PULSE infers proteins and metabolites from routine bloods, then predicts six diseases at 0.72-0.83

The training signal PULSE declines to use is the part worth reading twice. Instead of learning representations from diagnosis codes or case-control status, it reconstructs measurements it has hidden from itself, and the authors argue that label-trained features inherit the bias of those labels and cut a continuous physiology into two bins ^[10]. The temporal machinery does the rest: an earlier visit with paired modalities becomes a personal baseline, so a later unpaired panel is read as a deviation from that individual rather than from a population average ^[1][8].

That mechanism also defines who the method can serve. The advantage comes from historical paired measurements, so the patient must have been deeply profiled at least once already ^[16]. The framework is not a way to conjure proteomics for someone who has never had any; it is a way to stop paying for the second, third and fourth assay once the first exists.

On the headline result, the band is eleven points wide ^[13], and the abstract names neither the six diseases nor their individual AUCs ^[14]. An AUC of 0.72 sorts a population into rough risk order; it does not settle a case. "Comparable to ground-truth proteomic data" ^[5] is also a relative statement about how little the imputation lost, not a statement that the assay was strongly predictive to begin with. The metabolomic comparison is the harder number: measured against 251 real biomarkers, the generated profiles beat every benchmark method tested ^[4].

The awkwardness sits in the validation. Demonstrating that generated profiles match assayed ones requires a cohort where both were assayed, which is UK Biobank ^[3][4][5]. The paper's own motivation is that repeated deep profiling is too costly and operationally heavy for routine care, leaving most health systems with fragmentary laboratory panels ^[2]. So the settings with the most to gain are precisely the ones that cannot check the output against anything ^[15]. The accompanying News and Views in Nature Computational Science frames the contribution the same way, as filling in what was never directly measured using the patient's own history ^[12].

The claim against borrowed architectures is narrower than it sounds and better for it. MultiVI, scVAEIT and MIDAS treat modalities as views of a shared latent state, which works when each cell is measured once ^[7]; clinical records break that assumption because the same person returns, putting information both across modalities within a visit and across visits in time ^[8]. The alternative already in use, sequence models over one modality with concatenation or late fusion for the rest ^[9], throws away the same structure from the other direction. PULSE also took in retinal images and electronic health records alongside blood markers ^[6], which is what makes this an architecture argument rather than a blood-chemistry trick, though the abstract does not say what the images contributed.