No PriorsNo Priors Ep. 121 | With Chai Discovery Co-Founders Jack Dent and Joshua Meier
CHAPTERS
- 0:05 – 0:47
CHAI 2.0 launch: zero-shot antibody discovery and why it matters
Sarah introduces CHAI Discovery co-founders Joshua Meier and Jack Dent and frames CHAI 2.0 as a step-change in computational antibody discovery. The conversation sets up the promise: moving biology toward an engineering discipline, with antibodies as the initial wedge.
- •CHAI 2.0 positioned as a zero-shot antibody design platform
- •Claimed ~100x improvement in hit rate vs prior computational approaches
- •Context: founders’ backgrounds spanning AI, engineering, and bio
- •Thesis: best antibody engineers may become “expert prompt engineers”
- 0:47 – 3:56
Why start CHAI now: the bet on multi-molecule structure and portable platforms
Josh explains why this moment became company-timely: prior work felt like research without a near-term commercialization path. CHAI’s founding bet centered on better structure prediction for interactions (not just single proteins) plus a more generalizable, “portable” AI platform.
- •Timing rationale: too early vs too late—the “1–2 year sweet spot”
- •Key bet: structure prediction evolving from single proteins to interactions/complexes
- •Diffusion/LM-style advances enabling diverse generations (not single conformation outputs)
- •Critique of earlier AI-bio companies: overly tight lab integration and limited portability
- •Goal: deployable platform that scales across hundreds/thousands of projects
- 3:56 – 6:12
Jack’s motivation: the platform shift of “engineering molecules with atomic precision”
Jack describes being pulled from general AI/product problems into biotech by the magnitude of impact. Once the possibility of precise molecular engineering feels real, it becomes hard to justify working on anything else.
- •Personal background: product/engineering leadership, then choosing biotech
- •Friendship and long-running technical discussions with Josh as catalyst
- •Conviction: glimpsing the future makes the opportunity “impossible to unsee”
- •Societal scope extends beyond health into broader molecular applications
- 6:12 – 10:13
What CHAI 2.0 achieved: antibody hits in a 24-well plate, not billion-scale screens
Jack gives a lay explanation of the result: CHAI 2 designs antibodies that bind specified targets at a dramatically higher success rate, validated with a fast lab cycle. The chapter contrasts this with traditional discovery workflows that rely on massive random screening or animal immunization.
- •Antibodies are a major drug modality (large share of approvals/top-selling drugs)
- •Experiment loop: generate designs → synthesize/test → ~2-week validation cycle
- •Reported outcome: ~20% of tested designs bind as intended (vs ~0.1% computational baselines)
- •Traditional alternatives: yeast/phage display library screening, mice/llama immunization workflows
- •Claimed impact: faster/cheaper discovery plus unlocking previously inaccessible targets
- 10:13 – 12:54
Benchmarking like engineers: why 50+ targets and how targets were chosen
Josh explains why CHAI evaluated at scale instead of cherry-picking a handful of targets. The team prioritized generalizability and experimental turnaround speed—down to scraping vendor catalogs—while ensuring targets were held out from training.
- •Rationale for 50+ targets: robust benchmarking vs anecdotal demos
- •Selection method: choose in-stock proteins from vendor catalogs for speed
- •Holdout strategy: compare against SABDAB/PDB-related sources to avoid leakage
- •Sequence identity filtering (e.g., removing high-identity overlaps) to ensure novelty
- •Caveat: target set optimized for assessment, not necessarily therapeutic relevance
- 12:54 – 15:52
How CHAI 2 works under the hood: from structure prediction to generative design
Josh and Jack describe the conceptual stack: structure prediction provides an “atomic microscope,” then design uses generative modeling to place atoms and sequences that satisfy binding constraints. They map the shift from prediction to design to the shift from classification to generation in mainstream AI.
