Omnii-1 in the Loop: Accelerating Antibody Design with Experimental Feedback

An antibody's defining function is to bind a specific molecular target with high affinity. Yet strong binding alone does not make an antibody useful.

A successful therapeutic binder must recognize a specific molecular surface, such as an epitope on a disease-associated receptor, and bind in a way that produces the desired biological effect, such as blocking receptor signaling. It must also express efficiently, remain stable, and resist aggregation during manufacturing and storage. These requirements can compete: mutations that strengthen binding by introducing hydrophobic residues, for example, may also increase aggregation or reduce solubility, making the binder harder to produce and develop.

Antibody design is further complicated by the structural diversity of antibody formats. Conventional IgG antibodies contain two heavy and two light chains, while single-domain antibodies, commonly called VHHs or nanobodies, use a single heavy-chain variable domain to recognize their targets. The work described here focuses on single-domain binders.^bb The domains in Figure 2: VH and VL are the variable domains of the heavy and light chain. The six CDR loops they carry, three from each, form the paratope, which is the surface that contacts the antigen. CH1 and CL are the constant domains behind them, and the two arms hinge off the Fc, the constant stem that determines half-life and effector function. A VHH is a single variable domain with three CDRs and no light chain, so CDR3 does most of the binding.

These constraints can also be highly application-specific, encompassing specialized stability requirements, unconventional molecular architectures, or other design objectives that existing protein engineering tools are not equipped to optimize.

Our approach combines Omnii-1's protein sequence generation capabilities with predictive models and experimental validation.

Rather than designing antibodies in a single computational step, we use an iterative process in which each experimental round informs the next.

The workflow begins with a collection of experimentally characterized binders. We identify promising parent sequences based on their measured properties, including binding affinity and expression. Omnii-1 then proposes variants by identifying and resampling positions within those sequences.

Omnii-1 trait predictors evaluate the proposed variants and prioritize candidates across multiple objectives. A selected set is synthesized and tested in the lab, producing new measurements that inform subsequent rounds of candidate selection and model improvement.

Omnii-1 can explore variations around promising sequences, while experimental assays determine which proposed changes actually improve the properties of interest. As measurements accumulate, the system gains increasingly informative starting points and additional data for improving its predictive and generative components.

Since mid-September 2026, we have completed three rounds of antibody design and experimental validation with Aptamino. In each round, Omnii-1 uses patterns learned during training to generate biologically informed variations of promising experimentally characterized sequences. Omnii-1 then predicts relevant functional properties from sequence to help prioritize candidates for experimental validation. Selected designs undergo laboratory measurement for properties including expression and binding affinity. These experimental results inform subsequent rounds, allowing the design process to progressively refine promising candidates.

The campaign began with 1,620 designs generated by earlier pipelines, before Omnii-1 was introduced into the optimization process. Of these, 150 had measured binding affinities for the target receptor.^cc Binding affinity is reported as a dissociation constant, Kd: the concentration of antibody at which half the target is bound. A smaller Kd means a tighter binder, since less antibody is needed to occupy the target, and it is measured in nanomolar (nM) here. Because smaller is better, the axis in Figure 4 descends, so tighter binders sit higher on the plot. Omnii-1 was then used to guide three successive rounds of experimental optimization.

The resulting designs extend the observed Pareto frontier^dd The Pareto frontier represents the set of solutions for which improving one objective necessarily requires compromising another. of binding affinity and expression, demonstrating that the optimization process can identify candidates with more favorable combinations of these properties than those present in the starting cohort. Notably, two designs from the third experimental round exhibited binding stronger than the assay could resolve, establishing an experimental limit on their measured affinity rather than an exact Kd value. Improvements have also been observed in additional partner-held properties that are not reported here.

These findings are especially encouraging because Omnii-1 continues to improve throughout the process, using experimental feedback to refine the model and promising candidates from earlier rounds as starting points for further optimization. The experiment demonstrates the feasibility of using a general-purpose multimodal genome language model in an iterative experimental workflow, progressively exploring higher-performing regions of sequence space.

More broadly, Omnii-1's multimodal capabilities open the door to antibody design informed by additional biological context, including protein structure and other molecular inputs. These capabilities also create opportunities to combine sequence generation with predictions of binding, stability, and other properties within a single framework.

These early results from a single partner campaign provide a promising foundation for extending this approach to more complex antibody design challenges and other biological design problems.

Omnii-1 is pretrained broadly on biological sequence data, giving it representations of protein sequence patterns and biochemical regularities. However, antibody sequences were not included as a dedicated domain in its original pretraining mixture.

Adapting a general-purpose model to antibodies presents several challenges. Antibody repertoires contain extensive sequence redundancy, with many examples differing at only a handful of positions. Naively training on such datasets risks overemphasizing common sequence families or memorizing closely related examples instead of learning representations that generalize to novel binders.

To specialize Omnii-1, we use continued mid-training and post-training on a mixture of three data sources:

1. General protein sequences: A subset of the original pretraining distribution, included to preserve the model's broader protein knowledge.
2. Public antibody sequences: Antibody repertoire data that exposes the model to antibody-specific sequence patterns.
3. Proprietary partner designs: Experimentally characterized binder sequences and associated measurements relevant to the optimization task.

