Engineering multispecific antibodies with complete killing selectivity through the closed-loop integration of machine learning and high-throughput experimentation

Abstract

On-target, off-tumor toxicities remain a major barrier for T-cell engagers in solid tumors. We present EVA^TM, a closed-loop design platform integrating high-throughput functional assays with multi-objective Bayesian optimization to explore combinatorial T-cell engager (TCE) spaces. In a HER2×CD3 case study, iterative design-build-test-learn cycles traversed 44,160 designs defined by valency, topology, affinity and spacing. Compared with a Sobol baseline, EVA achieved 14-fold enrichment of potent, tumor-selective candidates. Multiple architectures reached sub-10 pM potency on HER2-high cells, near-complete efficacy, and $\ge$10,000-fold selectivity over HER2-low models, consistent with avidity gating. EVA™ recovered diverse high-performing topologies and generalized to a second target, supporting density-gated avidity as a design principle and providing an operational template for rapid, data-efficient optimization.

Keywords: Artificial Intelligence; Avidity; HER2; Multispecific antibodies; TCEs; VHH; bayesian optimization; machine learning; nanobody; selectivity; single-domain antibodies.

Figure 1.

Modular platform, combining synthetic biology, robotic automation and ml with user-defined assay choices and experimental evaluation. DBTL cycles employed across drug discovery projects for TCE lead optimization for tumor selectivity. Each cycle is completed in less than 6 weeks.

Figure 2.

Model performance across 4 cycles. The top panel shows key metrics for the model over cycles as more data is collected. The bottom panel shows the predicted and true (normalized log) values for each objective (potency above and selectivity below) over cycles 1–4 (left to right; initialization cycle is excluded).

Figure 3.

Key results for the study. (a) Box and whisker plot showing performance of the top 25 designs per cycle reported by the compound metric in the activation assay. The dotted line denotes the clinical benchmark (Runimotamab). Black diamonds represent extreme outliers (see methods) (b) 14-fold enrichment of high-performing molecules identified using MOBO (blue dots) versus Sobol quasi-random baseline (pink dots). Boxed area is defined by higher potencies and selectivities than the starting molecule (a rationally designed bivalent HER2×CD3 (2 + 1 format)). (c-h) Show killing dose–response curves for Runimotamab and our top 5 candidates. T47D is our healthy cell model (26K HER2 receptors per cell). SKBR3 is our cancer cell model (359K HER2 receptors per cell). The HER2 density on the surface of T47D cells is similar to cultured cardiomyocytes. Percent killing was measured after 4 days with effector-to-target ratio = 10:1 (n = 2).

Figure 4.

Details for a second project against an undisclosed target. (a) This project involves a different antigen target and a substantial increase in the complexity of the design space compared to our pilot project. (b) We identify a number of candidates with complete killing selectivity profiles across the range accessible with our assay. This is in the context of a five-fold difference in receptor quantification.

Figure 5.

The composition of the design space for this work, comprising of 48 unique HER2 SdAbs, a single αCD3 SdAb and 10 unique linkers, over 12 topologies resulting in 44,160 unique candidate designs.