Design new medicines from a prompt.

BIOS reads the literature, designs the binder, and builds it in the lab.

Featured news
Designing Medicine, One Binder at a TimeAnnouncing a pipeline that designs a therapeutic peptide binder in under an hour for under $7, and the autonomous robotic lab that synthesizes and validates those designs, closing the loop between idea and experiment.

Biology's legibility curve

1665–2026

    The vision: from a design on screen to a vial in your hand.

    1. Design

      Model a target, design a binder, and watch it fold on screen.

    2. Build

      BIOS hardware turns the best designs into real vials, unattended.

    3. Treat

      A validated candidate becomes a therapy. The loop closes.

    Software: Design binders with state-of-the-art AI

    Three AI models generate candidates, then six filters cut a hundred designs down to the ten most likely to work — about 35 minutes and $8 of compute per run.

    The design engine

    Generate → 6 orthogonal filters → shortlist → wet lab

    The core gate follows the Overath architecture: one DL confidence metric plus one orthogonal physicochemical descriptor, with cross-model agreement as the false-positive filter. Filters 1–4 are the fast core (<1 hr, <$7, mostly CPU). MD and ADMET are deeper validation that runs on the survivors.

    Select an engine or filter for detail.

    More useful tools for scientists

    Run the whole pipeline from a single command, or call any tool inline: public databases, image segmentation, target analysis.

    $ bios design --target GLP1R --n 100▍

    100 designs3 engines6 filters~35 min$8 compute

    Retrieve

    Bioinformatic Databases

    Segment

    Segment Anything

    Track

    Track Video

    Analyze

    Target Analysis

    Open Science Index

    Hardware: An autonomous lab that builds and tests

    Hardware closes the loop: a synthesis box mixes the top-ranked peptide vials and a dexterous robotic arm assays them on live cells, feeding every result back into the model.

    A 3d-printer box that mixes peptide vials.

    • 10 candidates = 10 vials
    • first run: test peptides for the arm
    • then: re-synthesises winning peptides for use

    From prompt to vial, one pipeline

    Nine stages, learning from every run.

    • 01Target definition
    • 02Data retrieval
    • 03Annotation
    • 04Binder generation
    • 05Scoring & filtering
    • 06Validation
    • 07Candidate selection
    • 08Wet-lab synthesis
    • 09Feedback loop

    $ bios design "a stable GLP-1R agonist" → ETA ~50 min

    A literature agent that does the reading

    Ask a research question. BIOS plans the work, reads the papers, runs the analysis, and gives you a cited answer, then suggests what to do next. You steer.

    Learn more
    1. 01

      Plan

      BIOS breaks your question into steps and hands each one to the right agent.

    2. 02

      Search

      It searches papers, patents, clinical trials, and biology databases for what matters.

    3. 03

      Analyze

      Analysis agents run the numbers: Python, statistics, and data processing over the evidence.

    4. 04

      Synthesize

      Everything comes back as one cited answer you can question and refine.

    5. 05

      Discover

      BIOS proposes new directions and experiments, so each loop goes deeper.

    Research notes and logs.

    Papers, lab notebooks, features we shipped, and the things that didn't work.

    Designing Medicine, One Binder at a Time

    announcement

    Introducing BIOS

    announcement

    Blog

    Frequently asked questions

    BIOS is an AI platform for drug discovery. From a single prompt it reads the literature, designs candidate protein binders with state-of-the-art models, filters them down to the most promising, and helps take them into the lab.

    Three generative models (RFdiffusion3, BoltzGen, and PXDesign) propose candidates, then six orthogonal filters — structural confidence, cross-model agreement, physics energetics, selectivity, molecular dynamics, and developability — cut roughly a hundred designs down to the ten most likely to work.

    BIOS runs open models including RFdiffusion3, Boltz-2, ESMFold2, mmseqs2, and Foldseek, all catalogued on the BIO Index (index.bio.xyz). It also exposes bioinformatics databases, image segmentation, and target-analysis tools inline.

    A typical run turns about a hundred candidate designs into a shortlist of ten in roughly 35 minutes for about $8 of compute.

    Yes. BIOS plans a research question into steps, searches papers, patents, clinical trials, and biology databases, runs the analysis, and returns one cited answer you can question and refine — then proposes new directions to explore.

    You can start chatting with BIOS at chat.bio.xyz. For partnerships, demos, or research collaborations, reach the team through the contact page.

    Bring medicines
    to patients faster

    Try BIOS