AI-Designed Antibodies in Human Trials: Evidence Tiers, Case Studies, and What's Next

For years, the promise of AI in drug discovery was measured in benchmark datasets and press releases. That phase is ending. As of 2025–2026, multiple programs where AI played a genuine role in molecular design have cleared the first-in-human threshold and generated clinical data. Among these, the most strategically important for the biologics field are the programs built on generative AI antibody design—where the sequence itself originates from a machine learning model rather than from immunization or library display.

This article examines the verified clinical cases, establishes a framework for grading what "AI-designed" actually means, and draws lessons for organizations considering AI antibody design as part of their discovery strategy.

What Does "AI-Designed" Actually Mean?

Before evaluating clinical evidence, it is essential to establish a taxonomy. The phrase "AI-designed drug" covers at least five distinct roles, and conflating them leads to systematically inflated or deflated assessments of the technology's maturity.

1. AI target discovery: Machine learning identifies a novel disease-relevant target (e.g., Insilico Medicine's PandaOmics flagging TNIK for fibrosis). The molecule itself may be designed conventionally.
2. AI de novo molecule generation: A generative model proposes novel sequences or structures that have not appeared in training data. This is the most technically demanding and highest-value category.
3. AI-guided optimization: AI selects from or ranks sequences derived from conventional discovery (e.g., affinity maturation of an existing hit using Bayesian optimization).
4. AI-assisted drug repurposing: AI identifies new indications for approved molecules (e.g., BenevolentAI's identification of baricitinib for COVID-19). No new molecular entity is designed.
5. AI trial design / patient stratification: AI optimizes clinical protocol design but plays no role in the molecular entity.

For evaluating the clinical progress of AI antibody design, categories 2 and 3 are the relevant benchmarks. Category 4—while commercially valuable—is not antibody design. The cases below are graded against this framework.

AI-Generated Antibodies That Have Reached Human Trials

Generate:Biomedicines — GB-0669

GB-0669 is publicly described by Generate:Biomedicines as an AI-generated viral neutralizing antibody—a program that went from computational design to Phase 1 first-in-human dosing. Phase 1 data in healthy adults demonstrated acceptable safety and a PK/PD profile consistent with therapeutic levels. This makes GB-0669 one of the first publicly documented cases where a generative AI model proposed the full antibody sequence that subsequently entered and completed Phase 1 in humans.

Key characteristics of this program:

• The antibody was generated by a deep learning model, not selected from an immune repertoire or phage display library
• The target is a viral antigen, a class where structural data and epitope definition are comparatively tractable
• Phase 1 safety data supports progression—the critical first proof that AI-generated antibody sequences can yield manufacturable, tolerable biologics in humans

Absci — ABS-101 and ABS-201

Absci's clinical pipeline offers two distinct antibody-design programs that have reached Phase 1 as of 2025–2026:

• ABS-101 (TL1A antibody, IBD): A TL1A-targeting antibody for inflammatory bowel disease. Absci uses a zero-shot generative design platform that integrates large language models trained on protein sequence databases with wet-lab screening loops. ABS-101 entered Phase 1 dosing as a direct output of this platform.
• ABS-201 (androgenetic alopecia): A distinct target in dermatology. ABS-201 received first-in-human dosing in Phase 1/2a. The program highlights that AI antibody design is not confined to oncology or infectious disease—the technology is beginning to address a broader range of therapeutic areas.

Both programs are noteworthy because Absci's platform is explicitly positioned as de novo generative design—the sequences are not derived from animal immunization. This matters for intellectual property (no hybridoma patents to navigate) and for the scalability of the design-build-test loop.

The Strongest End-to-End Evidence: Small Molecules as a Proof of Concept

While the antibody programs above represent genuine milestones, the most rigorously documented AI-designed clinical candidate to date is a small molecule: rentosertib (ISM001-055), developed by Insilico Medicine.

What makes this case the current gold standard:

• AI played a role at both ends: PandaOmics for target identification (TNIK, a kinase involved in fibrotic pathways) and Chemistry42 for generative molecule design and optimization
• The program progressed through Phase 1 and into a randomized, double-blind, placebo-controlled Phase 2a trial in 71 patients with idiopathic pulmonary fibrosis (IPF)
• Results were published in Nature Medicine in 2025, with a positive FVC signal in the 60 mg QD arm
• This is the first AI-designed molecule supported by a peer-reviewed Phase 2a RCT in a high-quality journal

For antibody design, rentosertib serves as proof that the end-to-end AI design paradigm—from target to clinic—is technically feasible and can generate reproducible, publishable clinical evidence. The antibody field is approximately 2–3 years behind in terms of published Phase 2 data, but the trajectory is clear.

