Antibody discovery: How to rethink the search logic for tens of billions of molecules?

In the biopharmaceutical field, antibody discovery is a systematic engineering problem centered around 'search.' The immune system naturally has the ability to screen for high-affinity binders from a library of astronomical diversity, while the core goal of in vitro antibody discovery technologies is to recreate and accelerate this evolutionary process in the lab. But in vitro reconstruction is never a 1:1 replication. Hybridoma technology, single B-cell cloning, and antibody display technology are the three main technical paths, which can essentially be divided into two categories: 'in vivo immune-driven' and 'in vitro construction and screening.' Each of these routes makes different trade-offs between the breadth of diversity coverage and the depth of screening precision. And this technological iteration that has lasted nearly half a century is now being reshaped by a completely new variable: computational biology.

1. The Gift and Limits of the Immune System

The starting point of antibody discovery is the most sophisticated molecular screening system evolved in nature. B cells generate an initial antibody library in the bone marrow through V(D)J gene recombination, with a theoretical diversity reaching the order of 10¹¹. When an antigen enters the body, it is captured by homologous B cells, which undergo somatic hypermutation and affinity maturation, ultimately producing high-affinity antibodies with dissociation constants as low as nanomolar (nM) or even picomolar (pM).

The logic of in vivo antibody generation has been translated into two core in vitro technical paths. One path is based on in vivo immunity, represented by hybridoma technology: immunize experimental animals with the target antigen, isolate splenic B cells and fuse them with myeloma cells, and then select in HAT medium: aminopterin blocks the de novo nucleotide synthesis pathway, HGPRT-deficient myeloma cells cannot survive via salvage pathways, and unfused B cells undergo natural apoptosis in vitro—only successfully fused hybridoma cells gain immortality and a functional HGPRT gene, allowing them to survive, proliferate, and yield positive clones upon further identification. This path preserves the natural antibody heavy and light chain pairing information but is limited by immune tolerance mechanisms—it's difficult to generate an effective immune response against highly conserved self-proteins (due to central/peripheral tolerance), targets with very high homology to the host animal (recognized as 'self'), and highly toxic antigens (which directly kill B cells or suppress immune responses), with screening throughput also directly limited by fusion efficiency and clone culture scale.

Another in vivo immune path is single B-cell cloning, which directly isolates antigen-specific B cells from the immune animal or patient, fully retaining natural chain pairing, with throughput and screening precision between hybridoma and display technologies. The transgenic mouse platform is an advanced humanized version of the in vivo immune path: this platform uses engineered mice carrying the full human antibody gene. After antigen immunization, it can directly produce fully human antibodies, retaining the natural advantages of in vivo affinity maturation while avoiding human anti-mouse antibody (HAMA) responses triggered by mouse antibodies.

Activation of B cells and production of antibodies

Another category is fully in vitro display technologies, covering branches like phage display, yeast display, and mammalian cell display: antibody gene fragments are inserted into display vectors, large antibody libraries are constructed in vitro, and through multiple rounds of 'bind-wash-elute-amplify' cycles, specific binders are enriched. This path bypasses the limits of immune tolerance, with library sizes reaching up to 10¹⁰–10¹¹, but it loses the natural chain pairing information, and the screened antibodies are artificially randomly paired heavy and light chains, which usually require additional affinity maturation optimization later on.

However, no matter which technical route you take, the core trade-off in antibody discovery has never really changed: the systems that cover the most diversity usually have lower selection precision—you can only prioritize screening high-affinity clones and find it hard to enrich molecules with specific epitopes and good stability; the higher the selection precision, the more you often have to sacrifice some sequence diversity.

2. Affinity and Developability: A system of equations that need to be solved together

Antibody discovery is often simplified as a single-objective optimization problem: find the clone with the highest affinity. But this framework ignores the key fact: antibody drugs must meet requirements for large-scale production, long-term in vivo stability, and clinical safety and efficacy—they’re not just lab reagents that only need to show binding activity.

High affinity is a necessary but not sufficient condition for druggability. A candidate antibody showing picomolar affinity in ELISA experiments might have extremely low yields in expression systems, rapidly aggregate under low pH, precipitate as particles at high concentrations, or cross-react with homologous proteins in serum. These failure modes often remain silent during early discovery, only surfacing when the project reaches the CMC (Chemistry, Manufacturing, and Controls) stage, after a lot of time and resources have already been invested.

An antibody’s developability is inherently determined by its sequence and structural features and is a core component of druggability. The extent of exposed hydrophobic residues in CDR regions directly correlates with aggregation tendency; the charge distribution in the complementarity-determining regions affects conformational stability at low pH; tendencies for deamidation, oxidation, and isomerization of specific residues in CDRs determine chemical stability and functional integrity. These properties can’t be directly quantified by affinity readings, yet they largely define the boundaries of whether a candidate molecule can ultimately become a drug.

Therefore, the decision-making in antibody discovery is essentially a multi-objective optimization problem: finding a globally acceptable balance among efficacy dimensions like affinity and specificity, developability dimensions like expression level, thermal stability, low-pH stability, chemical stability, and safety dimensions like immunogenicity risk. The traditional serial model of 'screen for affinity first, then test developability' results in many molecules being eliminated at the CMC stage, directly prolonging project timelines and wasting R&D resources.

3. The conundrum of exploring sequence space

The Exploration Dilemma of Sequence Space

The limitations of physical screening go beyond just evaluating dimensions later; the deeper problem lies in the huge gap between the astronomical size of sequence space and the actual screening throughput.

