Back to Blog
Anti-PD-1 Antibody Engineering: What 18 Globally Approved Antibodies Reveal About Epitope, Affinity, and Blocking Activity
Anti-PD-1 antibody engineering is the process of designing antibodies that block the interaction between the checkpoint receptor PD-1 (Programmed Death-1) on T cells and its ligand PD-L1 on tumor cells, thereby reactivating suppressed immune responses. Over 10 years (2014–2026), 18 anti-PD-1 antibodies have been approved by four major regulatory agencies — FDA, EMA, NMPA, and PMDA — making the PD-1/PD-L1 axis the most comprehensively characterized therapeutic antibody target in oncology. A 2026 systematic review by Almagro et al. in Frontiers in Immunology (17:1834585) assembled all available sequence, affinity, blocking, and crystal structure data for these 18 drugs, revealing engineering patterns invisible when studying any single antibody in isolation. This article summarizes the most actionable findings for antibody engineers and biologics developers.
Understanding the therapeutic mechanism clarifies what "a good anti-PD-1 antibody" means in practice. When tumor-expressed PD-L1 engages PD-1 on a cytotoxic T cell, the intracellular domain of PD-1 recruits SHP2 phosphatase into inhibitory microclusters. SHP2 preferentially dephosphorylates CD28, blocking the PI3K–AKT–mTOR pathway. The result is reduced glucose metabolism, impaired proliferation, and — with chronic antigen exposure — progressive transcriptional rewiring driven by TOX and NR4A transcription factors. T cells enter a state of exhaustion with partial epigenetic reprogramming that limits reversibility.
Anti-PD-1 antibodies interrupt this at the very first step: they physically block PD-1 from contacting PD-L1, preventing SHP2 recruitment and restoring CD28 co-stimulation. This makes the primary design objective clear — the antibody must sterically occlude the PD-1/PD-L1 binding interface. Affinity per se is secondary to interface coverage.
Pembrolizumab and nivolumab were both approved in 2014. In the decade that followed, 16 additional antibodies — the majority from Chinese developers targeting NMPA approval — reached market. The technology evolution follows a clear arc:
The common thread across all generations is an increasingly computational engineering phase: structure-guided humanization scoring, in silico developability filtering, and — in the most recent programs — AI-assisted CDR optimization. Discovery diversified; engineering converged on computational methods.
The PD-1 extracellular domain folds into an IgV topology with ten β-strands forming two β-sheets. The front β-sheet (strands CC'FGG' plus the FG loop) is the ligand-binding face where PD-L1 docks in an orthogonal IgV/IgV geometry. Four N-glycosylation sites decorate the domain, with the N58 site on the BC loop being functionally important for several approved antibodies.
Overlaying 22 antibody/PD-1 crystal structures (15 from approved drugs) onto a common PD-1 coordinate frame identifies three functional contact zones:
One structural rule with major engineering consequences: all approved antagonistic antibodies bind the distal membrane-facing surface of PD-1. Antibodies that bind near the membrane transmit agonist signals — the opposite of the intended therapeutic effect. Membrane-proximal binding therefore represents a hard exclusion criterion, not a matter of optimization.
The clearest and most counterintuitive finding in the Almagro et al. dataset is the weak correlation between equilibrium dissociation constant (KD) and functional blocking potency (IC50 for PD-1/PD-L1 disruption). The following representative data illustrate the point:
| Antibody | KD (nM) | IC50 PD-1/PD-L1 (nM) | KD/IC50 ratio |
|---|---|---|---|
| Toripalimab | 0.10 | 1.30 | 13× |
| Cemiplimab | 1.68 | 0.60 | 0.4× |
| Zimberelimab | 0.18 | 0.58 | 3.2× |
| Sintilimab | 0.32 | 29.2 | 91× |
| Pembrolizumab | 0.20 | 0.63 | 3.2× |
| Nivolumab | 0.14 | 6.60 | 47× |
The comparison between sintilimab and zimberelimab is particularly striking: their KD values differ by less than 2-fold (0.32 vs. 0.18 nM), yet their IC50 values differ by approximately 50-fold (29.2 vs. 0.58 nM). Similarly, toripalimab — among the tightest binders at 0.10 nM KD — is one of the weaker functional blockers at IC50 = 1.30 nM, while cemiplimab achieves better blocking (IC50 = 0.60 nM) at 13-fold lower affinity.
