How AI, subcutaneous dosing, and new formats are multiplying the data streams that R&D teams must track
Since the first U.S. Food and Drug Administration (FDA) approval of a monoclonal antibody therapeutic in 1986, this drug class has experienced a meteoric rise, now representing approximately a quarter of annual FDA approvals, and four of the top 10 anticipated best-selling drugs of 2026.1 Overall, the global monoclonal antibody market has been valued at $267.6 billion and is projected to grow at a 9.3% compound annual growth rate (CAGR) to 2033.2 It’s hardly surprising that such significant market evolution is accompanied by equally significant innovation; in format (bispecific antibodies, antibody fragments, and antibody–drug conjugates are already making waves in the market), in formulation (a move toward self-injectable dosing requires manufacturing and formulation adjustments), and in discovery, which is rapidly being transformed by artificial intelligence (AI). Yet as each bottleneck in the antibody development journey is tackled, it isn’t eliminated; just shifted further downstream, into validation, developability, and the clinic. Here, we posit that many of these bottlenecks are, at root, data problems; that innovation necessarily multiplies the volume of data that a team must manage and track; and that, currently, the science is moving faster than the infrastructure that supports it. We’ll argue that, as a consequence, the companies that will come out on top of this evolutionary shift in the antibody landscape aren’t necessarily those employing the most advanced science but those who successfully maximize their advantage through robust data management.
The Changing Face of Antibody Discovery
As the biologics market has expanded, enormous effort has been poured into antibody discovery in an attempt to keep the pipeline stacked with potential blockbusters. That effort has increasingly moved toward computational biology. Where discovery used to mean finding candidates that already existed in “the real world,” over the last few years, the definition of discovery has expanded to include methods for predicting, and then designing, de novo antibodies.
A significant step along this path came with the development of AlphaFold2, published in 2021.3 Developed by Google DeepMind, this deep-learning method predicts a protein’s three-dimensional structure from its amino acid sequence with near-experimental accuracy. Instantaneously, a decades-long bottleneck in drug discovery was unlocked, and the months (or even years) of effort required to solve a protein structure, the key determinant of its function, was reduced to minutes or hours of computational time. The hypervariable complementarity-determining region (CDR) loops of antibodies present specific structural challenges and prompted the development of antibody-specific structural prediction tools, including IgFold and ImmuneBuilder.4 Such monumental advances in our understanding of protein structure unlocked further computational development; the addition of a machine learning layer. This made it possible to predict how a protein’s structure might impact its functional properties.4
Thus, the stage was set for the generative revolution; all the elements in place for de novo antibody design in silico: the ability to define a target epitope and design a novel antibody from nothing. Bennett et al. demonstrated the feasibility of this approach in their 2026 Nature paper, in which they first used RFdiffusion to propose a 3D backbone shape for the antibody, sculpted to fit the target epitope.5 In the next step, they designed a CDR loop sequence that fit the proposed structural geometry, and then finally, independently predicted the structure of the final sequence, which they compared against the original geometrical structure proposed by RFdiffusion to confirm that the de novo sequence really did encode the intended shape. Bennett et al. used this methodology to design both single-domain antibodies (VHHs) and single-chain variable fragments (scFvs) against four different disease-relevant epitopes, with striking results. Cryo-electron microscopy (cryo-EM) confirmed that the proposed influenza hemagglutinin (HA) VHH design matched the physical antibody to a root-mean-square deviation (RMSD) of 1.45 Å, while the Clostridioides difficile toxin B (TcdB) scFv matched to 0.9 Å, with all six CDR loops agreeing to within 0.2–1.1 Å, demonstrating remarkable accuracy for structures with no antibody precedent in the training data.
