SynthyraBiosecurity screening
Find the sequences of concern that current screening misses.
DNA synthesis screening today compares an order against lists of known dangerous sequences. A sequence that is new or redesigned has no close match on a list, so it passes without a second look. Atlas reads what each sequence is and what it does, so those orders reach a reviewer, with the reason.
Illustrative. The sequences are real, from the Screener's public example set. The answers beside them are the kind the screen gives for each kind of input, and they are not the output of a live run. No detection performance is claimed.
Unobserved Share Index
v0.1.0 · 2026-10-09
0%
The share of synthesis orders that reach manufacturing without a person ever looking at them. The screen found no list match, so nobody was alerted.
Baseline band 90–98%. Most of these orders are harmless. The index measures how much goes unseen, and every input is cited in the methodology below.
The gap
Current screening only finds what is already on a list.
Screening of synthetic DNA orders is built on matching sequences against databases of regulated agents and toxins. The US framework for nucleic acid synthesis screening defines sequences of concern by their best match to regulated agents, and the industry consortium's harmonized protocol screens on the same principle.
Since 13 October 2026 the framework widens that definition to sequences known to contribute to pathogenicity or toxicity, even when they are not derived from a regulated agent. A screen built on matching has nothing to say about a sequence that matches nothing.
A sequence with no list match is never shown to a reviewer.
That limit is documented. A 2025 study in Science found that proteins of concern reformulated with generative design tools could pass screening software, and that patches developed with synthesis providers substantially closed the gap for the variants tested. Divergent natural homologs and proteins designed from scratch raise the same question again.
01
Orders with no list match
Screening compares an order against databases of known sequences of concern. In one published validation set, 96% of orders returned zero matches: the screen has nothing to say about them.
02
The share no person ever sees
About 5% of orders are flagged and routed to a reviewer. The rest clear without a person looking. The index is that unobserved share.
03
New and redesigned sequences
Detection drops to near zero below about 30% sequence identity. A redesigned protein or a distant natural relative has no close match, so a list cannot see it. Design tools make these sequences easier to produce every year.
The funnel
Where orders go, and where they do not.
Unobserved share
90–98%
Unobserved share 90–98%.Each dot is a synthesis order. Orders that match a screening list are sent to a human reviewer. Orders with no match go straight to manufacturing, and today nobody reads them. A list-based screen has nothing to say about a sequence that matches nothing on it. Atlas reads the sequence itself and predicts what the protein it encodes does, so a sequence of concern reaches a reviewer whether or not it is on a list.
Adding Atlas behind the existing screen sends part of the unseen stream to a reviewer. How large a part is still an assumption band, so the result is shown as a range. The methodology explains why.
How it works
How the screen reads a sequence.
Atlas reads a sequence once and asks four questions of it, listed below. Each answer comes back as a flag with its reason, in one API call. The reviewer sees the flags and the evidence behind them.
A second screen that runs beside the one you already have.
01
Relatives by sequence, structure and function
A search for the closest known relatives by sequence, by predicted structure and by function. The reviewer sees the alignment itself, with its score.
02
Taxonomic origin
Which part of the tree of life a sequence appears to come from. This helps a reviewer put a flag in context.
03
Natural or designed
Whether a sequence looks like something evolution produced or something a person designed.
04
Hazard-related function
Whether a sequence carries the signature of a function that contributes to pathogenicity or toxicity, including when it sits far from any known example.
The answer
Every sequence comes back safe or review.
Every screen ends in one of two words. Safe means none of the signals fired. Review means at least one did, and the level says how many and which, so a provider knows how hard to look.
We tune the screen to flag everything it can detect. A false alarm costs a reviewer a few minutes. A miss defeats the purpose of screening. When an order comes back for review, we recommend a conversation with the customer about what they are making.
01
No close relative
The sequence resembles nothing known by sequence, structure or function. The screen cannot say what it does, so a reviewer should look.
02
Natural-looking, with a flag
It resembles natural proteins, but it also triggered the designed signal or the hazard signal.
03
Relative of a known hazard
It is similar in sequence, structure or function to a protein of concern.
04
Designed, uncertain origin, hazard similarity
The designed signal, an unclear taxonomic origin, and similarity to hazardous material, together. The highest level.
Assess your own exposure
How much of your own order flow goes unseen.
Enter four numbers you already know. Nothing leaves your browser until you ask for an evaluation, and we never ask for a sequence.
5.0%
Approximate by design. Pooled oligo orders are not counted separately; if they are a large share of your volume, say so when you ask for an evaluation.
Your unobserved share
90–98%
Share of your orders that would reach manufacturing with no human having looked. A band, not a point: the inputs it is built from are themselves ranges.
- Percentile vs. baseline
- 98
- With a second stage
- 36–88%
The second-stage figure applies a wide assumption band, preliminary, internal evaluation, not yet third-party validated. It is not a measured Synthyra result, and no point inside the range is claimed.