The mechanistic explanation is structural: toripalimab's unusually long HCDR3 (16 residues — an outlier for this target class) forms a cavity that wraps around the FG loop, creating high affinity through multiple hydrogen bonds, but the contact residues overlap only partially with the PD-L1 binding surface. Cemiplimab engages the C'D loop and FG loop region with superior surface coverage of the PD-L1 footprint despite lower overall affinity.
The engineering implication is direct: optimizing binding free energy (ΔG) without simultaneously maximizing spatial overlap with the PD-L1 interface produces candidates with impressive affinity numbers and modest functional activity. Epitope coverage should be a co-optimization target, not a post-selection quality check.
Multiple sequence alignment of heavy and light chain variable regions from the 16 approved antibodies with available sequence data reveals convergent germline gene selection that is far too consistent to be coincidental. Three patterns stand out:
These patterns are not arbitrary. They represent the evolutionary filter of clinical development: the antibodies that reached approval are the survivors of a selection process that — unknowingly or by design — favored specific structural features compatible with both PD-1 binding and pharmaceutical manufacturability. AI-assisted antibody design can encode these structural priors explicitly from the start, rather than re-discovering them empirically through failed leads. Our AI antibody design service incorporates germline selection and HCDR3 length priors derived from approved biologics databases to bias generation toward developable candidates.
One of the more technically important findings from the crystal structure analysis concerns N-glycosylation at position 58 on PD-1's BC loop. For cemiplimab, camrelizumab, and penpulimab, HCDR2 residues make direct contacts with the N58 sugar chain. The glycan is structurally integrated into the binding interface — not a passive modification flanking the epitope.
This creates a reproducible experimental artifact: affinity measurements (SPR, ITC, BLI) performed using deglycosylated PD-1 recombinant protein will systematically underestimate binding for these three antibodies, and any assay that relies on deglycosylated antigen will misrank candidates during discovery. The practical implication extends to antibody discovery programs: when engineering antibodies against Region 2 of PD-1, use fully glycosylated antigen for primary screening and not E. coli-expressed or PNGase F-treated protein.
Two bispecific antibodies targeting PD-1 have received NMPA approval, offering the most direct evidence of how validated anti-PD-1 modules are being deployed in next-generation engineering.
Cadonilimab (AK104) — PD-1 × CTLA-4: A tetravalent (2+2) IgG-like format. Two PD-1-binding domains (derived from penpulimab's variable regions) are combined with two CTLA-4 scFv domains (from quavonlimab) via (G4S)₃ linkers. The Fc region carries LALA + G237A mutations to abrogate Fcγ receptor engagement, eliminating ADCC/CDC effector functions that would deplete Treg cells expressing CTLA-4 in normal tissue — a safety concern raised by earlier CTLA-4 antibodies.
Ivonescimab (AK112) — PD-1 × VEGF: Penpulimab's PD-1-binding scFv is fused to the C-terminus of bevacizumab's VEGF-targeting IgG1 heavy chain with an LALA Fc mutation. The resulting molecule achieves KD = 0.25 nM for PD-1 and KD = 0.33 nM for VEGF — affinities comparable to the parental monoclonals. Comparative clinical data show median progression-free survival extended from 5.8 months (pembrolizumab alone) to 11.1 months.
