AI Antibody Discovery: From Designed Binders to Validated Drug Candidates
The most telling AI drug discovery news this month was not the largest deal. On 9 October, GSK expanded its collaboration with Chai Discovery, and the order of events matters. Chai designed protein binders against a panel of GSK targets without any target-specific training, GSK tested them in its own laboratories and found binders for every target, and only then did GSK widen its access to Chai's platform [1]. The deal followed the data.
For biopharma scientists who have kept AI-driven drug discovery (AIDD) at arm's length, this shows a large company putting generative models into real programs. For AI companies with no laboratory, it shows what partners now expect before they sign.
October 2026 AI Drug Discovery Deals at a Glance
GSK was not alone. Five announcements in two weeks touched almost every stage of discovery.
*Terms as stated in company announcements [1–5]. Milestones are maximum potential payments, not committed funding.
The other four deals push AI in different directions. AstraZeneca's agreement with insitro places AI at the start of the pipeline, using the Virtual Human platform to find new targets in MASH and cardiomyopathies [2]. Generate:Biomedicines and Zealand Pharma will combine generative design with high-throughput testing to build multispecific proteins for obesity [3], while METiS will use AI to design ionizable lipids that steer LNPs toward specific tissues for Lilly [4]. Lilly has also opened insitro-built PK models trained on decades of its own preclinical data to selected biotechs through TuneLab [5]. None of these programs has clinical data yet, and goals such as better tissue selectivity are not yet results.
What the GSK and Chai Discovery Deal Shows About AI Protein Design
The GSK results are particularly encouraging in the context of Chai’s earlier findings. When Chai introduced Chai-2 in 2025, it tested up to 20 designs per target across 52 targets and identified binders for approximately half of them [6]. The GSK evaluation offers an encouraging example of AI-generated binders being tested independently within a pharmaceutical research organization. Identifying binders for every target tested in GSK’s laboratories, with what GSK’s Christopher Austin described as “strong binding affinities” [1], marks a meaningful step forward. GSK now joins Lilly, Novartis, Pfizer and Bristol Myers Squibb among Chai’s 2026 partners [7].
![Figure 1. 52 antigens targeted by Chai-2. Blue boxes indicate targets with at least one successful binder [6].](https://web.cdn.cyagen.com/cyagen-en-v2/pictures/162b8ca4b53694d0735067cafae1b630.webp)
Still, identifying a binder for every target tested does not mean that every design was successful. The announcement does not disclose the number of targets, the per-design hit rate, functional assay data or in vivo results. Without details of the target panel, it is also unclear how broadly these findings extend to more challenging targets, such as multi-pass membrane proteins. As Fierce Biotech noted, the approach has yet to be validated in the clinic [7].
Why an AI-Designed Binder Is Not Yet a Drug Candidate
Binding is only the first step. AI-designed binders span different molecular formats, but antibody discovery illustrates the validation required. An ELISA hit needs confirmation of binding in a biologically relevant context, kinetic characterization by BLI or SPR, and functional testing. Similar affinities can mask very different effects, including receptor blockade, activation or little functional activity.
Each lead also needs developability assessment, despite encouraging early findings for de novo antibodies [8]. In vivo studies provide evidence of pharmacological activity, with translational relevance depending on the model and mechanism. When species differences prevent engagement of the mouse target, humanized target models may support more relevant evaluation. AI can accelerate binder generation. Advancing those binders still depends on experimental validation.
How Bench Scientists and AI Companies Can Close the Lab Loop
For bench scientists who have not yet tried AIDD, GSK's approach is a sensible template, and it does not require building a model. Choose a few targets you know well, ask for a fixed number of designs, and run them through the same assays as your current leads. Your own assays make the most honest test set.
AI companies without a wet lab face the opposite problem. Partners will ask for native-target binding, function, developability and in vivo data they cannot generate themselves. For these teams the laboratory is effectively part of the product: fast turnaround, the same assays in every design round, and machine-readable results, negatives included, that can be used to retrain the model.