Your unobserved share is 90–98%, percentile 98 against the baseline; with a second stage, 36–88%.Exposure in dollars & orders
Flags to review / month
500
Reviewer-hours / month
500–2,000
Loaded review cost / month
$30,989–$173,365
Orders reaching manufacturing unobserved / month
9,000–9,800
Reviewer cost = flags × review-hours (1–4 h/flag, Applied Biosafety 2024) × a loaded hourly rate (BLS SOC 19-5011 median $90,150/yr, May 2025, loaded 1.43×–2.0×). Every assumption is shown, and each is listed with its source in the methodology.
Work with us
Run Atlas triage behind your existing screen for two weeks.
A free evaluation on your own flagged backlog. Atlas runs second, behind your current screen, and routes evidence to your reviewer — it issues no passes. We would rather be adjudicated on your data than grade our own homework.
We never ask for, accept, or process a sequence on this page. An evaluation is scoped and run under agreement.
Email us and we will reach out about an evaluation on your own flagged backlog. The values you entered above are included in the message; nothing is sent until you do.
Request a free evaluationOpens your mail client addressed to [email protected]. We do not store any sequence; none is collected here.
Methodology
Every input, its source, and what it does not say.
The index is a rate, not a count. No public source gives a global order volume, a global base-pair total, or a measured screening-coverage percentage. The one authoritative statement on coverage is that it is unknown (EBRC, 2022). So the headline is a share of order volume, reported as a band. Absolute numbers appear only in your own assessment, where the volume is your input.
Definition: unobserved = zero-match share × (1 − second-stage coverage). With no second stage, the index is the zero-match band itself. Most zero-match orders are harmless (over 99% of orders pass). The index measures what is unseen, and says nothing about what is harmful.
Version
v0.1.0 · 2026-10-09. Changelog: v0.1.0, initial publication after a two-track public-source review.
Zero-match share (the tail a list cannot read)
Peer-reviewed"an overwhelming majority (96%) featured zero matches." Zero-match ≠ hazardous; it is the share a homology screen cannot read.
Uncertainty: Point 96% on one validation set; banded 90–98% to reflect a single-set estimate.
SecureDNA, "Random adversarial threshold screening…", National Science Review 13(14) ↗ (accessed 2026-10-08)Orders that pass screening outright
Peer-reviewed"Over 99% of sequences passed screening." The index measures what is unseen, not what is dangerous.
Uncertainty: Stated as "over 99%" on the validation set.
SecureDNA, National Science Review 13(14) ↗ (accessed 2026-10-08)Share of orders flagged for review
Peer-reviewed"Approximately 5% of synthetic DNA orders are flagged and reviewing one flag can take several hours."
Uncertainty: Traces to a single 2015 JCVI self-estimate (Green 95% / Yellow 4.3% / Red 0.7%); not three independent measurements.
Rose, Alexanian, … Diggans, "Practical Questions for Securing Nucleic Acid Synthesis," Applied Biosafety ↗ (accessed 2026-10-08)Hours to review one flag
Peer-reviewed"reviewing one flag can take several hours." Expert adjudication plus customer-legitimacy verification.
Uncertainty: Source says only "several hours"; band is our reading of "several", stated as ours.
Rose, Alexanian, … Diggans, Applied Biosafety ↗ (accessed 2026-10-08)Identity floor where homology detection collapses
Peer-reviewedFlagging "dropped to near 0%" below ~30% identity: a novel sequence with no close relative is not shown to a human.
Uncertainty: Stated as "∼30%".
Wittmann, Wheeler et al., Front. Bioeng. Biotechnol. 14 (doi 10.3389/fbioe.2026.1858951) ↗ (accessed 2026-10-08)What a second stage can recover (a known set, and not Atlas)
PrimaryUsed only to bound what a second stage can do in principle. The page states this is not an Atlas result.
Uncertainty: Per-tool unflagged pre→post: 68→1%, 18→3%, 11→2%; AAAS "about 3% … still escaped." This is homology patches on a KNOWN set, not Atlas on the novel tail.
Wittmann et al., Science 390:82, Table S3 (author-hosted supplementary); AAAS release ↗ (accessed 2026-10-08)Providers with publicly identifiable screening
Peer-reviewedProvider share (~10%) is NOT capacity share. Both are rendered; neither is interchangeable.
Uncertainty: "69 of more than 700" ≈ 9.9%; "publicly identifiable" so a lower bound on screening; denominator open-ended.
Zakaria et al., Front. Bioeng. Biotechnol. 14 (doi 10.3389/fbioe.2026.1819575), citing IBBIS Global DNA Synthesis Map ↗ (accessed 2026-10-08)Reviewer compensation (loaded-cost basis)
Gov. statisticMedian annual $90,150; mean hourly $45.12. SOC 29-9011 is obsolete; 19-5011 is current.
