Sequence-First Biology: Less Structure, More Speed, Wider Reach
In preprints, LangAAI matches structure-based methods on affinity ranking using only sequences, VHAMSTeR finds 175 cryptic RNA bacteriophages, and pyreml fits a genomic prediction model more than 12 times faster than closed-source state of the art.
- In a preprint, LangAAI conditions exclusively on antibody and antigen primary sequences to model antibody hypervariable regions without 3D structures, and matches structure-based methods and inverse-folding models in affinity ranking benchmarks. Self-reported m:…3973
- In a preprint, VHAMSTeR outperforms existing host prediction methods for DNA viruses with a Matthews correlation coefficient above 0.91 and RNA viruses above 0.69, and identified 175 previously cryptic RNA bacteriophages in global metagenomic catalogues. Self-reported m:…0816
- In a preprint, pyreml fits a genomic best linear unbiased prediction model more than 12 times faster than current state-of-the-art closed source REML software, with outputs validated against lme4, rrBLUP, and BLUPF90 using more than 500 functional tests. Self-reported m:…2534
Antibodies without structures
A preprint on LangAAI removes a requirement that has long constrained antibody modeling. LangAAI conditions exclusively on antibody and antigen primary sequences to model antibody hypervariable regions without 3D structures, according to the preprint, which means it works where no structure is available. m:…3973
The preprint tests whether the antigen conditioning is genuine rather than decorative: LangAAI exhibits genuine antigen conditioning, evidenced by reduced sequence recovery in the third complementarity-determining region when cognate antigens are replaced with decoys. And it matches the performance of structure-based methods and inverse-folding models in affinity ranking benchmarks, according to the preprint, so dropping the structure requirement does not cost ranking accuracy. m:…3973
The decoy test is what separates a real conditioning signal from a model that merely looks antigen-aware. If swapping in a decoy antigen did not change recovery, the model would be generating plausible antibody sequences regardless of the target. Because recovery drops with decoys, the antigen input is doing work rather than decorating the output. m:…3973
Viral hosts across the tree of life
A second preprint widens the reach of sequence-based prediction to viral hosts. VHAMSTeR outperforms existing host prediction methods for DNA viruses, with a Matthews correlation coefficient above 0.91, and RNA viruses, above 0.69, according to the preprint. m:…0816
The method also produces new biology rather than just better scores. VHAMSTeR identified 175 previously cryptic RNA bacteriophages in global metagenomic catalogues, the preprint reports, and commensal protist viruses are depleted in the gut viromes of industrialized human populations compared to rural cohorts, according to the preprint. m:…0816
Finding previously cryptic phages matters because catalogues can only guide research on what they contain. A host-prediction method that surfaces 175 previously hidden RNA bacteriophages expands the known viral landscape itself, not just the accuracy of assignments within it, and the rural-versus-industrialized virome difference shows those catalogues still carry population blind spots. m:…0816
Speed for the everyday fit
A third preprint addresses the unglamorous bottleneck: fitting time. pyreml fits a genomic best linear unbiased prediction model more than 12 times faster than current state-of-the-art closed source REML software, according to the preprint, as a general-purpose, GPU-accelerated mixed model solver in Python. m:…2534
pyreml also implements a latent factor-analytic variance structure, a feature not currently available in other open-source REML solutions, according to the preprint, and its outputs are validated against lme4, rrBLUP, and BLUPF90 using more than 500 functional tests. Speed without that validation would be hard to trust; the test count is what makes the speedup usable. m:…2534
Leaner inputs, faster tools
The three preprints share a direction: do more with less. LangAAI drops the 3D structure requirement, VHAMSTeR extends host prediction across the tree of life from sequence data, and pyreml cuts fitting time by more than 12 times while staying open source. m:…3973 m:…2534
For computational biology teams, the implication is to question input requirements before adding compute. When a sequence-only model matches structure-based ranking and a GPU solver clears the fitting bottleneck, the constraint worth removing may be the workflow's assumptions rather than its hardware. m:…3973
The common thread is that none of these gains required a bigger model or a new data modality. LangAAI works with sequences alone, VHAMSTeR works across the tree of life from existing metagenomic catalogues, and pyreml works as a faster solver for a standard model class. The gains come from removing requirements and bottlenecks, which is usually cheaper and faster than scaling up. m:…3973 m:…2534
What is still unresolved
- Whether LangAAI's affinity ranking holds on antigens outside the tested sets
- Whether VHAMSTeR's host predictions hold on newly sequenced viral families
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.