How Cyagen Connects AI Antibody Design with Wet-Lab Validation
This is the gap Cyagen works in. GSK can test AI designs in-house; many AI-native biotechs cannot, and many biologics groups have no AI capability of their own. For AI teams, Cyagen's antibody discovery CRO acts as the wet-lab arm. Sequences come in, and decision-ready data goes back, covering recombinant expression, ELISA and flow cytometry binding, BLI kinetics, epitope competition, blocking and functional assays, and in vivo efficacy in humanized target models. Because the data is returned in a structured format, the same dataset can support a go/no-go decision and feed the next design round. Where appropriate, data formats and reporting requirements can be aligned with a partner's computational workflow.
For teams that want a second source of leads, immunization is still hard to beat. Immunization-based antibody discovery provides a complementary approach to computational design. Fully human antibody mouse platforms leverage biological diversification and affinity maturation to generate antibody repertoires, offering an additional source of candidates for subsequent functional and developability assessment. Cyagen's HUGO-Ab™ platform includes HUGO-Mab™ for conventional antibodies, HUGO-Light™ for common-light-chain bispecifics and HUGO-Nano™ for heavy-chain-only antibodies. HUGO-Ab-eKO™ adds a knockout of the mouse target for conserved antigens, where immune tolerance would otherwise blunt the response.
Cyagen's AbSeek™ AI-enabled analysis tools can support antibody sequence evaluation, structural assessment, and lead optimization alongside experimental screening and validation, so fewer clones need to go to the bench. For bench teams new to AIDD, this brings AI into a program without building it in-house. With design support, antibody generation, testing and in vivo work under one partner, AI-designed and mouse-derived antibodies can be compared head to head in the same assays and to the same data standards.
👉 [Download the HUGO-Ab™ Fully Human Antibody Discovery Brochure]
What October's Deals Mean for AI-Driven Antibody Discovery
October's deals show AI moving into every stage of discovery on terms that experiments set. AI-designed binders can now hold up in a partner's laboratory; the next evidence will have to show function, developability and in vivo efficacy, one lead at a time. Models can generate the molecules, but experiments still decide which ones become medicines, and the teams that keep the loop between the two short will move fastest.
👉 [Explore Our Antibody Library with Ready-To-Access Molecules for Drug Development]
👉 [Consult with Our Experts to evaluate your antibody designs]
References
[1] Chai Discovery. Chai Discovery announces collaboration with GSK following successful wet-lab evaluation [Internet]. Business Wire; 2026 Oct 9 [cited 2026 Oct 20].
[2] insitro. insitro announces a strategic collaboration with AstraZeneca for AI-enabled discovery of novel therapeutics for MASH and cardiomyopathies [Internet]. insitro; 2026 Oct 15 [cited 2026 Oct 20].
[3] Generate:Biomedicines. Generate:Biomedicines and Zealand Pharma enter collaboration and license agreement to develop protein therapeutics engineered with AI for metabolic diseases [Internet]. Generate:Biomedicines; 2026 Oct 13 [cited 2026 Oct 20].
[4] METiS TechBio. METiS TechBio announces strategic collaboration with Lilly to advance next-generation tissue-specific LNP delivery technologies [Internet]. METiS TechBio; 2026 Oct 19 [cited 2026 Oct 20].
[5] insitro. insitro-designed machine learning models for small-molecule in vivo pharmacokinetic behavior prediction now available in Lilly TuneLab [Internet]. insitro; 2026 Oct 6 [cited 2026 Oct 20].
[6] Chai Discovery Team. Zero-shot antibody design in a 24-well plate. bioRxiv [Preprint]. 2025 Jul [cited 2026 Oct 20]. doi: 10.1101/2025.07.05.663018.
[7] Taylor NP. GSK drinks in Chai Discovery's AI models after wet-lab validation. Fierce Biotech [Internet]. 2026 Oct 9 [cited 2026 Oct 20].
[8] Latent Labs Team, Kenlay H, Pretorius D, Crabbé J, Bridgland A, Schmon SM, et al. Drug-like antibodies with low immunogenicity in human panels designed with Latent-X2. arXiv [Preprint]. 2025 Dec 23 [cited 2026 Oct 20]. doi: 10.48550/arXiv.2512.20263.
