When a cancer research group needed ultra-high-affinity antibodies against an unstudied membrane protein — with no prior art, no commercial options, and an 8-week grant deadline — AntibodyLLM deployed its AI-driven CDR optimization platform paired with high-throughput wet lab screening to deliver a 50 pM binder that works across three assay formats.
The client, a university cancer biology lab, had identified a novel membrane protein consistently overexpressed in aggressive tumor samples. To validate its role as a therapeutic target and advance their grant-funded research program, they needed a high-quality research antibody — one that could reliably detect the protein in western blot, immunohistochemistry (IHC), and flow cytometry.
The problem: no commercial antibody existed. The protein had a structurally complex extracellular domain with limited immunogenic exposure, and prior attempts with a contract CRO using traditional hybridoma immunization had yielded only nanomolar-range binders — inadequate for the sensitivity required in IHC tissue sections at low expression levels.
The research team needed picomolar affinity (KD < 100 pM) to reliably distinguish target-positive from target-negative cells in flow cytometry, and they had 8 weeks before a grant reporting deadline.
Conventional immunization and hybridoma screening is largely a numbers game — produce thousands of clones, hope a few have good affinity. For structurally complex or weakly immunogenic targets, this approach routinely fails to cross the picomolar threshold.
Immunization generates antibodies to immunodominant epitopes — often not the functionally important binding surface. CDR sequences emerge stochastically, with no rational design for affinity.
Even large hybridoma campaigns screen hundreds of clones. The sequence space of possible CDR combinations is astronomically larger — most high-affinity solutions are never sampled.
CDR3 — the loop with the greatest structural diversity and the largest contribution to binding affinity — is the hardest to optimize by traditional means. It's also where the most value lies.
Two capabilities made this outcome possible: a computational AI platform that rationally designed CDR sequences with a structural bias toward high affinity, and a high-throughput wet lab screening pipeline that rapidly validated and refined the computational predictions into a real, working antibody.
AntibodyLLM's platform began by modeling the target protein's extracellular domain using structural prediction algorithms to identify high-value epitope surfaces — specifically regions with deep binding pockets accessible to CDR loops. This step alone is impossible in traditional immunization: the immune system cannot be directed to a specific epitope.
With a target epitope defined computationally, the AI then searched antibody sequence space to identify CDR combinations with predicted structural complementarity. The focus was deliberately placed on CDR3 of the heavy chain (HCDR3) — the longest, most structurally variable loop, and the primary determinant of binding specificity and affinity in most antibody-antigen interactions.
Rather than screening a random library, the AI generated a rationally constrained candidate set: ~200 antibody sequences with diverse HCDR3 sequences, each selected for predicted binding energy with the target epitope and compatibility with known stable antibody frameworks. This pre-filtering step eliminates the low-affinity noise that dominates random screening approaches.
| Parameter | Traditional Hybridoma | AntibodyLLM AI Design |
|---|---|---|
| CDR3 design strategy | Random somatic hypermutation | Structure-guided computational design |
| Candidate sequences evaluated | Thousands (random) | ~200 (pre-filtered, high-confidence) |
| Epitope targeting | Immunodominant (uncontrolled) | Rationally chosen functional epitope |
| Typical initial hit affinity | 1–100 nM | 100–500 pM (first round) |
| Path to picomolar affinity | Further immunization or phage display | Computational affinity maturation → wet validation |
Computational predictions are hypotheses, not guarantees. The AI-designed candidate set of ~200 sequences was immediately handed to AntibodyLLM's wet lab team for rapid experimental validation using a parallelized screening pipeline purpose-built for this workflow.
All 200 sequences were synthesized as gene fragments and expressed as scFv fragments in a 96-well HEK293 transient expression format within 5 days. Crude supernatants were screened in parallel by ELISA against the target antigen and two closely related family members (specificity counter-screen). Candidates with binding signal ≥ 10-fold over background on target and <2-fold on off-targets were promoted — this cut the field to 31 candidates in round one.
The 31 candidates were converted to full IgG1 format, expressed at small scale (5 mL), and subjected to biolayer interferometry (BLI) for quantitative KD measurement. This identified 8 clones with KD values between 80 and 420 pM — all within picomolar range. The top 3 clones were taken forward for application validation.
The top 8 BLI-confirmed binders from round one ranged from 80 to 420 pM. To push the best candidate into the sub-100 pM range required a second round of computational optimization — feeding the experimental binding data back into the AI model to refine its structural predictions.
Specifically, HCDR3 sequence variants of the top-performing clone were generated computationally by focusing on residues predicted to contribute most to binding energy. An additional 40 single- and double-point variants were synthesized and tested. Three of these showed further improved affinity, with the lead achieving a KD of 50 pM — a 4-fold improvement over the round-one starting point.
This closed-loop approach — AI design → wet validation → AI refinement → final validation — is fundamentally faster than iterative affinity maturation by phage display or error-prone PCR, which typically require multiple full library preparation and selection cycles.
Delivered at week 8 — validated in western blot, IHC, and flow cytometry with zero cross-reactivity to structurally related family members.
Concentrating computational design effort on HCDR3 — the loop that contributes most to affinity — is more efficient than optimizing all six CDR loops simultaneously. Targeted design + targeted screening delivers higher hit rates.
By pre-filtering sequence space computationally, the wet lab screened 200 pre-qualified candidates rather than thousands of random clones. Smaller, smarter library — higher affinity output with less material and time.
Feeding round-one BLI data back into the computational model allowed targeted affinity maturation in a second round — reaching 50 pM without a full library regeneration cycle. Speed advantage over traditional display methods is 3–4x.
Whether your target is structurally complex, weakly immunogenic, or simply hasn't yielded high-affinity binders by conventional methods — our AI-driven CDR design platform was built for exactly this problem.
Discuss Your Target