Does antibody development have a data problem?

Since the first US Food and Drug Administration (FDA) approval of a monoclonal antibody therapeutic in 1986, this drug class has experienced a meteoric rise. The global monoclonal antibody market has been valued at $267.6 billion, with a projected compound annual growth rate (CAGR) of 9.3% to 2033 (1). Growth on this scale brings innovation in format, in formulation and in discovery, which artificial intelligence (AI) is rapidly transforming.
Yet as each bottleneck in antibody development is tackled, it isn’t eliminated; it shifts downstream, into validation, developability and the clinic. Many of these bottlenecks are, at root, data problems. Innovation multiplies the data a team must manage, and the science is moving faster than the infrastructure that supports it. The companies that come out on top won’t necessarily be those using the most advanced science, but those who make the most of it through robust data management.

The changing face of antibody discovery

Discovery used to mean finding antibodies that already existed. Now, it increasingly means designing entirely new (de novo) ones. A major step came with AlphaFold2 in 2021: a deep-learning method from Google DeepMind that predicts a protein’s three-dimensional structure from its amino-acid sequence with near-experimental accuracy (2). Antibody-specific structure prediction tools and machine-learning models that link structure to function soon followed.
That set the stage for generative design. In their 2026 Nature paper, Bennett et al. used RFdiffusion to propose antibody backbones shaped to fit a chosen target site (epitope), designed the binding loops (complementarity-determining regions, or CDRs) to match, and confirmed by cryo-electron microscopy (cryo-EM) that the designs were accurate to within 1.5 ångströms (3).

‍
But the bottleneck has simply moved downstream. For each epitope, the team generated around 9,000 designs, too many to validate computationally, so an extensive shortlist still needed to be synthesised and screened in the lab, and since cryo-EM takes weeks or months per structure, this must be reserved for validating the most likely candidates. Most importantly, accurate fit doesn’t equate to reliable binding: every candidate was a weak binder, needing extensive affinity maturation before it could be therapeutically useful (3).

Drivers of change in developability assessments

At present, generative design can predict how an antibody will fit its target, but not how tightly it will bind, or how developable it will be. Part of the reason is data: public databases hold few antibody-antigen complex structures, and next to no standardised affinity or functional data, because that gets generated in different labs, with different assays, under different conditions. But this is starting to change. Arsiwala et al., for example, have built high-throughput assays for aggregation, viscosity and thermostability, producing the large, standardised datasets that predictive models need to train on (4). Others are developing new benchmarks for affinity prediction, and platforms that combine antibody generation with developability scoring, pointing towards a future where additional computational filters eventually reduce the number of candidates that must be made and tested, although the wet lab-in-the-loop remains a critical element of any validation process.

The science moves fast, but the market moves faster

As one question resolves, a new one arises. Namely, that what is developable for an antibody given by intravenous (IV) infusion isn’t necessarily developable for subcutaneous injection, which is fast moving into the mainstream, as evidenced by the September 2025 FDA approval of a subcutaneous formulation of the checkpoint inhibitor pembrolizumab (5).

There are two broad approaches to formulating for sub-cut administration: concentrate the antibody to 100–200 mg/mL, so a dose fits a self-injectable volume, or co-formulate it with hyaluronidase, an enzyme that helps a larger volume disperse under the skin (6). Either way, viscosity, solubility and aggregation at high concentration often can’t be predicted, and so must be measured, but high-concentration studies need large quantities of material that just don’t exist at candidate selection, when such problems would ideally be caught. Low-volume analytics and in silico prediction tools can help, but they in turn depend on having standardised data to train on, such as that generated by Arsiwala et al. (4).

Format adds another variable. The number of bispecific antibodies approved by the FDA doubled from seven to fourteen in just two years (7). Several of the first bispecific T-cell engagers, including teclistamab, are given subcutaneously (8), so format and high-concentration questions are being asked of the same molecule. Each new format brings its own developability issues, from chain mispairing in bispecifics to aggregation in antibody-drug conjugates (ADCs).