The ability to design a completely novel antibody targeting a specific epitope from scratch in silico is a transformational approach to discovery, opening up the potential to generate an almost infinite number of prospective new antibodies. However, this hasn’t “unblocked” the discovery bottleneck; just moved it downstream. In this case, into validation and then affinity maturation. For each epitope Bennett et al. targeted, they generated around 9,000 designs, and while the design and generation steps are, relatively speaking, quick and cheap, it’s not yet possible to validate all of those designs computationally; a large number had to be physically synthesized and screened using yeast display, and the final shortlist evaluated for accuracy using cryo-EM. Over the last decade or so, cryo-EM has emerged as the gold standard for atom-level structural resolution of antibody-antigen complexes. However, this is a highly resource-intensive process, with typical campaigns taking weeks, or more commonly months, per structure.
This research also highlights an additional complication for de novo discovery of a therapeutic antibody. An antibody that fits neatly to the target is not necessarily one that binds effectively. In this instance, all the candidates Bennett et al. identified were weak binders, requiring extensive affinity maturation before they could be considered therapeutically useful. This is a fundamental consequence of a computational approach; generative design can now accurately predict how a de novo antibody’s structure will map to the target epitope, but not how tightly it will bind. This still requires iterative wet-lab optimization, where high-throughput display campaigns and next-generation sequencing are being deployed at scale to clear the bottlenecks that high-volume design creates in validation and affinity maturation. In doing so, this also generates otherwise scarce training data on antibody sequence, structure, function, and how the three relate.
The lack of such connected data currently limits the ability of AI-enabled workflows to purposefully design antibodies that balance structural “fit” with therapeutically relevant binding affinity, and appropriate developability characteristics, from scratch. There aren’t many antibody–antigen complex structures in public databases, and next-to-no standardized comparable affinity or other functional data, because this is typically generated in different labs, assays, and conditions. However, there are significant efforts underway to change this.
Drivers of Change in Developability Assessments
To improve the prediction of binding affinity, AbBiBench evaluates machine-learning models on how well they predict binding within the full antibody–antigen complex, concluding that inverse-folding models outperform sequence-only models at ranking and generating higher-affinity variants.6 Simultaneously, Arsiwala et al. have developed high-throughput wet-lab assays to measure properties important for antibody developability, including aggregation, viscosity, and thermostability, and organized the output of these experiments into the kind of large, standardized data sets that predictive models need to be trained on.7 Meanwhile, Absci’s Origin-1 attempts to address de novo generation, developability “scoring,” and computational affinity maturation at once.8
Together, these point toward a future where generative models propose antibody structures against a target epitope, while computational filters score for affinity, developability, or both, reducing the number of final candidates that must be manufactured and tested experimentally.
The Science Moves Fast, But the Market Moves Faster
The nature of science is that as one obstacle is overcome, the market shifts and reveals a whole new challenge. In this case, what is developable for a monoclonal antibody delivered by intravenous (IV) infusion isn’t necessarily developable for subcutaneous administration, the demand for which has been growing steadily over the past decade; unsurprising, considering the benefits it delivers to patients, healthcare systems, and drug developers alike.9 For example, subcutaneous pembrolizumab can be administered in as little as one minute,10 and, with the original formulation due to lose patent protection in the next few years, the subcutaneous version extends its exclusivity.
Between September 2024 and September 2025, the FDA approved subcutaneous formulations of three widely used checkpoint inhibitors, atezolizumab, nivolumab, and pembrolizumab, each originally given by intravenous infusion.11,12 There are two broad formulation approaches for delivering an antibody subcutaneously. The first is to concentrate it to 100–200 mg/mL, so that a therapeutic dose fits into a volume small enough for self-injection.13,14 The second is to co-formulate the antibody with hyaluronidase, an enzyme that enhances dispersion and permeability in the subcutaneous tissue, so that a larger volume can be given;10,13 all three of the checkpoint inhibitor approvals used the latter.10–12
Whichever route is chosen, the same levers are available: engineering the molecule for lower viscosity, adjusting the formulation, and choosing the delivery device. Salts and amino acids are the main excipients reported to reduce viscosity, with arginine hydrochloride showing particularly strong results.13 Even so, there are limits: protein–protein interactions have held commercial formulations to around 150 mg/mL, and injectability is generally taken to require a viscosity below approximately 20 centipoise (cP), a measure of resistance to flow.15
Viscosity, solubility, and aggregation at high concentration cannot yet be reliably predicted and must be measured experimentally. The difficulty is a mismatch between when this information is needed and when the material to generate it is available. At the formulation stage, identifying optimal conditions at high concentration demands large quantities of material, which makes the process slow and costly and risks delaying progress to the clinic.16 Ideally, such liabilities would be identified earlier, at candidate selection, but only small quantities of material are available at that point. This tension is driving the development of low-volume analytics and in silico tools that predict high-concentration behavior,16 and those tools in turn depend on standardized data, such as that generated by the high-throughput viscosity and aggregation assays of Arsiwala et al.7
Meanwhile, the science, and the market, are already introducing additional variables, this time in format. An increasing proportion of antibodies in late-stage development have non-canonical formats designed to enhance their biological activity.17 The market is following: the number of bispecific antibodies approved by the FDA doubled from seven to 14 in just two years.18 To add another layer of complication, several of the first bispecific T cell engagers to reach patients, including teclistamab and epcoritamab, are given subcutaneously,19,20 so the developability questions raised by a new format and those raised by high-concentration dosing are being asked of the same molecule.
