BOAT: Navigating the Sea of In Silico Predictors for Antibody Design via Multi-Objective Bayesian Optimization
Authors: Jackie Rao, Ferran Gonzalez Hernandez, Leon Gerard, Alexandra Gessner
Organizations: MRC Biostatistics Unit, University of Cambridge, UK · Centre for AI, Data Science and Artificial Intelligence, R&D, AstraZeneca, Barcelona, Spain
Antibody lead optimization is inherently a multi-objective challenge in drug discovery. Achieving a balance between different drug-like properties is crucial for the development of viable candidates, and this search becomes exponentially challenging as desired properties grow. The ever-growing zoo of sophisticated in silico tools for predicting antibody properties calls for an efficient joint optimization procedure to overcome resource-intensive sequential filtering pipelines. We present BOAT, a versatile Bayesian optimization framework for multi-property antibody engineering. Our `plug-and-play' framework couples uncertainty-aware surrogate modeling with a genetic algorithm to jointly optimize various predicted antibody traits while enabling efficient exploration of sequence space. Through systematic benchmarking against genetic algorithms and newer generative learning approaches, we demonstrate competitive performance with state-of-the-art methods for multi-objective protein optimization. We identify clear regimes where surrogate-driven optimization outperforms expensive generative approaches and establish practical limits imposed by sequence dimensionality and oracle costs.
Figures & tables
Figure 1: Illustration of the BOAT loop for lead optimization (LO). LO campaigns typically consist of a few rounds of experimental testing of the order of a hundred sequences in the wetlab, where previous experiments inform the sequence design for the next iteration of the wetlab loop. The BOAT loop can be thought of as an inner loop in a single sequence design step of the wetlab loop, leveraging sophisticated tools used for computationally assessing sequences. The multi-objective optimization loop eliminates the common practice of generating a large pool of sequences and filtering them down by sequentially passing them through the different oracles, which is computationally inefficient and unaware of Pareto trade-offs. The oracles used depend on the requirements of the LO campaign, and can be exchanged flexibly. See Section 2 for details on the internal building blocks of BOAT.
Figure 2: Hypervolume vs. iterations (mean ± s.e., 10 seeds) for CDR1–3 under 2, 3, and 4 objectives. BOAT variants (q indicates batch acquisition functions) outperforms GA‑sum and NSGA‑II. Red dashed lines show ground‑truth hypervolume where available.
Figure 3: Discovered vs. ground truth Pareto fronts for 2-objective CDR optimization with 5 mutations. Each panel shows the best-performing seed for GA methods (left) and BOAT (right). We can see that even the most successful GA runs fall behind BOAT.
Figure 4: PLM evolution for the first 300 generations of 3-objective CDR3 optimization, with the best PLM score of all generated sequences and the mean score of the top 3 PLM scores recorded. Results averaged over 10 seeds with standard error. Note that the x-axis starts at 100 to omit the initial sequences.
Figure 5: (a) Scatter plot of hypervolume versus Shannon entropy for all seeds, methods, and every 100 iterations for CDR3. Initial solutions are highlighted in gold. Each point represents the diversity and multi-objective performance of a population at a given optimization step. (b) Evolution of Shannon entropy over optimization iterations for CDR3. Results averaged over 10 seeds with standard error bands.
Figure 6: Hypervolume evolution for 3-objective CDR optimization. Initial HV has been subtracted. Results averaged over 5 seeds with standard error.
Figure 7: Comparing relative hypervolume versus number of generated sequences between BOAT and LaMBO-2. The reference hypervolume of 1 has been assigned to the parental sequence.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Scatter plots of hypervolume versus Shannon entropy for CDR1 and CDR2 optimization. Each point represents the diversity and multi-objective performance of a population at a given optimization step, showing results for all seeds, methods, and every 100 iterations. Initial solutions are highlighted in gold.
Figure 9: Evolution of Shannon entropy over optimization iterations for CDR1, CDR2, and CDR3 regions. Results averaged over 10 seeds with standard error bands, showing how population diversity changes throughout the optimization process across all three CDR regions. Note BOAT-qEHVI performs almost exactly in line with BOAT-qNEHVI.
Figure 10: Plots comparing the hypervolume of the Pareto front by iteration of the BO algorithm for different encodings considered in Section 4.1 .
Figure 11: Plot comparing Pareto front by seed across different conditions for each CDR. All seeds are shown. It is clear especially in CDR3 runs that GAs tend to explore smaller Pareto fronts.
Figure 12: Plot comparing Pareto front after 300 generated sequences (200 oracle calls after 100 initial sequences) by seed across different conditions for each CDR. All seeds are shown. Early in the algorithm (especially for CDR3), BOAT is able to explore much closer to the ground truth Pareto front compared to GAs.
Figure 13: Plots for experiments when ablating batch size using the qEHVI acquisition function for CDR3 with 2 objectives and 5 maximum mutations, using batch sizes 2, 4, 8, 10. Other experimental details are kept the same as before. For (b), as before, each point represents the diversity and multi-objective performance of a population at a given optimization step, showing results for all seeds and batch sizes, and every 100 iterations.
Figure 14: Plot comparing Pareto front by seed across different conditions for each CDR. All seeds are shown. Similarly to the 5 mutations example, GAs visually do not explore as much of the ground truth Pareto front compared to BOAT.
Figure 15: Multi-seed comparison of BOAT and LaMBO showing final Pareto fronts for BOAT-EHVI, BOAT-qEHVI, BOAT-qNEHVI, and all LaMBO designs across 5 optimization runs. While LaMBO got to optimize the 2 objectives for affinity and expression, the BOAT variants additionally optimized for naturalness (ESM-2 likelihood). As the hypervolumes are fully comparable across seeds, we sorted them by descending terminal hypervolume for each method.
Figure 16: Hypervolume progression for LaMBO-2 and BOAT variants when omitting naturalness constraints on BOAT.
Figure 17: Multi-seed comparison of BOAT and LaMBO showing final Pareto fronts for BOAT-EHVI, BOAT-qEHVI, BOAT-qNEHVI, and all LaMBO designs across 5 optimization runs, where both LaMBO and BOAT got to optimize the two objectives of affinity and expression predicted by LaMBO’s discriminative head. As the hypervolumes are fully comparable across seeds, we sorted them by descending terminal hypervolume for each method.
LIAC, ISIC, EPFL, Lausanne, Switzerland · Department of Chemistry, University of Zurich, Zurich, Switzerland · Centre of Competence in Research (NCCR) Catalysis, Zurich, Switzerland