Omnii Antibody

Antibodies are among the most important classes of modern therapeutics¹, but designing an effective antibody remains a difficult problem. A promising candidate must not only bind its intended target, but also satisfy a range of constraints involving stability, expression, specificity, and manufacturability. Optimizing these properties simultaneously requires searching an enormous space of possible protein sequences.

Most antibody design approaches focus on the antibody sequence or structure in isolation. We envision a different approach: designing antibodies with a richer understanding of their biological context, from the molecular environment of their targets to the experimental conditions and functional properties that determine their success. Omnii, our multimodal genome model, offers a path toward this vision through its ability to learn from diverse biological modalities and long sequence contexts.

But can a general-purpose biological model, rather than one built specifically for antibodies, be adapted to solve a practical antibody engineering problem?

In collaboration with Aptamino, we set out to test this idea. We adapted Omnii to single-domain antibody design and developed a lab-in-the-loop workflow that combines sequence generation, functional property prediction, and iterative midtraining on experimental measurements. Omnii-generated designs advanced the measured tradeoff between binding affinity and expression yield.

These results are an early demonstration of a broader capability: turning a general-purpose multimodal genome model into a design system that learns from experiments and adapts to bespoke biological objectives.

An antibody's defining function is to recognize and bind a molecular target. Yet strong binding alone does not make an antibody useful.

A successful therapeutic binder must recognize the correct molecular surface, achieve the desired biological effect, and exhibit favorable properties throughout production and development. These requirements can compete with one another. Changes that improve binding affinity, for example, may compromise expression or stability.

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.

The central challenge is therefore not simply to generate protein sequences that might bind a target. It is to identify sequences that satisfy multiple biological constraints, many of which cannot be reliably determined without experimental measurement.

This creates a natural opportunity for AI systems that learn from experimental feedback.

Our approach combines Omnii'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 then proposes variants by identifying and resampling positions within those sequences.

Functional 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.

The value of this approach comes from the interaction between computation and experiment.

Omnii 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.

Each round follows the same general process: Omnii generates variants of promising experimentally characterized sequences, predictive models rank the candidates, and selected designs undergo laboratory measurement for functional properties including expression and binding affinity.

The campaign began with 1,620 designs generated by earlier pipelines, before Omnii was introduced into the optimization process. Of these, 150 had measured binding affinities for IL23R. Omnii was then used to guide three successive rounds of experimental optimization.

The resulting designs extend the observed Pareto frontier 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 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 therefore tests more than whether Omnii can generate plausible antibody sequences. It demonstrates the feasibility of using a general-purpose multimodal genome model in an iterative experimental workflow, progressively exploring higher-performing regions of sequence space.

More broadly, Omnii's multimodal capabilities open the door to antibody design informed by additional biological context, including protein structure and other molecular inputs. Its ability to support multiple post-training tasks also creates opportunities to combine sequence generation with predictions of binding, stability, and other properties within a single framework.

These are early results from a single partner campaign, rather than evidence of universal improvements across antibody targets. Nevertheless, they provide a promising foundation for extending this approach to more complex antibody design challenges and other biological design problems.

Omnii 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, we use continued midtraining 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 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's broad biological priors with bespoke knowledge relevant to the design problem.

Successful antibody design depends on both readily computable biochemical properties and complex functional behaviors that require experimental measurement.

Understanding how Omnii represents these properties is important for both generating better antibody candidates and evaluating them in silico before committing to costly experiments. We therefore examine what the model learns during pretraining and how effectively its representations support these two roles: design and prediction.

We next asked whether Omnii learns to represent the physical properties of antibody sequences during pretraining, even though it was never explicitly trained to predict them. We focused on five properties that can be calculated directly from sequence: charge, hydrophobicity, aliphatic index, aromatic content, and molecular weight.

Measuring this is less straightforward than it might seem. All five properties depend on amino acid composition, and even an untrained transformer can recover that information surprisingly well. To distinguish learned biochemical information from this baseline ability, we compared Omnii's internal representations before and after training. We also tested each property against 500 controls created by randomly reassigning its values across amino acids. These controls preserve the mathematical structure of the property while removing its biochemical meaning.

The results suggest that training makes Omnii's representations more selective for biochemically meaningful information. Although training generally reduces the ability to recover arbitrary combinations of amino acid counts, charge and hydrophobicity remain just as accessible as before training. Both outperform all 500 shuffled controls, while aromatic content and aliphatic index show smaller but significant effects. This suggests that Omnii preserves information about biologically relevant properties while becoming less sensitive to arbitrary patterns in sequence composition. These findings demonstrate what can be recovered from the model's representations, rather than establishing that the model directly uses these properties when making predictions.

We next evaluate whether protein language model scores correlate with more complex antibody properties that cannot be directly calculated from sequence.

Using 327 experimentally characterized Aptamino designs against the same target, we evaluate relationships between model-assigned sequence likelihoods and two experimentally measured traits: expression and binding potency.^aa 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 model 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 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.

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 and of three established protein language models: ESM-C², AbLang-2³, and IgLM⁴, each adapted using continued pretraining.

For every antibody sequence, we extract the model's final-layer hidden states and average them across residue positions, then fit a ridge regression from that embedding to the experimentally measured properties. We evaluate the probes using five-fold cross-validation with sequence-cluster-aware splits, so closely related sequences never appear across training and evaluation folds. All models use the same 327 designs, regression approach, and evaluation splits, providing a more direct comparison of the information available in their representations.

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

The results suggest that Omnii'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.

The significance of this work extends beyond antibody discovery.

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 necessarily imperfect.

The same fundamental challenges arise across protein engineering, enzyme optimization, and other applications of biological sequence design.

Our work with Aptamino demonstrates an approach to these problems: begin with a broadly pretrained biological model, adapt it to a partner's sequence and experimental data, generate candidate designs, and incorporate laboratory measurements into an iterative optimization process.

The antibody campaign remains at an early stage. Further work is needed to improve generation strategies, predictive accuracy, experimental efficiency, and generalization to additional targets. In particular, separating the contributions of the generative model, functional predictors, and iterative parent selection will be important for understanding how best to improve the system.

Still, the initial results show how Omnii can participate in a real experimental design workflow, rather than serving only as an offline sequence representation or prediction model.

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

Omnii for antibody design is in early access.

We're looking for partners working on antibody discovery, protein engineering, and other biological design challenges where AI-driven sequence generation can be coupled with experimental feedback.

If you have proprietary biological data, experimental systems, or the ability to validate designed molecules, we'd like to explore how Omnii could support your research.

Sign up for early access above, or contact 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.