Platform

Xelari is not a language model. The pipeline combines: 

- Geometry-aware neural networks for structural physics 

- Stochastic and deterministic algorithms for conformational sampling 

- Energy-decomposed scoring with thermodynamic constraints 

- Machine-learning components where they outperform classical methods 

Neural networks are used where they provide a measurable advantage: geometry, ensemble modeling, contact prediction. The platform does not rely on sequence-only language models or chatbot wrappers around third-party tools.

No. The platform is fully de novo. No SELEX rounds are performed, and no SELEX-derived enrichment data are used as training input. This eliminates PCR amplification bias, GC-content artifacts, and the false-negative trap inherent to enrichment-based selection.

Used as a building material rather than as genetic information, nucleic acids offer:

- High chemical and thermal stability 

- Predictable, programmable folding 

- Low synthesis cost 

- Site-specific modification 

- Biocompatibility 

- Reproducibility without batch variation

Sequence-only language models, transformer or GPT-style architectures trained on aptamer sequences, operate on 1D nucleotide strings without explicit 3D geometry. Three concrete biophysical reasons make them unreliable for aptamer-protein binding: 

1. No geometric inductive bias. Binding is dictated by 3D surface complementarity, electrostatics, and steric exclusion. Linear-sequence models do not enforce these constraints.

2. Conformational polymorphism. Aptamers, particularly G-rich DNA and short RNAs, exist as ensembles of distinct folds in thermodynamic equilibrium. A sequence-only model that emits a single deterministic score cannot represent this ensemble. 

3. Sparse productive contacts. Aptamer-protein contacts are sparse. Without geometric supervision, the model cannot reliably separate productive contacts from incidental sequence co-occurrences. 

Xelari operates on structure first. Sequence is the output, not the input.

SELEX-rescue and SELEX-extension models (VAEs, diffusion models, transformers trained on HT-SELEX rounds) improve sequences within an enriched pool. Two structural biases are baked into this approach:

1. PCR amplification bias. SELEX uses PCR between rounds. Sequences with highly stable 3D folds are difficult for polymerases to amplify and are systematically depleted in late-round sequencing data. Models trained on enriched pools therefore learn to prefer easy-to-amplify sequences over high-affinity binders. 

2. False-negative trap. A typical SELEX library covers a tiny fraction of the theoretical sequence space. Sequences absent from the final pool are usually labelled "negative" when they are simply unobserved. Generalization beyond the narrow distribution of the starting library breaks down. 

Xelari does not start from a pool. It builds the binder from the target structure outward, with no SELEX dependency.

AlphaFold succeeded on protein folding because of two things that do not exist for aptamers:

1. Hundreds of thousands of solved protein structures in the PDB, plus deep multiple sequence alignments providing evolutionary constraints. 

2. A monolithic end-to-end architecture suited to that data abundance. 

For aptamer-protein recognition, the data regime is the opposite. The number of solved aptamer-protein complex structures is in the low hundreds, accumulated across three decades of structural biology. The effective number of unique aptamer-protein systems is even smaller, because the data clusters around a few well-studied targets. 

A monolithic end-to-end model trained on this corpus runs into a data ceiling, not a model-capacity ceiling. The right response is decomposition into physically interpretable subproblems, each independently constrained. That is the architectural choice behind Xelari.

A small number of structure-based de novo aptamer design tools exist in academic settings. They are scope-limited: 

- RhoDesign (MIT, 2024). RNA only, small-molecule targets only. Inverse folding via geometric encoder + transformer. 

- AiDTA (preprint, 2025). DNA only, protein targets. Fragment assembly via Monte Carlo Tree Search. 

- MAWS / AptaLoop. DNA or RNA, small-molecule or protein targets. Fragment growth with free-energy minimization. Pre-ML era tools, narrow scope. 

Xelari covers DNA and RNA modalities, protein targets as the primary track plus small molecules and cells, and runs a decomposed pipeline with ensemble physics, thermodynamics, and off-target screening. 

