generative AI working alongside traditional medicinal chemistry — took that drug from target discovery to Phase II clinical trials in eighteen months.
Eliza Ward: Eighteen months. Okay. The conventional version of that journey is closer to ten years.
Brian Reed: So the number that won't leave me alone: what does it actually mean that a machine can learn the grammar of molecular chemistry — binding, stability, safety properties — from huge chemical datasets, and then use that grammar to invent? Not select. Invent.
Eliza Ward: That's generative AI for de novo molecule design — that's the actual term, de novo, from nothing. The system learns patterns from what already exists and then explores chemical space beyond what any library contains. It's the chef analogy, genuinely: read every recipe ever written, now create a dish that has never existed.
Brian Reed: And the people watching that — the medicinal chemists in the room — they're seeing structures appear that they didn't draw and wouldn't have drawn.
Eliza Ward: Which is — wait, that's the part I want to hold for a second. Because it's not replacing the chemist. Insilico still ran this through conventional medicinal chemistry validation. The AI generated the candidates; humans interrogated them.
Brian Reed: And the result was rentosertib — INS018_055 — a TNIK inhibitor, which made it to Phase IIa.
Eliza Ward: With positive results. That's confirmed.
Brian Reed: Positive Phase IIa results. So the eighteen months isn't just a speed record — it's eighteen months and then something worked. That gap between those two facts is the whole story we're in today.
Eliza Ward: Yeah — and the question underneath it is whether that's a proof that AI changed what's *possible*, or just a proof that AI changed how fast we arrive at the same odds.
Brian Reed: But that same odds question — it has a test case, right? Because Exscientia did this too. DSP-1181, AI-designed small molecule, into Phase I. And then it stopped.
Eliza Ward: Discontinued. Yeah. That's the word.
Brian Reed: So why? Efficacy? Safety signal? Someone ran the commercial math and walked away?
Eliza Ward: The research doesn't say. That's — actually, that gap is the thing. We know it entered Phase I, we know it stopped. The reason isn't in the record. And that silence is doing a lot of work.
Brian Reed: Because if it's commercial math, that's almost a different problem than if it failed on the biology.
Eliza Ward: Completely different. But here's what DSP-1181 and INS018_055 have in common structurally: both got celebrated at the speed milestone. Into Phase I, into Phase II — that's where the press release lives. What happens after is quieter.
Brian Reed: The part I keep getting stuck on — isn't one success still meaningful? Like, eighteen months to Phase II is not nothing.
Eliza Ward: It's one data point. That's — I mean, no source gives us a median timeline across a population of AI-assisted programs. We don't have that. So we're extrapolating a pattern from a single case, and then Exscientia's discontinuation is sitting right next to it.
Brian Reed: So the scoreboard has one confirmed win and one confirmed stop, and we're calling it a trend.
Eliza Ward: And the scoreboard was designed by the people playing the game. That's the part that actually bothers me. AI was optimized for binding affinity, safety markers, oral bioavailability — the things you can measure early. Clinical survival in humans is not that. Patient variability, side effects that only show up at scale — AI didn't train on that because that data barely exists in clean form.
Brian Reed: Which means — wait, so you're saying AI might have just accelerated the pipeline without changing what makes drugs fail?
Eliza Ward: Faster at failing is still useful. I want to be clear about that. But it's not what the INS018_055 headlines suggested. The claim in circulation is that AI improved success rates. And we don't — there's no systematic data that actually shows clinical survival improving across AI-assisted programs versus conventional ones.
Brian Reed: So the Exscientia discontinuation isn't a footnote to the Insilico story. It's the question the Insilico story can't answer by itself.
Eliza Ward: And that's exactly where I want to pull back — because the Exscientia gap points at something bigger than one discontinuation. The whole pipeline. Where does AI actually move the needle, and where is it just... arriving faster at the same wall?
Brian Reed: Walk me through the stages. Because I think they're not equal.
Eliza Ward: They're really not. So — okay, start at the beginning. Target identification. BenevolentAI runs knowledge graphs across genomic and proteomic data, mines the literature at a scale no single researcher can touch. That's not incremental. A human team reading papers to find a disease-associated target is — I mean, that's months. BenevolentAI's approach compresses that by doing it across the whole literature simultaneously.
Brian Reed: And then protein structure — AlphaFold changes what you can actually see.
Eliza Ward: Fundamentally. Google DeepMind's AlphaFold predicts a protein's 3D shape from its amino acid sequence. Before that, you're working from a flat diagram, essentially guessing which pocket a drug molecule might bind to. Now a researcher can look at the actual structure and say — there, that's the site, design toward that.
Brian Reed: That's not a speed improvement. That's a different kind of information entirely.
Eliza Ward: Right. Then virtual screening — AI filters millions of compounds in silico before a single one gets synthesized. No lab time, no reagent cost at that stage. The hit discovery step used to eat enormous resources just eliminating obvious dead ends. Now you eliminate them computationally.
Brian Reed: The part I keep wondering about, though — that's all pre-human. Target ID, structure prediction, screening. Lead optimization too. All of that is before a drug ever sees a patient. So where does the complement framing start to get... convenient?