More to do, in less time

These pressures converge on the same teams, while venture capital concentrates on later-stage, de-risked assets (9), leaving early-stage companies less room to iterate. Generative design delivers thousands of candidates per target. Developability assessment is moving upstream. Subcutaneous delivery adds high-concentration assays on small quantities of material. And every new format brings variants, whether chain pairings, linkers or drug-to-antibody ratios (DARs). More to do, in less time, with the same number of people.

This is where mistakes happen. When conjugates at different DARs carry near-identical names across spreadsheets, it’s easy to attribute a result to the wrong molecule. When formulation conditions (concentration, pH, excipient, temperature and timepoint) live in separate files, or only in a notebook, results can’t be compared, or reused to train predictive models. None of this is new. The FAIR principles (data should be Findable, Accessible, Interoperable and Reusable) are well established, and poor reproducibility in US preclinical research has been estimated to cost around US$28 billion a year (10). But the challenge isn’t storing data; it’s connecting it, so that three years and four senior scientists later, a team can still trace a lead candidate back through its developability, affinity maturation and cryo-EM validation to its original design.

The science will keep advancing, and each advance multiplies the data a team must connect. The teams best placed to benefit will be those whose data layer keeps pace. In practice, that means:
•          a single record linking each candidate’s sequence, structure, construct, batch, assay conditions, analysis and results
•          information captured in a structured form at the point of work, so that it is comparable and ready for model training
•          a full audit trail that survives long after the scientists who generated the data have moved on.

‍

This article was first published in full in Pharma’s Almanac, where the full reference list is also available.

‍

References

1. Grand View Research. Monoclonal antibodies market (2026–2033). https://www.grandviewresearch.com/industry-analysis/monoclonal-antibodies-market

2. Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583–589. https://doi.org/10.1038/s41586-021-03819-2

3. Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature. 2026;649:183–193. https://doi.org/10.1038/s41586-025-09721-5

4. Arsiwala A, Bhatt R, van Niekerk L, Quintero-Cadena P, Ao X, Rosenbaum A, et al. A high-throughput platform for biophysical antibody developability assessment to enable AI/ML model training. mAbs. 2025;17(1):2593055. https://doi.org/10.1080/19420862.2025.2593055

5. US Food and Drug Administration. FDA approves pembrolizumab and berahyaluronidase alfa-pmph for subcutaneous injection. 19 September 2025. https://www.fda.gov/drugs/resources-information-approved-drugs/fda-approves-pembrolizumab-and-berahyaluronidase-alfa-pmph-subcutaneous-injection

6. Mijangos L, Harding S, Darton N. Developing high-concentration monoclonal antibody formulations for subcutaneous administration to improve patient treatment. Biophys Rev. 2025. https://doi.org/10.1007/s12551-025-01346-2

7. Strohl WR. Structure and function of therapeutic antibodies approved by the US FDA in 2024. Antib Ther. 2025;8(3):197. https://doi.org/10.1093/abt/tbaf014

8. Johnson & Johnson (Janssen). FDA approves TECVAYLI (teclistamab-cqyv), the first bispecific T-cell engager antibody for the treatment of patients with relapsed or refractory multiple myeloma [press release]. October 2022. https://www.jnj.com/u-s-fda-approves-tecvayli-teclistamab-cqyv-the-first-bispecific-t-cell-engager-antibody-for-the-treatment-of-patients-with-relapsed-or-refractory-multiple-myeloma

9. J.P. Morgan. Q1 2026 biopharma licensing and venture report. April 2026. https://www.jpmorgan.com/content/dam/jpmorgan/documents/cb/insights/outlook/jpm-biopharma-deck-q1-2026.pdf

10. Freedman LP, Cockburn IM, Simcoe TS. The economics of reproducibility in preclinical research. PLoS Biol. 2015;13(6):e1002165. https://doi.org/10.1371/journal.pbio.1002165

‍