It is the only platform in this subcategory that combines DNA + RNA modality, ensemble-aware design, thermodynamic optimization, multi-objective output, peer-reviewed wet-lab validation on a live pathogen, and a 24-hour design cycle.

Static rigid-body docking treats both partners as fixed objects and scores their geometric fit. This works for some small-molecule cases. It fails for aptamer-protein binding because of induced fit. 

When an aptamer engages a target, both molecules adapt. Loops rearrange, G-quadruplex topologies stabilize, protein side chains rotate, water networks reorganize. A binder selected by rigid docking on free-state structures often turns out, in solution, to prefer a different pose, or no pose at all. 

Xelari incorporates ensemble-level flexibility on the aptamer side and induced-fit awareness on the protein side. The output is a binding mode consistent with realistic complex behavior.

A single scalar "binding score" tells you that the model predicts binding. It does not tell you why. If the score is wrong, there is no diagnostic path. 

Xelari decomposes the predicted binding free energy into physically interpretable components: electrostatic interactions, Van der Waals contacts, water-mediated hydrogen bonds, and (where applicable) cation-mediated stabilization. 

A designer can see which interactions drive predicted affinity, which are weak, and where a sequence modification is most likely to improve specificity or stability. Predicting binding is one thing. Understanding binding is the basis for rational optimization.

1. Project setup. Define the target protein and the desired Xelamer functions in natural language. 

2. Automated structural analysis. The platform parses the request, prepares the target structure, and identifies candidate binding regions. 

3. Design. Multi-stage pipeline generates and scores Xelamer candidates against binding, specificity, stability, and environmental constraints. 

4. Delivery. Final designs are returned with sequence, predicted folded structure, binding mode, and functional annotations.

Yes. A built-in chat interface accepts plain-language project briefs, follow-up refinements, and technical questions. No special syntax required.

- Target protein (UniProt ID or structure file) 

- Desired binding region, if known 

- Required Xelamer function(s) 

- Operating conditions (pH, temperature, ionic environment)

The platform accepts both experimental structures (X-ray, cryo-EM, NMR) and computationally predicted structures (AlphaFold-class models). For targets without a deposited experimental structure, an AlphaFold-predicted model is a fully supported input. Quality flags from the prediction are integrated into downstream design decisions where relevant. 

If the overall model quality is poor but a region of interest is well-resolved, you can scope the design to that high-confidence fragment, leaving the rest of the protein out of the binding-site definition. This is the recommended approach for partially predicted targets.

The platform models environment-dependent behavior across: 

- pH range 

- Ionic composition (Mg²⁺, Na⁺, K⁺). Ionic environment is a primary determinant of nucleic-acid folding. 

- Temperature 

- Buffer-specific stability (serum, vitreous humor, CNS fluid, tumor microenvironment) 

- Co-factor presence 

Optimization happens against the real environment of the intended application, not as a post-hoc filter.

A ready-to-order Xelamer in 24 hours, depending on target complexity. Antibody development takes 6 to 12 months. SELEX takes 1 to 6 months.

Computation runs on Xelari infrastructure with hosting available in both US and EU regions to meet client compliance requirements (data residency, GDPR, sector-specific frameworks). Deployment region is selected at project setup. The web interface works from any standard browser, no specialized hardware required on the client side.

Xelamers are produced under Research Use Only (RUO) terms by default. Within that scope, they are used across: 

- Diagnostics research: assay development for cancer detection, pathogen identification, organ-function markers, allergy testing, aging biomarkers. 

- Therapy research: target validation and lead-finding work on antidotes, antiviral and antipathogenic agents, surgical adjuncts, direct therapeutic candidates, drug delivery vehicles. 

- Basic research: protein function studies, cell signaling, structural biology, novel assay development. 

Translation to diagnostic or therapeutic products requires separate development and regulatory pathways, which Xelari can support under bespoke agreements.

- Detection. Signal the presence and amount of the target. 

- Inhibition. Block the target's function. 

- Activation. Stabilize the target's active state. 

- Linking. Tether the target to a surface or another molecule. 