Both bispecifics share a key engineering philosophy: the PD-1-binding module was not re-engineered for the bispecific format — it was transplanted intact from an approved monoclonal. Engineering effort was invested in the format architecture (linker design, Fc engineering, spatial arrangement of the two binding arms) rather than re-optimizing the PD-1 binder from scratch. This approach minimizes the introduction of new developability liabilities and leverages an existing clinical safety profile. It is a principle directly applicable to any bispecific design program: re-use validated modules where possible, engineering novelty only where necessary.
For teams considering bispecific development, our monoclonal antibody production capabilities include experience with complex multi-domain formats, Fc engineering, and expression optimization for bispecific IgG constructs.
Synthesizing the structural, sequence, and functional data from the full approved anti-PD-1 dataset yields five principles with broad applicability beyond the PD-1 target:
The 37 PD-1 crystal structures in the Protein Data Bank as of 2026 — including 25 antibody complexes — represent an exceptional structural training dataset. AI tools including AlphaFold3, ProteinMPNN, and RFdiffusion can leverage this data to model binding modes for novel CDR sequences, predict epitope coverage without requiring a new crystal structure, screen large CDR libraries for FG-loop engagement, and identify HCDR3 conformations compatible with the known pocket geometry.
The key insight from the Almagro et al. review is that these tools work best when biological priors — epitope requirements, loop length constraints, germline preferences — are explicitly encoded into the generation and ranking process. AI that searches the full antibody sequence space without structural constraints will rediscover, inefficiently, the same patterns that 25 years of experimental work already mapped. AI that starts from validated structural priors can focus its search where the data says candidates are most likely to succeed.
For programs targeting PD-1 or related immune checkpoint receptors, combining structural bioinformatics (epitope mapping from available structures), AI sequence generation with built-in developability filtering, and experimental validation in native-glycosylation assay formats represents the most efficient current path from concept to candidate. The 18 approved antibodies are not just therapeutic products — they are a curated engineering dataset that any new anti-PD-1 or checkpoint antibody program should study before screening the first construct.
Primary source: Almagro JC, Lund J, Peterson NA, Shelton D, Benson N, Seefeldt M, Chaparro-Riggers J. "Engineering strategies and binding mechanisms of therapeutic anti-PD-1 antibodies approved by regulatory agencies globally." Frontiers in Immunology 17:1834585 (2026). doi:10.3389/fimmu.2026.1834585
Anti-PD-1 antibody therapy is a form of cancer immunotherapy that reactivates tumor-suppressed T cells by blocking the PD-1/PD-L1 inhibitory axis. PD-1 is a checkpoint receptor on T cells that, when engaged by PD-L1 on tumor cells, recruits SHP2 phosphatase and shuts down CD28 co-stimulation — progressively driving T cells into a state of functional exhaustion. Therapeutic anti-PD-1 antibodies physically block PD-1 from engaging PD-L1, restoring T cell effector function. As of 2026, 18 anti-PD-1 antibodies have been approved globally by FDA, EMA, NMPA, and PMDA across multiple tumor types including melanoma, NSCLC, and gastric cancer.
Affinity (KD) measures binding tightness; blocking activity (IC50) measures how effectively the antibody prevents PD-L1 from engaging PD-1. These are governed by different structural determinants. An antibody can bind PD-1 very tightly at a site with little spatial overlap with the PD-L1 contact surface, yielding high affinity but weak blocking. For example, toripalimab has a KD of 0.10 nM but an IC50 of 1.30 nM, while cemiplimab achieves better blocking (IC50 = 0.60 nM) at 13-fold lower affinity (KD = 1.68 nM). The most striking case: sintilimab and zimberelimab have nearly identical affinities (0.32 vs 0.18 nM) but their blocking IC50 values differ 50-fold. Epitope coverage of the PD-L1 binding footprint is a more reliable predictor of functional potency than KD.
