Monoclonal antibodies are one of the workhorses of biopharmaceutical development, with over 100 FDA-approved drugs and well-established manufacturing, regulatory, and clinical-development pathways. Yet conventional antibody discovery remains hampered by mounting costs and long timelines, typically six to twelve months to get from a target to a lead candidate.
By designing and characterizing therapeutic antibodies computationally, AI promises to make development cheaper, faster, and more flexible. But scientific questions abound.
Development of an antibody-based drug hinges on three factors: the best binding site on the target, which candidates bind to it most tightly, and whether any of them can survive manufacturing and the clinic. For each, the field has predictive models that do well on familiar targets and assays but considerably worse on unfamiliar ones. Benchmarks built around in-distribution accuracy have made that gap difficult to measure — and to close.
Three papers from our science team at Amazon Bio Discovery, an AI-powered application that gives scientists access to biological AI models and integrated lab services to design and test novel drug candidates, tackle research questions about each of these three factors. Two are peer-reviewed journal papers on prediction: ranking candidates by binding strength and flexibly predicting developability. The third brings prediction into an end-to-end design process, navigates the selection of binding sites with an agent, and delivers experimentally validated antibody hits against a novel cancer target.
Ranking binders from sequence alone
One of the biggest questions in antibody design is which candidates bind the best. In "A systematic evaluation framework for universal antibody-antigen binding affinity prediction and candidate recommendation", published in iScience, we propose a new framework to assess binding affinity predictors and train a new sequence-based predictor, MochiBind.
Most affinity predictors are evaluated on their ability to predict the absolute binding affinity, on antigens that appear in their training data, against test sets that contain few or no nonbinders. Each of these characteristics makes the evaluation easier than the intended application. Absolute affinity values are not comparable across assays, and performance degrades for antigens the model has not seen. The practical use case, meanwhile, involves ranking a pool of thousands of candidates, most of which don’t bind to the target at all, to pick the ones worth testing in the lab. Surveying seven prior studies, we found that none satisfied all the conditions necessary to train a reliable universal predictor.
We therefore reframed the task. Rather than predicting an absolute number, MochiBind predicts which of two antibodies against the same antigen binds more tightly. We begin by using a pretrained protein language model (ESM-2) to embed residues of antibody-antigen complexes in a representational space. We then compute the mean of each complex’s residue embeddings, to give it a single embedding.
A specially trained network layer projects these embeddings into a lower-dimensional space, and predicts relative binding strength from the difference between the two projections.[HL2] Pairwise comparisons are then aggregated into a global ranking over the candidate pool using TrueSkill, a Bayesian rating algorithm originally developed for ranking video game players based on match outcomes. No structural input is required at any stage.
This formulation has two practical advantages: relative orderings are more consistent across assays than absolute values, so the training signal is less sensitive to measurement noise, and the output is the ranked list the discovery process needs.
Our paper also presents a novel evaluation framework. We used the AlphaBind dataset, which covers four antigen systems (targeting TIGIT, PD-1, HER2, and theSARS-CoV-1 RBD) with roughly 30,000 experimentally characterized variants for each and pairwise sequence similarity between antigens that’s close to zero. The protocol is strictly cross-antigen: train on two antigens, validate on a third, and test on the fourth, rotating so that each serves as the held-out system once. We then standardized two metrics: (1) pairwise accuracy and (2) retrieval accuracy and precision at top K, which measure how many of a model's K recommendations are experimentally confirmed strong binders.
MochiBind achieved higher pairwise accuracy than every structure-based baseline on all four held-out antigens, outperforming the closest competitor by almost 10% on average. In terms of ranking performance, MochiBind also achieved the highest retrieval accuracy on all four antigens and the highest retrieval precision (lowest false-positive rate) on three out of four. It also scored 200,000 antibody pairs in roughly 13 seconds on a CPU, a more than 100-fold inference speedup over competing methods that should enable the screening of very large design libraries.
Learning to predict antibody properties in context
Proteins that bind tightly to their targets but clump together or degrade in the bloodstream or provoke an immune response are not effective or safe as drugs. Most attempts to predict such properties from biological data encounter the same problem: batch effects, or systematic differences in the way different labs handle samples or conduct experiments that lead to predictable deviations in measurement — deviations known as batch offsets. A model fine-tuned on one lab's data quietly inherits its offsets.
In "Context-aware multi-property antibody predictor: A novel framework integrating text and protein language models", in npj Systems Biology and Applications, we address batch effects during inference. Our model — the context-aware multiproperty antibody predictor, or CA-MAP — takes a prompt containing a variable number of example antibodies with their measured properties, followed by a query antibody and the name of the property to predict. When the examples come from the same lab as the query, their measured properties capture the batch offset. The model’s input — its context — thus includes the information it needs to adjust for batch effects without retraining.
Getting a model to use that context, however, is not straightforward. A model trained on data from a single source can learn to ignore the examples — whose measurements are systematically skewed, after all — and rely on the query sequence alone. Our training strategy, AB-context-aware, prevents this by applying a hidden random transformation to both the context properties and the expected answer, resampled for every prompt. Under this scheme, the transformation can be recovered only from the context, so the model must use it.