Uncertainty: Best proxy; BLS has no dedicated biosafety/biosecurity occupation (confirmed absent).
BLS OEWS, SOC 19-5011 Occupational Health and Safety Specialists, May 2025 ↗ (accessed 2026-10-08)Loaded-cost multiplier
Gov. statisticReader chooses within the band; both edges are cited.
Uncertainty: No single government-wide mandated multiplier exists; 2.0× is a widespread agency convention.
BLS ECEC Jun 2026 (1.43×, benefits only); 91 FR 48588, 85 FR 42986, 89 FR 17741 (2.0× benefits+overhead convention) ↗ (accessed 2026-10-08)Self-assessment credit: zero-match orders reviewed
AssumptionCredited when zero-match orders are routed to a human reviewer. Affects your own score only, never the published index.
Uncertainty: No public measure exists of how much of a zero-match queue human review examines; 0.2 is a deliberately modest credit.
Synthyra modelling assumption (self-assessment only) (as of 2026-10-09)Self-assessment credit: non-homology screening
AssumptionCredited when screening includes function- or provenance-based methods. Affects your own score only.
Uncertainty: Function- and provenance-based screens vary widely and none publishes tail coverage; 0.1 is a deliberately modest credit.
Synthyra modelling assumption (self-assessment only) (as of 2026-10-09)Atlas second-stage coverage (the one unverified input)
UnverifiedAn assumption applied as a whole band, not a Synthyra claim. The band is wide on purpose; no point inside it is asserted.
Uncertainty: Atlas's effect on the novel tail has not been measured independently; the band is bounded below by caution and above by the not-Atlas patch analogue.
preliminary, internal evaluation, not yet third-party validated, to be replaced with a third-party-adjudicated figure.
Internal — see the derivation below (as of 2026-10-09)Derivation of the second-stage band
We assert no Atlas detection number, and the UI states none. The “with Atlas triage” view applies the whole preliminary band and shows the range that results, so no single value inside it is ever stated. That a second stage can recover much of a tail is anchored to a published analogue that is explicitly not Atlas: Wittmann et al. (Science, 2025) showed homology patches cut the unflagged rate on a known redesign set from a ~26% mean to a ~3% mean. Atlas acts on the novel tail, where its effect has not been measured independently, so the band starts low and stays wide. The counterargument is on the record and we print it: current design tools are “largely incapable of producing the extremely divergent (<30% sequence identity) sequences” that would most stress a list-based screen. When a third-party-adjudicated figure exists, the band is replaced and the Unverified label is dropped.
The two self-assessment credits are modelling choices of the same kind, marked Assumption: they credit a reader who already sends zero-match orders to a person, or whose screen reads function or provenance, with examining part of the tail. They move only your own score, never the published index.
Where this stands
What we claim today.
- It is a beta. No detection performance is claimed on this page until it is published with its evaluation.
- It is a screening aid. Its output is a flag for a human reviewer, and the reviewer decides about the order.
- It is built to run beside homology screening. It does not replace existing screening, regulatory obligations, or review.
The test it must pass is simple to state. On held-out taxa and families, does it flag the divergent variants that similarity search misses, at a stated false-positive rate? If it does not, we narrow it to the categories that work, or we publish the negative result.
These models are built to find concerning sequences, and only to find them. We do not evaluate, publish, or describe ways to get a sequence past a screen. Red-teaming happens under agreement with a partner, on held-out families, and results are reported as capability and false-positive and false-negative rates. Whether to release weights or data is decided one artifact at a time. Our terms prohibit using the platform to design, optimize, or test agents intended to harm human health, agriculture, or the environment.
Next step
Current screening leaves most orders unread. See what Atlas finds in yours.
A pilot runs Atlas beside your existing screen for two weeks, on a set of orders we scope together, under agreement. You see every flag and its reason, and you keep the report. We never ask for a sequence to start the conversation.
A scientist replies within two working days.
Sources
- OSTP, Framework for Nucleic Acid Synthesis Screening (rev. Sep 2024)
- IGSC, Harmonized Screening Protocol v3.0
- Wittmann et al., Strengthening nucleic acid biosecurity screening against generative protein design tools, Science 390 (2025)
Index sources: Wittmann et al., Science 390:82 (2025); Wittmann, Wheeler et al., Front. Bioeng. Biotechnol. (2026); SecureDNA, National Science Review (2026); Rose, Alexanian, Diggans et al., Applied Biosafety (2024); IGSC Harmonized Screening Protocol v3.0 (2024); OSTP Framework (rev. Sep 2024) and HHS/ASPR; BLS OEWS (May 2025). Full quotes, dates, and the uncertainty of every figure are in the methodology above. The amber colour marks list-matched orders sent to a person; the unobserved colour marks orders no person sees: unseen, not dangerous.