- Reformation. Confer new properties on the target.

Yes. The platform supports bispecific and higher-multispecific constructs that engage two or more targets, or two distinct epitopes on the same target, within a single aptamer architecture. Design optimizes geometric compatibility, linker length, and individual binding modes in parallel. 

This capability is used for receptor-pair engagement, dual-pathway modulation, and proximity-induced effects. 

Bispecific design is available on request through direct engagement with the Xelari team.

In many research applications, yes. This is especially true where antibody limitations are a bottleneck: cost, batch variation, cold-chain logistics, single function, slow development. The platform can match or exceed antibody binding properties for most protein targets used in research and diagnostics work.

Yes. The design pipeline accounts for non-specific interactions, off-target proteins, nuclease stability, and the operating environment of the intended application.

Yes. Xelamers can be designed with fluorescent labels, enzymatic domains, surface-attachment chemistries, click handles, conjugation linkers, and other modifications standard in oligonucleotide chemistry.

Register, choose a plan, and create your first project. The team is available to assist with onboarding and project scoping.

No. The interface is designed for biologists, not for ML engineers. Natural-language input and guided project setup remove the need for separate training.

Yes. The first peer-reviewed prospective wet-lab validation is published: 

Bruno, J.G., Nasaev, S., Ufaev, D. et al. Evaluation of Artificial Intelligence-Generated DNA Aptamers Against Treponema pallidum Surface Proteins. Journal of Fluorescence, 2026. DOI: https://doi.org/10.1007/s10895-026-04717-4 

Xelamers designed in silico on the Xelari platform were synthesized and validated against live Treponema pallidum by fluorescence microscopy and spectrofluorometry. Multiple in-flight validation programs across oncology, infectious disease, and rare-disease targets are also active under partner agreements.

For specific client projects, additional wet-lab validation can be arranged through partner laboratories.

- Modalities: DNA and RNA aptamers (single-stranded, structured). 

- Targets: primarily proteins (intracellular, extracellular, membrane-bound). Small-molecule and whole-cell targets are supported in development tracks. 

- Target structure source: experimental (X-ray, cryo-EM, NMR) or computationally predicted (AlphaFold-class models).

Xelamer designs are produced under Research Use Only (RUO) terms by default. The client receives the right to use the designs for internal research purposes.

Commercial rights, including downstream development, licensing, royalty structures, and full IP transfer, are discussed and structured case by case as part of the service or partnership agreement. There is no blanket commercial-rights clause; terms are tailored to the project, scope, and intended end use.

- Authenticated, role-restricted access 

- Encryption of sensitive data in transit and at rest 

- Strict project-level access controls 

- Regular security audits - On-request deletion of all client data 

- US and EU hosting regions available for data-residency compliance

Yes. In addition to platform access, Xelari offers bespoke design programs for specialized targets, multi-objective campaigns, bispecific and multispecific constructs, and integrated wet-lab validation.

Xelamers produced on the platform are intended for research use only (RUO) and are not subject to FDA or equivalent regulatory oversight in this form. 

For diagnostic or therapeutic development paths, the client is responsible for obtaining the necessary regulatory approvals. Xelari can provide design dossiers, predicted-property documentation, and supporting characterization data to accompany regulatory submissions.

- New target classes (membrane proteins, intrinsically disordered proteins, multi-protein complexes) 

- New modalities: siRNA, ASO, small-molecule binders 

- Expanded modification space (modified backbones, chimeric chemistries) 

- Improved ensemble and thermodynamic models 

- Continuous incorporation of new wet-lab validation data into the platform feedback loop 

- New functional capabilities for Xelamers

No. Xelari uses DNA and RNA as building materials for binding tools. The platform does not edit, deliver, or interact with the cellular genome. Xelamers are functional molecules that operate outside genetic contexts.

Yes. Xelamers are produced by standard solid-phase oligonucleotide synthesis, a mature commercially available chemical process that scales linearly from milligram research quantities to industrial production. No cell lines, fermenters, or biological production systems are required.