We additionally use cluster-aware training and evaluation splits^ee To reduce data leakage via homology, we cluster antibody and designed-binder datasets by sequence similarity and hold some clusters out of training for evaluation. to reduce sequence similarity between training and held-out evaluation examples. This is particularly important for antibody data, where conventional random splits can substantially overestimate generalization by allowing near-identical sequences to appear in both sets.

The goal of this specialization is to augment Omnii-1's broad biological priors with bespoke knowledge relevant to the design problem.

Successful antibody design depends on both biochemical properties that can be calculated directly from amino acid sequences, such as molecular weight, charge, and hydrophobicity, and complex functional properties, such as binding affinity, expression yield, and stability, that often require experimental measurement.

Understanding how Omnii-1 represents these properties is important for generating better antibody candidates and predicting their behavior before committing to costly experiments. We therefore evaluate the biological properties learned during pretraining and the utility of these representations for antibody design and property prediction.

We investigated whether Omnii-1 learns representations of antibody sequence chemistry during pretraining, despite receiving no explicit supervision on these properties. We focused on five properties that can be calculated directly from sequence: charge, hydrophobicity, aliphatic index, aromatic content, and molecular weight.

All five properties can be calculated directly from amino acid composition, information that can be recovered even from untrained transformer representations because random projections preserve aspects of the input signal. To distinguish learned biochemical information from this baseline, we evaluate how well representations from an early layer of Omnii-1 linearly predict each real property relative to 500 fictitious control properties. These controls are generated by randomly reassigning property values across amino acids, preserving their mathematical structure while removing their biochemical meaning.

The results suggest that training makes Omnii-1's representations more selective for biochemically meaningful information. Charge and hydrophobicity outperform all 500 shuffled controls, while aromatic content and aliphatic index show smaller but significant effects. This suggests that Omnii-1 preserves information about biologically relevant properties while becoming less sensitive to arbitrary patterns in sequence composition.

We evaluate whether Omnii-1's sequence likelihoods correlate with more complex antibody properties that cannot be directly calculated from sequence. Using 327 experimentally characterized Aptamino designs against the same target, we examine the relationships between Omnii-1's assigned protein sequence likelihoods and two experimentally measured traits, expression yield and binding potency.^ff Binding potency is defined using a partner-specific combination of IC50 and Kd measurements. Sequence scores are calculated from token log-probabilities, and normalized across residue positions to reduce biases associated with sequence length.

Table 1. Correlation between Omnii-1 likelihood scores and experimentally measured antibody properties.

Reported values are Spearman correlations with 95% cluster-bootstrap confidence intervals. The evaluation includes 327 designs grouped into 172 sequence clusters. Cluster-level resampling accounts for the substantial sequence similarity within the dataset; the median design shares 94% identity with its nearest neighbor.

Omnii-1 assigns higher likelihood to the designs that express and bind better, while the random scorer is indistinguishable from zero on both traits.

Sequence likelihood is only one way to evaluate what a language model knows, as a model may encode information relevant to antibody function within its hidden representations even when that information is not directly reflected in sequence probabilities. To investigate this, we train supervised ridge regression probes on the learned representations of Omnii-1 and two established protein language models: ESM-C 600M⁴ and AbLang-2⁵, each adapted using continued pretraining.

Table 2. Correlation between representation-based predictions and experimentally measured antibody properties.

Spearman correlations from out-of-fold ridge predictions, with 95% cluster-bootstrap confidence intervals. Predictions are ranked within evaluation folds before pooling. The one-hot baseline replaces model embeddings with a one-hot encoding of the sequence, and the random baseline replaces them with Gaussian noise; both preserve the evaluation procedure.

Omnii-1 achieves the highest point estimates on both expression and binding potency.^gg Omnii-1 and the ESM-C models were trained on mixtures including general protein and Aptamino sequences, while AbLang-2 used Aptamino-specific adaptation data. This was chosen to give the baselines the most competitive performance, and because AbLang-2 is trained primarily on antibody data, a mixture which pushed it towards general protein risked degradation of antibody-specific priors.

The results suggest that Omnii-1's learned representations are competitive with those of protein and antibody-specific language models for predicting experimentally relevant properties.

These findings provide encouraging evidence that representations learned by a general-purpose biological foundation model can be adapted to specialized antibody design tasks.

Antibody optimization is an example of a broader class of biological problems where successful design requires balancing multiple functional properties, experimental measurements are expensive, and computational predictions are imperfect. Similar challenges arise in enzyme engineering, where catalytic activity must be balanced with stability and expression, and in protein engineering, where binding specificity, solubility, and manufacturability may compete.

Our work with Aptamino demonstrates how Omnii-1, a broadly pretrained multimodal genome language model, can be adapted to a partner's specific objectives, sequence data, and experimental results. These initial results illustrate its potential to support iterative, laboratory-guided optimization rather than serving only as an offline prediction model.

Our longer-term goal is to extend this capability across biological domains, allowing partners to adapt Omnii-1 to their own experimental systems, proprietary datasets, and design objectives.

Partners can bring antibody candidates, targets, screening data, and experimental validation capabilities. We aim to combine Omnii-1's generative capabilities with iterative laboratory feedback to optimize antibody properties and accelerate discovery.

Or write to us at health@radicalnumerics.ai.

We're also hiring researchers and engineers interested in building biological foundation models and connecting AI systems with real-world scientific experimentation.