Each non-conventional format raises its own developability questions. Bispecific antibodies, for example, can suffer from reduced biophysical stability where additional binding domains are fused to the scaffold, and from chain mispairing, which creates antibody-related impurities that are difficult to remove.21 Antibody-drug conjugates face a different challenge: the more hydrophobic the payload, the greater the destabilization of the antibody structure and the greater its tendency to aggregate.22 Aggregation and viscosity are also the properties that high-concentration dosing puts under the most strain, so a format that was developable for infusion has to be assessed again for subcutaneous delivery.
More to Do, in Less Time
These pressures are converging on the same teams, and on timelines that the funding environment is arguably compressing. With venture capital concentrating on later-stage, de-risked assets,23,24 early-stage companies have fewer chances to raise capital and less room to iterate before they must show the data that unlocks the next round. Generative design delivers thousands of candidates per target (around 9,000 per epitope in the Bennett et al. study), each of which needs validating. Developability assessment is moving upstream. Subcutaneous delivery adds high-concentration assays that must be run on small quantities of material. And every new format brings its own variants, whether chain pairings, linkers, or drug-to-antibody ratios (DAR), the average number of payload molecules attached to each antibody. More to do, in less time, with the same number of people.
This is exactly where problems occur, and where mistakes are made without a platform to take the burden off. When variants of a bispecific, or conjugates of one antibody at different DARs, carry near-identical names across spreadsheets, it becomes easy to attribute a result to the wrong molecule. When the conditions of a formulation study (concentration, pH, excipient, temperature, and timepoint) live in different files, or only in a notebook, results can’t be compared or reused to train the predictive models the field increasingly relies on.
None of this is new: the FAIR principles, which call for data to be Findable, Accessible, Interoperable, and Reusable, were published a decade ago,25 and poor reproducibility in U.S. preclinical research has been estimated to cost around US$28 billion a year.26 However, as the science gets more sophisticated, the data has become more convoluted. Discovery projects can span years and move through multiple team members. The complexity isn’t in recording or storing all these disparate data but in connecting them, such that three years and four senior scientists later, it’s still straightforward to reconstruct the decision-making process or trace a lead candidate back through its developability profile, affinity maturation data, experimental, and cryo-EM validation to its original predicted structure and sequence design.
The science will keep advancing, and the market will keep moving faster still; but each advance multiplies the data a team must connect, and 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; and a full audit trail that survives long after the scientists who generated the data have moved on.
About Lab Thread Ltd.
Founded in 2024 and headquartered in Silverstone Park, UK, Lab Thread Ltd develops next-generation software for life-science research laboratories. The platform addresses the challenge of fragmented workflows — where project management, lab operations and scientific data often live in separate silos. Lab Thread brings them together into a unified environment designed specifically for biologists.
References
1. Twenter, Paige. “10 drugs poised to be best-sellers in 2026.” Becker’s Hospital Review. 13 Mar. 2025.
2. Monoclonal antibodies market (2026–2033): Size, share and trends analysis report. Grand View Research. Accessed 26 Oct. 2026.