Eliza Ward: That's — yeah, that's the fracture line. Nature Reviews Drug Discovery frames AI as improving development mostly as a complement to conventional chemistry and biology. Not a replacement. Fine. But those same journals also call it 'revolutionizing' drug discovery and describe it as a 'strategic component at the core of innovative drug discovery.' Those two framings can't both be precisely true. Either you're at the core or you're a complement. That's a meaningful difference.
Brian Reed: So 'complement' is load-bearing. It's doing — wait, is it doing two things at once? Claiming credit for the wins and deflecting when it doesn't translate?
Eliza Ward: That's what I can't rule out. INS018_055 still needed conventional medicinal chemistry validation. Generative AI did not eliminate the human chemists. So when it works, AI led the design. When DSP-1181 stops in Phase I, the framing quietly absorbs it — human biology is complicated, conventional methods missed something. The word 'complement' conveniently survives both outcomes.
Brian Reed: There's baricitinib too — AI-assisted repurposing, an existing approved drug identified for a new target. That's a genuine win. But it's also — I mean, that's a different problem than de novo design. Repurposing is finding a known thing in a new place. It doesn't actually test whether AI can predict clinical success from scratch.
Eliza Ward: Exactly. It extends the range of what AI is doing without resolving the attrition question. And the attrition question is what determines whether this is transformation or acceleration toward the same failure rate.
Brian Reed: And there's a deal structure coming that I think crystallizes exactly what the industry actually believes versus what it's willing to say publicly — the XtalPi numbers with DoveTree are where that gets real.
Eliza Ward: XtalPi is exactly where the belief system shows its hand. So — two deals. Eli Lilly, 2023, $250 million. Real cash, committed collaboration. That's the one nobody disputes.
Brian Reed: And then DoveTree, 2024, up to $6–10 billion.
Eliza Ward: Up to. That phrase is doing enormous work. It's milestone-contingent — meaning XtalPi doesn't see $6 billion, or $8 billion, or $10 billion unless their AI-designed candidates hit specific clinical targets. It's not capital on the table. It's a payment schedule tied to proof.
Brian Reed: So the headline number is — it's a ceiling, not a floor.
Eliza Ward: Think about it this way. A contractor tells you they'll pay up to $10,000 for your kitchen renovation — if every inspection passes, if the tile is perfectly laid, if it's done by March. That's not $10,000. That's hope structured as a number.
Brian Reed: Right — but the part that doesn't fit is, who structures a $250 million deal two years before a $10 billion one unless they genuinely believe something changed?
Eliza Ward: Belief is not the same as certainty, and that deal structure is actually the proof. If DoveTree were certain, the $6–10 billion would be committed. The milestone contingency is the industry admitting — quietly, in contract language — that the infrastructure underneath is not solid yet.
Brian Reed: What infrastructure, specifically?
Eliza Ward: Three things that nobody has actually solved. Data quality — I mean, the same literature that praises AI for analyzing vast datasets also cites data quality as an unresolved bottleneck. Pharmaceutical data is siloed, proprietary, often biased toward compounds that failed in predictable ways. A more sophisticated model trained on bad data is still just — it's a faster way to be wrong. Then interpretability. Regulators at the FDA cannot approve a compound because an AI recommended it and the researchers can't fully explain why. That's not a philosophical problem, that's a pathway-to-market problem.
Brian Reed: The black box is literally a regulatory blocker.
Eliza Ward: And then ADMET — pharmacokinetics, toxicity prediction. Models exist. XtalPi uses them, Insilico uses them. But they haven't eliminated late-stage safety failures. A compound can look perfect on absorption and metabolic stability in silico and then — something shows up at scale in a human population that the training data never contained. The prediction-to-reality gap is still there. That's why CLADD exists, actually — it's a RAG-based multi-agent LLM system, retrieval-augmented, multiple agents pulling from biomedical knowledge bases dynamically, specifically because no single model has clean enough data to answer complex drug discovery queries on its own. You build the workaround because the underlying problem isn't fixed.
Brian Reed: So the binding constraint isn't — hang on, it's not whether the models are smart enough. It's what the models have to grow in.
Eliza Ward: Exactly that. Model sophistication is not the ceiling anymore. Data soil and regulatory readiness are. And that's what the DoveTree milestone structure is quietly encoding — we believe in the models, we are not yet certain the soil is ready.
Brian Reed: Which means the XtalPi number isn't evidence of conviction. It's evidence of a very expensive bet on conditions that don't fully exist yet.
Eliza Ward: And that's — I mean, that's where I want to leave the infrastructure argument, because the bet is real and the conditions aren't yet. But there's something underneath all of it that I keep not quite saying out loud.
Brian Reed: The patient at the end of it.
Eliza Ward: Yeah. Because we've been talking about timelines, deal structures, attrition rates — and at some point there's a person in a Phase III trial who has no idea whether INS018_055 came out of a generative model or out of a medicinal chemist's notebook. They don't know Insilico Medicine. They don't know what de novo design means. They're just — they're waiting to find out if it works.
Brian Reed: And that's the test AI hasn't actually faced yet. Not at scale. The eighteen months from target to Phase II is real — I'm not dismissing it — but speed only matters if it gets you somewhere worth going. A Phase III result. A drug that works better, or is safer, or reaches people who couldn't access the old one. That's the silence at the end of the pipeline. The long wait before the data comes back.
Eliza Ward: We're still in that silence. That's the honest place we're at. Thanks for pushing through all of it with me.
