Life sciences / HERALD / PARALLAX
HERALD: connect measured antigenic drift to a conditional risk estimate
HERALD began as a theoretical surveillance framework linking an antigenically meaningful representation to a drift statistic and a conditional dominance-risk bound. Later project descriptions discuss retrospective replays; those are separate records.
Research settingIndependent
TL;DR
In brief
The question
Can a geometric representation of viral sequences signal an immune-escape variant early enough, and specifically enough, to be useful for public health?
What the work does
The HERALD preprint constructs a Riemannian pullback metric over a protein-language-model representation, regularizes it, fuses drift signals through a probability-integral transform, and states conditional guarantees plus evaluation and governance protocols.
What is still open
The guarantees are conditional and the manuscript sets out protocols rather than outcomes; prospective lead time, false-alert rates and calibration are not established.
Available recordConceptual manuscript plus later project summaries; the theoretical paper itself reports no surveillance performance.
From the research record
Inspect the work
Original figures, recorded values, and explicit model details. Each exhibit identifies its source and what the comparison means.
What each stage of the risk argument requires
Scroll horizontally to inspect the full table →
| Stage | Mathematical object / condition | Scope |
|---|---|---|
| Representation | (1 − ε) dg ≤ δ ≤ (1 + ε) dg | Only the stated bounded-distortion path family |
| Separation | Distance gap Δ | Depends on the contrastive objective and its assumptions |
| Tail bound | Finite class-conditional second moments | Cantelli bound, not a Gaussian-noise assumption |
| Composition | (1 − ε₁)(1 − α)(1 − ξ) | Product form requires the stated conditional independence |
| Fallback | 1 − (ε₁ + α + ξ) | Conservative union bound when independence is unavailable |
These are ingredients of a conditional argument, not fitted surveillance probabilities. Negative lower bounds provide no positive guarantee. Assay linkage, calibration, and abstention remain necessary.
herald.tex · Bounded-distortion assumption, Cantelli proposition, compositional error budget
The research in context
Public-health groups watch viral genome sequences as they are deposited, hoping to notice a variant that evades existing immunity before it spreads widely. The hard part is not spotting change - sequences change constantly - but deciding which changes matter. Most mutations are functionally silent, counting mutations does not say whether antibodies still bind, and the data arrive late and unevenly: a specimen collected one week may not be uploaded for several. Any early-warning claim therefore has to keep three different things apart: sequence distance, antigenic change, and epidemiological significance such as whether a lineage comes to dominate.
HERALD, deposited as a preprint in November 2025, proposes a geometric reader for that problem. It starts from a protein language model's representation of viral sequences and pulls a Riemannian metric back onto it, so that distance in the resulting space is intended to measure functional rather than lexical change; a contrastive objective with Jacobian and Laplacian regularization shapes that metric. Drift signals are then combined through a probability-integral-transform fusion scheme intended for real-time use. Around the construction the manuscript places formal guarantees, an evaluation protocol spanning SARS-CoV-2, influenza and HIV, and an ethics and governance framework. The deposit carries a non-commercial licence and a provisional-patent notice.
What the record defines is a metric and a fusion rule, not a demonstrated detector. Each link in the chain - that the regularized representation keeps distortion bounded relative to functional distance, that a contrastive margin corresponds to separation, that separation maps to escape risk - is an assumption carried into the next, so an end-to-end risk statement holds only if all of them do. The manuscript's own contribution includes the protocol for testing it: what would have to be frozen before an evaluation window, and what would have to be audited, including when each sequence actually became available, before any lead-time figure could mean something prospective.
The chain of claims in the theoretical paper
The reviewed manuscript proposes a pullback metric, regularization of the learned representation, and bounded distortion between latent displacement and functional distance. It then combines drift signals through a probability-integral-transform scheme and connects a contrastive margin to separation.
A further step uses a finite-second-moment probability bound to relate separation to immune-escape risk. The end-to-end dominance bound is conditional on all of these links, including calibration and coverage. Detecting a sequence change and predicting epidemiological dominance are not the same event.
A replay needs its own clock
Later summaries describe an encoder frozen on escape data and a weekly sequence replay. An early-looking alert in such a reconstruction depends on when sequences and labels were available to the system, not only on the specimens’ collection dates.
The original conceptual manuscript explicitly reports no retrospective or prospective performance, latency, or calibration. Its DOI should not be used as if it validates a later “days before naming” headline. A retrospective alert also needs false-alert and missed-event accounting.
The next research step
The next representative record should identify the exact replay version, freeze the encoder and thresholds, and retain submission-date availability. That would make the timing claim inspectable without turning hindsight into a prospective result.
Background mathematics and definitions
Attach collection, submission, and availability times to each sequence. Build each historical analysis window from information available at that date. Keep divergence detection, antigenic interpretation, and public-health significance as distinct endpoints.
- Availability time
- The earliest time the evaluation assumes the system could access a record.
- Lead time
- Difference between a precisely defined reference event and the first qualifying alert.
- False alert
- An alert not meeting the predeclared outcome criterion.
Illustrative example
A specimen collected on day 1 but uploaded on day 8 cannot support a day-3 alert in an availability-dated replay. Moving it backward to its collection date creates seven days of information the live system did not possess.
This background example is separate from the source records and the main technical figure.
Source record
Additional source versions reviewed
-
mirador/theory/herald.texLocal project manuscript or report · not republished here -
parallax/README.mdLocal project manuscript or report · not republished here
No new experiments are reported in this research note. Figure captions distinguish saved research records, source-model calculations, and explanatory diagrams.