Sources: Nature Medicine 2025 — rentosertib Phase 2a RCT; Nature Biotechnology 2024 — TNIK inhibitor preclinical/clinical

Why AI Antibody Design Is Harder Than Small Molecule AI Design

The small molecule and antibody design problems share the same computational paradigm—generative models, structure prediction, multi-objective optimization—but differ in ways that matter practically:

Sequence Space and Structural Complexity

An antibody variable domain contains ~110–130 amino acids, with 6 CDR loops determining most of the binding specificity. The combinatorial sequence space (20^~35 for CDRs alone) is vastly larger than typical drug-like chemical space. Generative models must learn the coupled constraints of VH/VL pairing, CDR loop geometry, framework stability, and antigen complementarity simultaneously. Structure prediction tools—AlphaFold-Multimer, IgFold, ABodyBuilder2—have dramatically improved CDR loop modeling, but the accuracy gap for long CDR-H3 loops remains an active research area.

Immunogenicity and Humanization

AI-generated antibody sequences may contain T-cell epitopes or unusual sequence features that increase immunogenicity risk in human patients. Computational deimmunization tools (e.g., EpiSweep, Optizyme) can flag high-risk peptide sequences, but in silico prediction of immunogenicity for de novo sequences still carries significant uncertainty. This is an area where tight integration between AI design and experimental assays—humanness scoring, MHC-II binding prediction, PBMC stimulation assays—is critical before clinical advancement.

Expressibility and CMC

A computationally elegant antibody sequence that cannot be expressed at adequate yield in CHO cells, or that aggregates under formulation conditions, has no path to the clinic. AI antibody design workflows that do not include expressibility-aware scoring—or that are not tightly coupled to a CHO expression and characterization platform—will generate a high rate of false positives at the computational stage. This is why AI antibody design services that are vertically integrated with CHO expression infrastructure have a significant practical advantage over pure computational platforms.

Bispecific and Multi-Domain Formats

The antibody field has moved beyond monospecific IgG toward bispecific, trispecific, and antibody-drug conjugate formats. Generative AI for bispecifics must simultaneously optimize two binding interfaces, correct chain pairing, and linker geometry—a substantially harder problem than monospecific design. Progress is being made (several bispecific-specific design tools have been published since 2024), but bispecific AI design remains at an earlier maturity level than monospecific design.

The Broader AI Drug Discovery Landscape: Key Players to Watch

Beyond antibody-specific programs, the broader AI drug discovery ecosystem provides important context for where the technology is heading:

• Iambic Therapeutics (IAM1363): A highly selective AI-designed HER2 inhibitor (small molecule) that has completed Phase 1 dose escalation. The program is notable for its brain-penetrant design objective—a multi-parameter optimization challenge that plays to AI's strengths.
• Recursion + Exscientia: The merged platform combines large-scale phenotypic screening (Recursion OS) with AI-guided small molecule design (Exscientia's Centaur Chemist). REC-1245 progressed from target identification to IND in under 18 months—a benchmark that, if replicated across their pipeline, would represent a structural shift in development timelines.
• BenevolentAI + baricitinib: The clearest example of AI-assisted drug repurposing resulting in an FDA-approved indication (COVID-19 hospitalized adults). This is AI-assisted repurposing, not de novo design, but it validates the clinical impact of AI-guided hypothesis generation at the indication level.
• Exscientia + DSP-1181: The first publicly documented AI-designed small molecule to enter human clinical trials (Phase 1 for OCD, 2020). Discovery phase completed in under 12 months vs. the ~4.5-year industry average—the milestone that established the commercial viability of AI-designed clinical candidates as a category.

Five Lessons from the First Wave of AI-Designed Clinical Candidates

1. AI accelerates discovery but does not change clinical attrition rates—yet. The first wave of AI-designed candidates is clearing Phase 1 safety bars, but Phase 2 efficacy data is still accumulating. Rentosertib's Phase 2a signal is encouraging, but a larger Phase 2b/3 is required. The fundamental biology of drug development has not changed; AI has compressed the front end.
2. End-to-end integration outperforms point solutions. The programs with the most complete evidence chains—Insilico Medicine's rentosertib, Absci's pipeline—use AI at multiple stages, not just at the design step. Platforms that cover target identification, sequence generation, and hit-to-lead optimization in an integrated loop produce more defensible claims.
3. Data quality remains the rate-limiting step. All generative AI antibody design platforms are bounded by the quality and diversity of their training data. Programs with access to proprietary assay data—binding kinetics, thermal stability, expression yields, in vivo PK—generate models with meaningfully better predictive accuracy than those trained exclusively on public sequence databases.
4. Expressibility and manufacturability must be co-optimized at design time. For antibodies specifically, a sequence that scores well on affinity and selectivity but poorly on CHO expression yield creates a bottleneck that no downstream optimization can fully resolve. Stable cell line development constraints should inform the design objective function from day one.
5. The taxonomy of "AI-designed" will increasingly matter for IP and regulatory purposes. As AI-designed biologics become more common, the distinction between AI-generated sequences (novel IP position) and AI-optimized sequences (dependent on prior art) will have significant downstream consequences for freedom-to-operate and patent strategy. Clear documentation of the AI's role in generating or selecting the clinical candidate sequence is becoming a regulatory and commercial hygiene requirement.