- •CHAI 1.0: structure prediction engine, open-sourced and widely used
- •Structure prediction as enabling layer: visualize atomic positions in 3D
- •Design framing: prompt with a target and constraints; generate sequence + structure that binds
- •Analogy: structure prediction as “ImageNet moment,” design as “Midjourney for molecules”
- •Lock-and-key intuition: can’t design a key if you can’t reliably see the lock
- 15:52 – 18:15
Emergent concepts: what the model seems to “know” about molecular interaction
Sarah probes for an equivalent of emergent language in LLMs; Josh points to the surprising generalization across distant protein families. The team observes similar hit rates even on much more dissimilar sequences, implying the model captures deeper interaction regularities than classic biological similarity heuristics suggest.
- •Observation: models generalize beyond near neighbors in training distribution
- •Biologist notion of “dissimilar” targets may not reflect model’s learned representation
- •Harder split test: pushing down to very low sequence similarity with similar success
- •Implication: interaction “signatures” may be embedded in data in a learnable way
- •Open question: how to extract new scientific principles from these learned rules
- 18:15 – 20:33
Expected impact and deployment: faster discovery + partnering to scale reach
Josh distinguishes two impacts: speeding up what’s already possible and enabling programs that were previously unreachable. Jack adds that CHAI is opening access to academics and industry because the opportunity is too large for one company to pursue alone.
- •Impact #1: compress months/years of discovery by shifting screening to compute
- •Impact #2: tackle targets where traditional methods fail or stall
- •Failure modes differ between AI-driven search and traditional lab workflows
- •Strategy: broad access via partnerships to maximize real-world adoption
- •Operational reality: heavy inbound demand and prioritizing early access
- 20:33 – 22:17
Reception and skepticism: what people question and what the benchmark answered
Josh outlines typical critiques—whether results are real, general, and high-quality—and argues scale benchmarking helped address them. The remaining objection is pragmatic: does faster discovery truly change what’s feasible, or just accelerate existing workflows?
- •Common skepticism: reproducibility, data trust, generalization beyond well-studied targets
- •Scale benchmark as a credibility mechanism vs single-target demos
- •Pragmatic pushback: “we can already discover drugs—does AI change targetability?”
- •Counterpoint: biggest unlock is solving the programs that previously failed
- •Market dynamic: growing belief that teams must adopt or risk falling behind
- 22:17 – 24:22
Wet-lab screening in an AI era: sampling at scale and hybrid workflows
The discussion turns to whether massive wet-lab screening becomes obsolete. Josh and Jack argue labs still matter—especially to test large batches of AI-generated candidates—potentially enabling enormous sampling that humans couldn’t manually evaluate.
- •Models are probabilistic: more samples can yield better candidates
- •Paper used up to 20 designs/target; future could test 10x–1000x more
- •Lab as evaluator: can test thousands of candidates even if humans can’t rank them directly
- •Hybrid future: CROs and screening expertise remain valuable, now paired with AI generation
- •Goal: better drugs (e.g., dosing/formulation improvements) via broader search
- 24:22 – 26:53
Biotech 25 years out: from doom-and-gloom to “CAD for biology”
Jack predicts a shift from slow discovery toward software-like design suites for molecules, analogous to tools like SolidWorks or Photoshop. He frames current biotech pessimism as cyclical and argues that rapidly improving success rates could fundamentally reshape feasibility and economics.
- •Current context: weak biotech markets and long investment cycles
- •Trajectory argument: 0.1% → ~20% in a year suggests room for rapid improvement
- •Vision: computer-aided design suite for biology and molecular engineering
- •Implications: new targets, new modalities, and new markets become reachable
- •Cultural takeaway: “Bullish on biotech” as a near-term inflection belief
- 26:53 – 31:37
Beyond binders: optimizing developability and next-gen antibody formats
Josh explains that binding is only the first gate; real drugs must satisfy manufacturability, stability, and other constraints. He also describes how easier antibody generation could accelerate more complex therapeutic designs, like biparatopic antibodies and multi-target optimization.