A typical antibody CDR-H3 region consists of 8–16 amino acids, with a theoretical sequence space of 20⁸–20¹⁶ (this theoretical space doesn’t consider the natural preferences of V(D)J recombination, the physiological distribution limits of CDR-H3 lengths, or the structural constraints of immunoglobulin folding on sequences— the actual functional sequence space is much smaller, but still enormous compared to experimental screening throughput)—and that’s just one of the six CDR loops in an antibody. Even if a display library has a capacity of 10¹¹, it only covers a tiny fraction of the theoretical possibilities. This means that the 'best' antibodies found by traditional methods are only local maxima in affinity within the explored sequence range, not absolute global optima across the entire sequence space.

A deeper challenge is that the path of affinity maturation is highly sensitive to the starting sequence. The attainable maximum affinity of a given lead molecule is largely predetermined by its sequence and structural features. If the CDR conformational space of the lead molecule is insufficient to accommodate mutations that could lead to higher affinity, later affinity maturation is likely to hit a plateau. And these structural constraints cannot be intuitively recognized or predicted from sequence alone.

It's worth noting that the ultimate goal of antibody development is not to achieve global optimal affinity, but to find a 'satisfactory solution' that balances multiple attributes; yet traditional screening models struggle even to explore the global affinity landscape, further limiting the chances of discovering high-quality candidates.

4. Moving from 'Hitting By Chance' to 'Defined Search'

The underlying logic of traditional antibody discovery is to 'hit by chance' the desired target molecules from massive physical libraries—relying heavily on sufficiently large library sizes and sensitive screening methods. Essentially, it's a 'random generation and passive screening' probability path. The involvement of computational methods is rewriting this fundamental logic: moving from passive screening in physical libraries to active searching and design in computational space.

MatwingsVenus™ (Xiaowu™) AI represents the core application direction of computational biology in antibody discovery—not merely expanding library size or increasing screening throughput, but fundamentally changing the paradigm of 'finding antibodies':

intelligent navigation of sequence space.

The protein function prediction capability of the MatwingsVenus™ (XiaoWu™) agent (VenusX/VenusG) can, before the construction of an antibody library, predict which CDR features—such as length, charge distribution, and hydrophobic patterns—are most likely to interact with high affinity with a target epitope based on the structural characteristics of the antigen. This is not meant to replace the natural immune system or the diversity-generating capacity of display libraries. Instead, it helps to narrow down high-probability regions of functional molecules in the sequence space before diversity generation, improving the effective coverage of the library.

Structure-guided rational affinity maturation

Traditional affinity maturation relies on random mutation libraries and repeated selection, which is time-consuming and limited in efficiency. The protein design capability of MatwingsVenus™ (XiaoWu™) (VenusREM/VenusPrime) provides a new path: by using the 3D structure of the antibody-antigen complex, it predicts which sites in the CDR regions and what types of mutations are most likely to enhance binding free energy while minimizing disruption to the overall molecular structure. In an antibody engineering case for a certain cytokine target, VenusPrime recommended only three mutations located at CDR-H2 and CDR-L1, which improved antibody affinity from the nM level to the pM level, without building and screening a large-scale mutation library. This process shifts affinity maturation from an iterative cycle of “generate diversity → screen → regenerate” to a linear path of “computational prediction → targeted validation.”

Early pre-assessment of developability

Combining protein function prediction and structural prediction tools (AlphaFold2/ESMFold/Protenix) can provide key developability parameters even before locking in candidate molecules. CDR hydrophobic patch exposure, conformational stability under low pH, and post-translational modification hotspots—risks traditionally not revealed until the CMC stage—can be assessed during the discovery phase. Candidates ranked high in affinity but prone to aggregation can be deprioritized, while those with slightly lower affinity but excellent stability can be reconsidered. This early pre-assessment significantly reduces failure risks at the CMC stage.

Epitope-driven targeted antibody design

For most drug targets, antibody biological activity depends not only on affinity but also on the binding epitope. The structural prediction capability of the MatwingsVenus™ (XiaoWu™) agent can annotate functional key regions on the antigen’s 3D structure—such as receptor binding interfaces, ligand-competitive sites, and specific conformational epitopes—and incorporate these constraints into the computational framework for antibody screening and design. This elevates the goal of antibody discovery from “finding molecules that can bind the antigen” to “finding functional molecules that bind in a specific way.”

Conclusion

Antibody Discovery

The history of antibody discovery is essentially a history of technological evolution that keeps expanding the boundaries of the search. Hybridoma technology moved the search space from inside the body to the lab; phage display boosted library capacity from 10⁷ to 10¹¹; transgenic mice upgraded foreign antibodies to fully human ones—each step broke through the limits of the previous stage, but the underlying logic of 'passively screening from diverse libraries' never changed.

The core value of computational methods lies precisely in that they begin to rewrite this decades-old underlying logic. When an antibody's affinity, stability, and developability can be initially predicted at the sequence design stage, when affinity maturation no longer relies on random mutation but on structure-guided rational design, and when candidate screening not only asks 'how strongly does it bind' but also 'can it be developed into a drug'—antibody discovery gradually shifts from a screening-centered probability science to a design-centered protein engineering discipline.

This doesn't mean screening is no longer important; rather, the core goal of screening changes from 'finding that one in billions of possibilities' to 'validating the one we've purposely designed.' Faced with the vast sequence space, what we're learning isn't just how to stumble upon the answer by chance, but how to define the answer rationally.