What the Next Three Years Will Look Like for AI Antibody Design

Based on the clinical programs currently in Phase 1 and Phase 2, the field is on track for several milestones by 2028:

• The first Phase 2 efficacy read-out for an AI-generated antibody (likely in oncology or immunology, given the current pipeline distribution)
• Generative AI tools for bispecific antibody design reaching clinical-candidate quality on a routine basis
• Regulatory agencies (FDA, EMA) issuing clearer guidance on documentation requirements for AI-derived sequences—likely requiring design provenance as part of IND submissions
• Convergence of AI sequence design with CRISPR-based stable cell line development: AI designs the sequence; CRISPR-integrated CHO cell lines provide the manufacturing vehicle. This end-to-end digital-to-physical pipeline represents the next competitive moat in contract biologics.

AntibodyLLM's AI antibody design service is built around exactly this integrated model—combining generative AI sequence design with our proprietary UCOE-stabilized CHO expression platform and CRISPR site-specific integration capability. Explore our technology platform to understand how the computational and manufacturing layers connect.

Frequently Asked Questions

What is AI antibody design?

AI antibody design is the use of machine learning models—including large language models, generative neural networks, and structure-prediction algorithms—to computationally design novel antibody sequences with desired properties such as high affinity, selectivity, bispecificity, or improved developability. The approach differs from traditional hybridoma or phage display methods by generating candidate sequences in silico before experimental validation, significantly compressing the discovery timeline.

Have any AI-designed antibodies entered human clinical trials?

Yes. As of 2025–2026, at least two AI-generated antibody programs have reached Phase 1 with publicly available data. Generate:Biomedicines' GB-0669, described as an AI-generated viral neutralizing antibody, reported Phase 1 safety and PK/PD data in healthy adults. Absci's ABS-101 (a TL1A antibody for IBD) and ABS-201 (androgenetic alopecia) have both entered Phase 1/2a studies. These represent the first wave of generative-AI-origin biologics to enter the clinic.

What is the difference between AI-generated and AI-assisted antibody design?

AI-generated antibody design means the model proposes novel sequences de novo—the CDR sequences or full variable regions are computationally generated and have not existed in prior databases. AI-assisted design means AI is used to screen, optimize, or prioritize sequences initially derived from conventional discovery methods (immunization, display libraries). Both are valid approaches but should not be conflated when evaluating clinical evidence or patent landscapes.

How does AI antibody design compare to hybridoma technology in speed?

Traditional hybridoma workflows typically require 6–12 months from immunization to a validated lead candidate. AI antibody design platforms can compress the in silico discovery phase to weeks, generating and ranking hundreds of thousands of candidate sequences before any wet-lab work begins. Exscientia's published data for DSP-1181 showed a 4-fold reduction in discovery time for a small molecule; antibody-specific platforms report similar acceleration for the computational phase. Downstream expression, developability testing, and CMC development timelines remain comparable to conventional approaches.

What role does AlphaFold play in AI antibody design?

AlphaFold2 and its successors (AlphaFold-Multimer, RoseTTAFold-All-Atom) predict antibody–antigen complex structures with near-crystallographic accuracy in many cases. In AI antibody design workflows, structure prediction is used to filter generated sequences by predicted binding geometry, model CDR loop conformations, assess epitope coverage, and guide affinity maturation. The combination of generative sequence models and structure prediction has become the de facto standard for computational antibody design as of 2025.

What challenges remain before AI-designed antibodies can be broadly translated to the clinic?

Several challenges persist: (1) Predicting immunogenicity in vivo remains difficult—computational deimmunization tools have limitations for truly novel sequences. (2) Expressibility and manufacturability of AI-generated sequences are not guaranteed; sequences that score well computationally can have poor CHO expression yields. (3) Off-target binding is harder to predict for antibodies than for small molecules due to the large binding interface. (4) Clinical attrition rates for AI-designed antibodies are not yet well-characterized. Integrating AI design with downstream expression and stability testing early in the workflow is the key mitigation strategy.

Can AI design bispecific antibodies?

Yes. Bispecific antibody design is a particularly well-suited application for AI because it requires simultaneous optimization of two binding interfaces, chain pairing compatibility, and linker geometry—a problem that is computationally tractable but experimentally expensive to screen exhaustively. Generative models trained on bispecific formats can propose sequences with correct heavy/light chain pairing and reduced chain-mispairing propensity. Several AI platforms now include bispecific-specific design modules, though this remains at an earlier maturity level than monospecific design.