- •Drug design is multi-objective: binding plus stability, manufacturability, developability
- •CHAI 2 reduced search over binding space; next is deeper property optimization
- •Next-gen formats: biparatopic/multi-epitope concepts become more practical
- •Multi-target prompting: design for cross-species binding or selectivity constraints
- •Industry effect: lower discovery risk may reduce target crowding and expand opportunity
- 31:37 – 36:20
Where CHAI invests next: from ‘hits’ to true drug candidates and a usable product layer
Sarah asks about defensibility and strategy; Jack and Josh emphasize converting hits into clinic-ready molecules and building the product/software interface around the models. The team frames CHAI 2 as more than a model: it’s an end-to-end pipeline with increasingly complex “prompting” and workflows.
- •Near-term: turn hits into viable therapeutics via deeper characterization and assays
- •Long-term: generate full drug candidates zero-shot, not just initial binders
- •Defensibility via productization: workflows, interfaces, and multi-constraint specification
- •Design requires better UX than folding: prompts are complex and scientist-driven
- •Closed-loop learning: incorporate lab outcomes to guide next design iterations
- 36:20 – 39:57
Advice to scientists: learn the tool, master prompting, and rethink what’s possible
Josh argues the key skill shift is learning how to express constraints and intent to the system—prompting and problem formulation become central. Both founders describe how hands-on access changes creativity: users start to see new strategies around epitope choice, selectivity, and multi-target design.
- •Practical advice: get access, iterate on prompts, and explore capabilities directly
- •Prompting as domain leverage: epitope choice can drive selectivity/cross-reactivity
- •Creativity unlock: realistic success rates change how people plan programs
- •Analogy to LLMs: hard to imagine use cases until you interact with the system
- •Raising the bar: domain experts become more effective with faster feedback loops
- 39:57 – 43:14
Building an interdisciplinary company: how Jack ramped on bio and why focus mattered
Jack describes ramping into biology as an intense push, accelerated by surrounding himself with exceptional interdisciplinary teammates. Josh highlights a culture of focus: a small team moving in one direction with strong engineering discipline, rather than fragmented research efforts.
- •Ramping on a new domain: deliberate study to reach the frontier
- •Team leverage: complementary expertise across bio/chem/physics/AI/engineering
- •Small-but-mighty: ~dozen-person team executing like an engineering organization
- •Culture: shared focus on a single hard problem vs “pet projects”
- •Execution emphasis: shipping breakthroughs into partner-facing workflows
- 43:14 – 46:31
Platform investment and engineering rigor: avoiding entropy in deep learning codebases
Sarah probes CHAI’s early platform mindset; Jack explains why architecture stewardship, modularity, and tests are critical—especially when bugs can waste weeks and large compute budgets. This chapter connects classic software practices to high-stakes model training and iteration speed.
- •Scaling lesson: without stewardship, software entropy slows progress to a crawl
- •Principles: simplicity, modularity, minimizing context needed for contributors
- •Deep learning risk: regressions may surface weeks later after expensive runs
- •War story: bisecting git history via repeated training to find a hidden bug
- •Rigor: unit tests and disciplined engineering as a competitive advantage
- 46:31 – 49:27
Scaling experiments, staying scrappy, and hiring to meet demand
Josh explains a pragmatic approach to compute scaling: invest aggressively once there are clear signs of life, but avoid scaling before conviction. The episode closes with CHAI’s hiring push across research, product engineering, antibody engineering, and go-to-market roles.
- •Early scrappiness: cloud credits and uncertain early company logistics
- •Compute strategy: scale fast when signals justify it; stay focused when uncertain
- •Work style: high intensity plus disciplined prioritization
- •Hiring across functions: research, product engineering, antibody engineering, BD/sales
- •Goal: operationalize CHAI 2 for partners and broader industry adoption