3. Jumper, John, et al. “Highly accurate protein structure prediction with AlphaFold.” Nature. 596: 583–589 (2021).
4. Zheng, Jiayao, et al. “The application of machine learning on antibody discovery and optimization.” Molecules. 29: 5923 (2024).
5. Bennett, Nathaniel R, et al. “Atomically accurate de novo design of antibodies with RFdiffusion.” Nature. 649: 183–193 (2026).
6. Zhao, Xinyan, et al. “AbBiBench: A benchmark for antibody binding affinity maturation and design.” arXiv. Preprint 2506.04235. 23 May 2025; revised 10 Oct. 2025.
7. Arsiwala, Ammar, et al. “A high-throughput platform for biophysical antibody developability assessment to enable AI/ML model training.” mAbs. 17: 2593055 (2025).
8. Levine, Simon, et al. “Origin-1: A generative AI platform for de novo antibody design against novel epitopes.” bioRxiv. Preprint. Jan. 2026.
9. Viola, Margarida, et al. “Subcutaneous delivery of monoclonal antibodies: How do we get there?” Journal of Controlled Release. 286: 301–314 (2018).
10. “European Commission Approves Subcutaneous Administration of KEYTRUDA® (pembrolizumab) for All Adult Indications Approved in the European Union.” Merck. 19 Nov. 2025.
11. “FDA approves pembrolizumab and berahyaluronidase alfa-pmph for subcutaneous injection.” U.S. Food and Drug Administration. 19 Sep. 2025.
12. Jaber, Nadia. “FDA approves injectable nivolumab, an alternative to IV infusion.” National Cancer Institute, Cancer Currents Blog. 19 Feb. 2025.
13. Mijangos, Laura RR, Stephen E Harding, and Nicholas J Darton. “Developing high-concentration monoclonal antibody formulations for subcutaneous administration to improve patient treatment.” Biophysical Reviews. 17: 1013–1031 (2025).
14. Mosca, Ilaria, et al. “Multiscale diffusion, dynamic cluster formation, and intermolecular interactions in pharmaceutically relevant monoclonal antibody formulations.” Molecular Pharmaceutics. 22: 5373–5388 (2025).
15. Ramesh, Anujan, et al. “Ultra-high concentration low-viscosity subcutaneous antibody formulations using ionic liquids.” Journal of Controlled Release. 388: 114295 (2025).
16. Zarzar, Jonathan, et al. “High concentration formulation developability approaches and considerations.” mAbs. 15: 2211185 (2023).
17. Crescioli, Silvia, et al. “Antibodies to watch in 2025.” mAbs. 17: 2443538 (2025).
18. Strohl, William R. “Structure and function of therapeutic antibodies approved by the US FDA in 2024.” Antibody Therapeutics. 8: 197–237 (2025).
19. “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.” Janssen Pharmaceutical Companies of Johnson & Johnson. 25 Oct. 2022.
20. Muley, Ajit. “FDA approves EPKINLY® for relapsed or refractory follicular lymphoma.” Cancer Research Institute. 28 Jun. 2024.
21. Madsen, Andreas V, et al. “Design and engineering of bispecific antibodies: Insights and practical considerations.” Frontiers in Bioengineering and Biotechnology. 12: 1352014 (2024).
22. Buecheler, Jakob W, et al. “Impact of payload hydrophobicity on the stability of antibody–drug conjugates.” Molecular Pharmaceutics. 15: 2656–2664 (2018).
23. “Q1 2026 biopharma licensing and venture report.” J.P. Morgan. Apr. 2026.
24. Taylor, Nick Paul. “Early-stage biotechs feel the squeeze as funding favors derisked assets: JPM.” BioSpace. 12 Jan. 2026.
25. Wilkinson, Mark D, et al. “The FAIR guiding principles for scientific data management and stewardship.” Scientific Data. 3: 160018 (2016).
26. Freedman, Leonard P, Iain M Cockburn, and Timothy S Simcoe. “The economics of reproducibility in preclinical research.” PLOS Biology. 13: e1002165 (2015).
