Chai Discovery: From Protein Prediction to Molecular Engineering

Paint a binding site, generate a molecule

Load a molecule, paint the region where an antibody should attach, and generate candidates. Chai’s molecular design suite turns that request into a visual workflow: a paintbrush specifies an epitope, the binding site on a target, and an operation resembling content-aware fill generates binders—molecules intended to attach there. Neil Patil compares the interface to computer-aided design (CAD) tools and Figma. The scientific problem becomes a spatial specification: design something that engages this part of another molecule.

In the Latent Space conversation, Patil describes leading platform and product work across training infrastructure, model serving, and the design suite. His path ran from App Store development through robotics and self-driving cars, early work at Vanta, and founding a security company before joining Chai. The co-founder introduced as Matt moved from theoretical computer science into AI and biology during his PhD, working through the AlphaFold 1 and AlphaFold 2 eras. Their shared bet is a general software layer for drug discovery: Patil positions Chai as a supplier to pharmaceutical companies, and Matt names Eli Lilly, Pfizer, Novartis, and Argenx as partners.

Matt argues that earlier attempts to sell this software layer were constrained by inadequate models. Two advances helped change that: AlphaFold 2 Multimer modeled multiple proteins together, while ProteinMPNN demonstrated experimentally validated inverse folding—finding a sequence that folds into a supplied shape. Chai sought models that could work across targets. Its early Chai 2 challenge produced antibody hits against half of 50 selected targets. Half the targets had hits; that is not a 50% success rate per design. The set came from laboratory contractors’ already validated catalogs after difficulties with earlier target preparations.

Binding is only the first constraint

An antibody has a relatively constant framework and variable binding regions. In the simplified Y-shaped picture, design concentrates on the fingertips. Predicting how an arbitrary antibody binds a target is difficult, but design can choose easier binding arrangements: the model need only find a suitable solution among the possibilities. That freedom helps explain why designing an antibody can sometimes be easier than predicting an arbitrary antibody–target interaction.

Traditional approaches discussed include mouse immunization and large yeast-display screens, which search candidate populations for molecules that stick to a target. An initial hit may reveal neither its binding site nor whether it has suitable drug properties. Intentional design aims to control the site because different therapeutic effects require different interactions: - Block an interaction: occupy a region needed for an unwanted protein interaction. - Activate a receptor: engage a membrane protein precisely enough to trigger downstream signaling, a behavior called agonism. - Bring targets together: bind molecules to induce proximity between them. - Deliver a payload: direct a toxic payload toward a cancer cell with an antibody-based therapeutic. Finding a binder alone does not establish any of these effects.

The specification must also express which other molecules to bind or avoid: - Desired cross-reactivity: a candidate intended for humans may also need to bind the corresponding monkey protein for animal testing. One proposed approach targets a conserved region—one that changes little between the two versions. - Selectivity: the candidate should avoid similar human proteins whose unintended binding could cause toxicity or side effects. Chai describes these as both modeling and interface problems. Its claim is bounded: users can specify particular binding and avoidance targets, and selective design can be attempted in some cases. This does not amount to reliable counter-screening against a large, diverse protein set.

From folding a sequence to designing one

Chai’s model history begins with infrastructure. Around the appearance of AlphaFold 3, its early team needed a production multiple sequence alignment (MSA) pipeline. Aligning related protein sequences exposes conserved positions and correlated mutations, which help infer which amino acids lie near one another in three dimensions. The five-person team used its first structure-prediction model, Chai 1, to force the supporting systems into place. It also built a web interface so biologists could use the model without downloading a Git repository; one team member recalls roughly 48 hours awake finishing the report and server.

Chai 1 turns molecular inputs into atomic coordinates. Its inputs combine sequence information, the atoms associated with each token, and properties such as charge and element type. A molecular tokenizer groups that information for a transformer trunk. The transformer produces representations that condition a diffusion model, which generates the three-dimensional geometry.

Chai 2 generates candidates for a target structure. The team describes an all-atom diffusion model that can choose which atoms are present as well as where to place them; those choices map back to amino-acid identities. Three tasks clarify the change: - Structure prediction: start with a sequence and predict its shape. - Inverse folding: start with a shape and seek a sequence that folds into it. - Co-design: generate a sequence and compatible structure together. Conceptually, co-design repeatedly adjusts shape and sequence toward a self-consistent pair: a structural change alters which sequence could support it, and the sequence constrains the structure. This is the team’s explanation of the mechanism, rather than a full architecture specification.

Self-consistency needs an experimental check

A plausible sequence–structure pair creates an evaluation problem. One computational check sends the generated sequence through a separate folding model and compares its prediction with the intended structure. Agreement raises confidence, but self-consistency can reward repetition: nearly identical proteins may score well while solving few distinct problems. Evaluation therefore needs several signals: - Structural agreement: does the separate model recover the intended geometry? - Prediction confidence: how certain is that model about its output? - Generation diversity: does the generator produce different sequences and structures? High agreement and confidence can coexist with poor diversity. These signals help select candidates; they do not prove laboratory behavior.