We measured the effect on a fine-tuned domain-specific multimodal LLM, TxGemma, predicting hydrophobicity. Without batch effects, standard fine-tuning and AB-context-aware training perform comparably, a correlation with ground truth of 0.99 (according to Spearman’s rank correlation coefficient, where 1 is perfect correlation). With a simulated additive batch effect in the 0–0.3 range, standard fine tuning falls to a 0.58 correlation, while the context-aware model remains at 0.99.
CA-MAP has a relatively small multimodal architecture combining text and proteins. Sequences (encoded with ESM-2), property names (encoded with sentence embeddings), and numerical values each have dedicated encoders and projectors, and a state space model based on the sequence-modeling architecture MAMBA composes them. Trained on a synthetic dataset of 876,898 antibody-heavy chains covering six developability properties, CA-MAP achieves a Spearman correlation (denoted ρ) greater than 0.8 on several of them and outperforms the fine-tuned TxGemma baseline across all four properties tested jointly.
The architecture is also considerably cheaper to train and run, with roughly 182,000 trainable parameters to TxGemma’s 40 million, and it’s about 200 times as fast per prompt at inference.
Because properties are specified as text, CA-MAP can also be queried for properties absent from its training data. In one set of experiments, we trained CA-MAP on only four of the dataset’s six developability properties and tested it on the other two (positive-charge heterogeneity, or PosCh, and immunogenicity). When we used only the two target properties as context, immunogenicity prediction reached ρ = 0.25; with all six correlated properties in the context, ρ = 0.73. PosCh improved from ρ = 0.08 to ρ = 0.73 under the same comparison. These gains indicate that the model is drawing on correlations between developability properties, which suggests that expensive assays could be estimated in part from cheaper ones.
Designing antibodies with AI, validating them in the lab
In our third paper, "Agent-guided de novo design of nanobody binders against a novel cancer target", which was presented as a Spotlight at the ICML 2026 Workshop on Generative and Agentic AI for Biology and received the Best Paper Runner-Up Award, we bring predictive and generative antibody models together to design therapeutic nanobodies from scratch in a real drug discovery project.
The target antigen for the design project — or “campaign”, as it’s known in the industry — was chosen to reflect real clinical need: a cell surface target for desmoplastic small round-cell tumors, a rare and aggressive pediatric cancer. Our collaborators at the Dr. Nai-Kong V. Cheung’s Lab at Memorial Sloan Kettering Cancer Center in New York identified it by sequencing patient tumor specimens for proteins that (1) sit on the tumor cell surface, (2) are driven by a specific genetic error, and (3) are largely absent from healthy tissue. The target has no experimental structure and no public antibody information, so there was no template to graft, no prior campaign to affinity-mature from, and no possibility that the design models encountered this antigen during training.
One of the key decisions at the outset of a de novo design campaign is which specific regions on the antigen surface, known as epitopes or hotspots, to target. We designed a hotspot recommendation agent that orchestrates seven bioinformatics tools, which do things like determine solvent-accessible surface area, secondary structure, hydrophobicity, and sequence uniqueness against user-specified negative targets; match epitopes against 500,000 entries in NIAID’s Immune Epitope Database; and annotate domains according to the categories in the protein families (Pfam) database. Our model synthesizes these tools’ outputs into hotspot recommendations with an explicit biophysical rationale for each.
Grounding the recommendations in deterministic tool outputs focuses the search on evidence-supported regions rather than relying on the model's parametric knowledge of protein biology. Evaluated on antibody-antigen complexes from the SAbDab benchmark, the agent recovered at least one true epitope residue within its top five proposed regions about 80% of the time on a diverse holdout set. For the target antigen in our design campaign, it proposed eight hotspot regions.
We then used three generative models with different design principles — RFantibody (diffusion over protein backbones), IgGM (joint sequence-structure diffusion), and mBER (backpropagation through a structure prediction model) — to generate antibody designs that target those hotspots. Each model produced 96,000 designs, and each design was scored on properties like folding confidence (how likely the antibody is to fold into the shape necessary to bind to the target), complex quality (how likely the antibody is to form the correct binding interface with the target), and sequence liabilities (how likely the antibody sequence is to cause development or manufacturing problems), and MochiBind's sequence-based affinity estimate. Our candidate selection agent applied multi-objective Pareto filtering to ensure the retention of designs excelling on different metric combinations, and it prioritized 100,000 candidates for experimental screening.
Each candidate was synthesized and displayed on the surface of a yeast cell to be screened for whether it stuck to the target, and the designs that stuck most strongly were carried forward through two rounds of sorting and filtering. None of the 116 candidates that survived these rounds bound to an unrelated control protein, indicating that they bind specifically to the intended target, rather than being generally sticky. All 116 were then individually measured to determine how tightly they bind to the target antigen, and 46 were identified as strong binders.
These 46 binders, along with the binder and nonbinder labels from the full screen, become training data for the next design cycle: a lab-in-the-loop workflow where each round of experiments sharpens the models that propose the following round. Amazon is uniquely well positioned to run that loop , with the scientific expertise to build foundational ML for biology, the computational capacity to design and score hundreds of thousands of candidates, and a path to deliver these methods, including those like MochiBind and CA-MAP that aren’t available today, to customers through Amazon Bio Discovery, an AI-powered application that connects these biological AI models with integrated lab services so scientists can move from design to experimental validation in a single workflow.