Absci

Generative design is the business. Origin-1 designs antibodies against a chosen epitope on a target structure, lead optimization models tune candidates for developability, half life and immunogenicity, and the wet lab exists to test those designs and return data that retrains the models. Every internal program, from ABS-201 to ABS-202, and every partnered program starts on that platform.

Absci sells no software, so what a partner buys is the output of the models and the laboratory working together.

Scientists choose the target and the epitope, Origin-1 and the optimization models propose sequences, and every design is made and tested in Absci's own laboratory before it advances. Candidates then move through preclinical studies and trials under protocol, and designs that fail in the lab return as training data, so the loop itself checks the models.

No published rule sets how many designs are tested, or what result a design needs, before it moves forward.

Origin-1 is the named generative model for de novo antibody design against novel epitopes, presented at RosettaCon in 2026, and the platform pairs it with multiparameter lead optimization models. Training data is characterized: proprietary data from SoluPro cell lines and the ACE assay, which screens millions of antibody variants, combined with public biological datasets, and the mAbs paper describes fine tuning protein language models on 240,000 variants for polyreactivity.

Model versions and release practice are not published.

The models are Absci's own. Origin-1 and the optimization models are trained on data its laboratory generates, with public biological datasets added, and AMD and NVIDIA are named compute partners; nothing is installed in a partner's environment, so the chain from target to designed antibody runs inside one company.

No list of service providers or cloud hosts is published.

ABS-201 is in a named clinical study with a disclosed timeline. In the HEADLINE trial, blinded interim data from all four single ascending dose cohorts in June 2026 showed the antibody well tolerated with an estimated half life of at least 65 days, and interim proof of concept data in pattern hair loss is due in the second half of 2026, with full results in early 2027. ABS-101 reported interim Phase 1 data showing a longer half life than first generation anti TL1A antibodies. On method, a 2026 paper in mAbs by Absci scientists describes fine tuning protein language models to predict antibody polyreactivity.

The de novo design work behind Origin-1 appears in preprints and conference abstracts rather than peer reviewed journals.

Little patient data passes through the platform. Designs come from target structures and laboratory data, and patient information enters only through clinical trials, which carry their own privacy notices given at the point of collection. The website privacy policy describes technical, organizational and physical safeguards and international transfers of personal information.

How partner materials and data are kept, and whether they train Absci's models, is not stated.

The privacy policy covers personal information collected through Absci's websites and its other interactions, and supplemental notices apply in specific settings such as clinical trials. The platform works with laboratory and partner data rather than health records, and trial data is handled under trial protocols and their own notices.

No position is taken on whether Absci acts as a business associate.

Absci's annual report describes a cybersecurity program run by its head of IT under audit committee oversight, with third party risk assessments including annual penetration testing, a managed detection and response service that monitors the network continuously, security questionnaires for vendors, disaster recovery exercises and phishing training for staff.

No SOC 2 report or ISO 27001 certificate is named.

The platform is research software used inside Absci and with partners, never sold as a product, so no device question arises. Its candidates follow the drug pathway: ABS-201 has completed dosing in four single ascending dose cohorts of its Phase 1/2a trial, and the company describes a possible expedited route from Phase 1/2a into Phase 3 that depends on planned discussions with the FDA.

Each program carries its own regulatory status as a drug candidate, while the design models that produced it sit outside regulation.

Performance is reported task by task. The mAbs study showed that a model fine tuned on local data predicted polyreactivity far better within its own sequence space, while its performance on 80 clinical antibodies outside that space was largely unchanged, and Absci's de novo design preprints report binding rates against named targets.

No evaluation of model behavior across target classes or epitope types is published.

A published success rate exists for one of the platform's models. In Saunders et al., mAbs (2026), which Absci lists among its peer reviewed publications, protein language models fine tuned on 240,000 variants were asked to infer 18 low or non polyreactive antibody variants, and the fine tuned model's inferences succeeded 66.6 percent of the time against 0 percent for the base model. Every design is tested in Absci's laboratory before it advances.

No comparable figure is published for Origin-1's de novo designs.

The platform works on protein structures and laboratory measurements, so no clinical record system sits in its workflow, and partners exchange targets, designs and assay data with Absci directly. AMD and NVIDIA are named as partners for the compute behind the models.

No integration with a partner's research systems, such as electronic lab notebooks or data platforms, is described.

Absci runs its models and laboratory itself, with the wet lab and headquarters in Vancouver, Washington, AI research in New York and Belgrade and an innovation center in Zug, and partners receive designed candidates and data. The privacy policy says personal information may be transferred to the United States or other countries.

Where models are trained and where partner data is stored is not stated.

Revenue comes two ways: drug creation partnerships with pharmaceutical, biotechnology and technology companies, structured around milestones and royalties, and the internal pipeline, which Absci develops and may partner or out license. As a Nasdaq company it reports revenue ($0.3 million in the second quarter of 2026), spending and cash ($201.1 million at June 30, 2026, expected to fund operations into the second half of 2028), and Eli Lilly's $40 million investment carries no program rights.

Terms for individual partnerships are not disclosed.

Antibodies are the modality, and the programs show where the platform has been applied: pattern hair loss and endometriosis for ABS-201, a confidential immunology indication for ABS-202, inflammatory bowel disease for ABS-101, and partnered work in oncology with AstraZeneca, dermatology with Almirall and companion animal medicine with Invetx. The internal pipeline centers on cytokine and prolactin receptor biology.

Clinical results so far come only from ABS-201 and ABS-101, so reach across other target classes rests on preclinical work.