Binding assays return in weeks; resolving structures can take months. These reported timescales make functional testing a faster improvement loop, though still much slower than computational evaluation. Structure prediction also has a convenient benchmark: hold out sequences and close relatives, then compare predictions with known structures. A newly designed binder has no such answer key. Confirming that it binds requires making it and testing it.

For one Chai 2 design, experimental structure validation using cryo-EM produced such close overlap that the team initially suspected it had received its own design back. The reported error was 0.33 angstrom for that comparison, with no specified error definition, atom set, or distribution across designs. Separately, the team recalls approximately 20% average binding hit rate in the 50-target study, without specifying the averaging method or assay criteria. That measure differs from obtaining hits on half the targets. Asked about leakage, the team says it chose validation targets with no known antibody binder; it does not describe a broader training-data audit.

The product must express the biology

After Chai 2, the team debated studying the 25 unsolved targets individually versus investing in general model improvement. It chose the latter and reports progressive gains through Chai 2.5, Chai 2.7, and Chai 3. The goal extends beyond tight binding to developability: stability, ease of manufacturing, self-aggregation, and safety considerations that affect a molecule’s usefulness as a therapeutic. The team describes qualitative progress and a Chai 2.5 developability study; generating near-therapeutic-grade molecules remains the objective.

Demand following Chai 2 prompted a product build-out and efforts to secure serving compute. Pharmaceutical intellectual-property concerns shaped deployment: Chai describes aggressive data segmentation and single tenancy, approaching a separate deployment or account per customer. Eli Lilly worked closely on the first design suite. The interface must turn biological requirements into usable model context—for example, specifying an epitope on a membrane receptor while preventing the designed antibody from colliding with the cell membrane. Visual tools, plots, and scientific analysis help users define those constraints and inspect the results.

A previously difficult target gives partners a meaningful first test. Scientists can evaluate the model without immediately sharing sensitive current work. Chai describes collaborators with years or decades of experience on individual targets; their detailed feedback improves the product. In one anecdote, a scientist became emotional after Chai helped obtain an initial binder for a target she had pursued for ten years.

Internal scientists provide another check. An early lab-science hire, Nathan Rollins, reportedly insisted on rigorous validation when Chai 2 results arrived. The science team continually expands an evaluation set containing targets with existing therapeutics and deliberately challenging cases. Its experiments seek initial binders to test and improve the models, rather than to establish Chai’s own drug pipeline.

Better candidates move the bottleneck outward

An interviewer asks whether later optimization can be as general as initial hit discovery. The response questions how permanent those stage boundaries are: many exist because the initial molecule is not good enough. Chai wants to generate increasingly drug-like molecules directly, while acknowledging obstacles in specifying behavior and learning multiple molecular properties. That supports two complementary goals: - Research seeks stronger first attempts: generate new candidates that need less subsequent work. - Product supports model–lab cycles: use results from one binder to condition the next model run. As simpler blocking tasks improve, iteration could move toward harder functions: activating receptors, making bispecific antibodies whose arms bind different targets, or designing antibody–drug conjugates that carry a payload.

The software could change with that progression. The team compares today’s product to Cursor: scientists inspect molecular details much as programmers inspect code. If those outputs become sufficiently reliable, future interfaces could orchestrate campaigns across epitopes, targets, or pathways. But choosing a therapeutically useful epitope requires broader biological understanding than finding a binder. It requires knowing which proteins drive disease, how they interact, and which interaction an intervention should change. A confident structural prediction might suggest a site to block in a particular viral-threat scenario; it would not establish therapeutic success.

Structural prediction also remains uneven across protein classes. The discussion distinguishes AlphaFold 2’s success on a class of single proteins with useful MSAs from the harder antibody–target problem. Antibodies’ diversity and rapid recombination to recognize unfamiliar pathogens do not supply the same conserved evolutionary information. Identifying useful epitopes on new targets consequently remains difficult even when some cases are tractable.

This shapes the economic case. A host asks why saving a few million dollars on antibody generation matters within a much larger development campaign, using rough illustrative costs. Chai’s answer emphasizes new capabilities and more ambitious targets. It reports receptor agonist activity from Chai 2, without quantitative results or assay details here. Bispecifics illustrate another opportunity: requiring two arms to bind different targets compounds the search burden. The platform strategy aims to support harder designs across a partner’s portfolio, alongside savings on existing discovery work.

Molecular generation depends on reliable systems

Declarative molecular design rests on difficult data engineering. Structural files can contain multiple copies, unresolved regions, alternative positions, and metadata about experimental methods. A suggestion to delegate parsing to language models meets two objections: missed edge cases and an accumulation of special cases that makes code difficult to understand. Chai favors principled parsing that humans can inspect, using concrete examples to check behavior and expose failures.