Analysis of 22 anti-PD-1 antibody crystal structures identifies three functional contact zones on the PD-1 extracellular domain. Region 1 (residues 25–43, N-terminal loop + β-strand A'A) is engaged by nivolumab with limited PD-L1 overlap. Region 2 (residues 59–100, C'D loop + N58 glycosylation site) is targeted by cemiplimab, camrelizumab, and penpulimab with partial PD-L1 overlap; the N58 glycan directly contacts HCDR2 residues in these antibodies. Region 3 (FG loop, residues 121–144) has the highest spatial overlap with the PD-L1 binding surface and is targeted by the majority of high-potency blockers. Critically, all approved antagonistic antibodies bind the distal membrane-facing surface — membrane-proximal binding results in PD-1 agonism, the opposite therapeutic effect.
Multiple sequence alignment of 16 approved anti-PD-1 antibodies reveals strong convergence: 75% use IGKV1 or IGKV3 family light chains, and heavy chains predominantly derive from IGHV1, IGHV3, or IGHV4 families. These germline genes have high natural B-cell repertoire frequency and encode frameworks with favorable solubility and thermal stability — properties validated at clinical manufacturing scale. Heavy chain humanness ranges from 77% (pembrolizumab) to 96% (dostarlimab). Average HCDR3 length is ~7 residues, roughly half the human antibody average, reflecting geometric constraints of the PD-1 binding pocket. Toripalimab's 16-residue HCDR3 is the outlier; its extended loop enables a distinct binding mode at the cost of reduced blocking efficiency.
Two NMPA-approved bispecifics demonstrate different combination strategies. Cadonilimab (AK104) is a tetravalent PD-1 × CTLA-4 bispecific — penpulimab's PD-1 binding domains combined with quavonlimab's CTLA-4 scFvs, on an LALA+G237A Fc to minimize effector function. Ivonescimab (AK112) combines penpulimab's PD-1 scFv with bevacizumab's VEGF-targeting domain (LALA Fc), achieving KD = 0.25 nM for PD-1 and 0.33 nM for VEGF. Clinical data show ivonescimab extended median PFS from 5.8 to 11.1 months versus pembrolizumab alone. Critically, both bispecifics transplant validated PD-1 binding modules intact from approved monoclonals — engineering novelty is confined to the bispecific architecture, preserving existing safety and manufacturability data.
For cemiplimab, camrelizumab, and penpulimab, HCDR2 residues make direct contacts with the N58 glycan on PD-1's BC loop. The glycan is a structural participant in the binding interface, not a passive bystander. As a practical consequence, SPR or BLI affinity measurements using deglycosylated PD-1 protein (E. coli-expressed or PNGase F-treated) systematically underestimate binding affinity for these antibodies. Any anti-PD-1 discovery program should perform primary screening with fully glycosylated, mammalian cell-expressed PD-1 antigen, and specifically confirm glycosylation status by mass spectrometry before interpreting affinity data. This issue extends to all targets with glycosylated binding interfaces — the N58 example is a well-documented case that generalizes broadly.
The 37 PD-1 crystal structures in PDB (25 antibody complexes) constitute a high-quality structural training dataset. AI tools including AlphaFold3, ProteinMPNN, and RFdiffusion can use this data to predict epitope coverage for novel CDR sequences, enforce HCDR3 length priors consistent with the PD-1 pocket geometry (4–8 residues), screen large sequence libraries for FG-loop engagement before synthesis, and score candidates for developability using germline-derived frameworks (IGHV1/3/4 + IGKV1/3). The critical point from the 18-antibody dataset is that AI works most efficiently when structural and biological priors are explicitly encoded — not when searching the full unconstrained antibody sequence space. AI that integrates known epitope constraints and loop length distributions reduces the search space by orders of magnitude and focuses design toward candidates most likely to achieve both high affinity and potent functional blocking.
Our AI-powered platform encodes structural priors from validated clinical antibodies to bias your program toward high-affinity, high-blocking candidates from day one.
Start Your Project