Compute presents both a supply problem and a workload problem. Chai reports intermittent shortages after relying heavily on spot and on-demand inference, prompting advance capacity purchases. It also argues that hardware and software optimized for large language models leave substantial optimization work for biological models.

For a Chai-1-type architecture, the team describes a pair representation with `L²` entries, where L is sequence length. The described attention computation scales as L³. Despite that heavy arithmetic workload, moving the representations through memory remains a bottleneck; even LayerNorm can consume substantial execution time. Chai therefore optimizes basic operations and evaluates chips on its own workloads, where relative performance can differ between training and inference.

A single design request can spread computation across many GPUs. Storage outages, database interruptions, and GPU failures make orchestration a core requirement. Chai adopted Temporal for durable execution—coordinating long-running work with managed retries—across application database side effects, model calls, and data pipelines. The benefit is avoiding an expanding collection of custom queues and retry logic. With another $400 million reportedly raised and another compute cluster planned at the time of the conversation, the team presents reliable execution as a prerequisite for larger experiments.

Simpler architectures face a data tradeoff

Research failures can be less visible than infrastructure crashes. A loss curve stops improving or gradients behave strangely without identifying a faulty component. Chai’s preference for simplicity follows from that opacity: fewer interacting modules make changes easier to understand and optimize. Another participant qualifies the argument: AlphaFold 2 and AlphaFold 3 gained data efficiency from human-designed inductive biases—architectural choices that help them learn the problem. Removing those choices may require new data sources or more efficient use of existing data.

Sequences are abundant; experimental structures are scarce. Sequence data supports transformer approaches such as ESM, but applying the same approach to structural data is harder. The discussion credits triangle layers with useful architectural bias while emphasizing their cost. Small hidden dimensions and large sequence dimensions fit GPUs poorly, and large pair representations consume memory that could otherwise support more parameters. Chai’s proposed direction is to extract useful principles from AlphaFold and modify them, rather than preserve the entire architecture.

The product faces a related choice: separate workflows for cross-reactivity, selectivity, and bispecifics, or a general interface for marking regions to bind and avoid. Chai favors exploring the general formulation while acknowledging biological complications. A team member connects that approach to an early theoretical-computer-science result: an n^20 algorithm met a polynomial-time requirement while leaving practicality unresolved. The research habit is to allow enough freedom to find a solution, then simplify it.

General models still need customer-specific work

Chai’s competitive thesis is that some design capabilities will become commodities while frontier models pursue harder tasks. The product layer remains consequential: scientists need to inspect particular structural regions, visualize results, and carry out specialized workflows. Meanwhile, public sequences cannot replace biological measurements. Useful experimental labels remain slow to obtain and iterate on, limiting the ambition to generate nearly clinic-ready candidates in one attempt. Chai sees substantial room for both better models and more data, and says it invests in turning compute into additional data without disclosing the mechanism.

Chai says pharmaceutical partners are already running campaigns with its product, while reiterating that it has no plans for its own drug pipeline. It also describes training or fine-tuning model versions for partners. Experimental data may improve performance on particular target or candidate classes, and customization can include a property the customer wants its designs to satisfy.

An interviewer interprets this customization as evidence that scientific AI companies are consulting businesses. Chai maintains its goal of general models and products while acknowledging integration work with every customer. The team sees successful integration as a source of defensibility: earning trust during the first year can lead to longer partnerships and more ambitious projects.

Allocate experiments, attention, and talent

The closing economic discussion treats pharma as a portfolio business: allocate capital across uncertain targets, with the hope that better design tools make more ambitious bets feasible. Participants worry about rising development costs and the funding model; reversing those costs remains a hope, not a reported result. Chai applies the portfolio analogy internally, describing roughly ten researchers within a company of thirty at the time of the conversation. Staff allocate compute and attention to promising ideas.

One participant contrasts scientific work, constrained by lab capacity and compute, with software work constrained mainly by people. Another adds a useful qualification: cheap implementation still consumes attention. An API creates ongoing support, product complexity, and work for customers learning to use it. As AI makes building easier, choosing which features and experiments deserve that attention becomes more consequential.

Asked which bottleneck they would remove, the guests identify two distinct limits: - Validation latency and uncertainty: researchers want immediate feedback on whether a design works; the experimental loop still leaves substantial uncertainty. - Biology’s obscurity: specialist vocabulary, difficult visualization, and perceived credential requirements discourage potential contributors. Chai wants its website and product to make the work easier to understand visually.

The closing ambition is declarative precision engineering for biology: specify the desired behavior and use computation to fill in the molecular design. The team argues that useful models and commercial applications already coexist with substantial unsolved research. Its invitation to generalists is concrete—3D geometry, diffusion models, and transformer trunks all offer entry points for machine-learning researchers. Greater computational control could also make the field more accessible to software engineers, while specialist biological expertise remains necessary.