Machine learning for psychiatric biomarkers has a double blur. On the input side, the cheapest scalable blood assay, bulk RNA sequencing, grinds thousands of cell types into one averaged measurement. Rare or shifted cell populations, such as immature natural killer (NK) cells, vanish into that average. On the label side, the target is usually a diagnosis like "major depressive disorder," a subjective category that lumps together patients whose biology differs. We propose to remove both blurs together and to test a simple, falsifiable claim. Resolving blood into cell-type states helps only when the target is also resolved into specific symptoms. Against a diagnosis label, the benefit of cell-level resolution is largely invisible.
Concretely, a large language model reads clinical notes and extracts symptom dimensions, each tied to quoted evidence: cognitive slowing, fatigue, anhedonia, sleep disruption. A deconvolution model, meaning a method that estimates which cell types and states produced a bulk measurement, turns bulk blood expression into cell-state abundances. A small interpretable predictor maps those abundances to the symptom dimensions. The contribution is not another pipeline. It is a factorial experiment showing whether, and by how much, the two blurs interact, together with a test of whether the cell-state layer carries real meaning.
Core Claim
Two findings in the file motivate the claim. Parsing clinical and neurobiological sources of heterogeneity in depression used normative brain modeling on 6,121 UK Biobank patients to show "equifinality": distinct neurobiological subtypes produce identical clinical depression presentations yet differ in cognition (p < 0.001). Altered NK Cell Receptor Profiles and Immune-Inflammatory Markers in Adolescent Major Depressive Disorder found that NK cell depletion (p = 0.005) and a maturation shift (lower CD57, higher HLA-DR) track objective cognitive deficits rather than the diagnosis as a whole. Taken together, these suggest that peripheral cell states map onto symptom dimensions, and that the diagnosis label averages over them.
If that is true, a standard study that deconvolves bulk blood and regresses against diagnosis will see little gain over bulk signatures and wrongly conclude that cell resolution does not matter. We state the claim as an interaction. Let $R(\text{input}, \text{label})$ be held-out predictive performance, where input is either bulk expression summaries or deconvolved cell states, and label is either the diagnosis or the extracted symptom dimensions. The quantity of interest is
$$\Delta = \big[R(\text{cell},\text{sym}) - R(\text{bulk},\text{sym})\big] - \big[R(\text{cell},\text{dx}) - R(\text{bulk},\text{dx})\big].$$
The hypothesis is $\Delta > 0$, with a confidence interval that excludes zero. A null result is informative too. It would say that cell-level deconvolution of blood adds nothing for psychiatric phenotypes at the resolution current reference atlases allow.
Relation to Prior Work
The closest architectural precedent is A Concept-based approach to Voice Disorder Detection. That work uses a concept bottleneck model, a predictor forced to route its decision through human-readable intermediate variables, with concepts that a language model extracted from anamnesis text. There, text supplies the intermediate concepts. Here, text supplies the target, and the intermediate concepts are biological cell states. BRI-MH likewise extracts "cognitive markers" from text but fuses them with behavioral sensing and has no biological layer.
Two deconvolution methods supply our input-side machinery; neither touches clinical text or subjective endpoints. Cross-modal transfer learning for mapping bulk transcriptomes at cellular level (POPPY) aligns bulk and single-cell profiles in a shared space structured by the Cell Ontology. Count Bridges deconvolves integer counts so that the recovered cell profiles sum exactly to the bulk measurement. POPPY also flags a constraint we take seriously: cell states missing from the reference atlas get misassigned to existing categories. Exhausted or immature NK states are exactly the kind likely to be under-represented.
The label-side machinery comes from Self-Verification Improves Few-Shot Clinical Information Extraction. Its pipeline of extraction, omission detection, evidence grounding and pruning gives each extracted symptom a verbatim source span. A reluctant additive model framework (RAITR) supplies the idea that a flexible component must "earn" inclusion under a heavier penalty. We reuse that idea to measure how much signal escapes the cell-state bottleneck.
Research Program
Label side. A language model reads a patient's notes and produces a score for each symptom dimension, a presence/absence/uncertain status, and a quoted evidence span. It runs inside the self-verification loop. The dimensions are fixed in advance from standard rating-scale item content: cognitive complaints, fatigue, anhedonia, sleep, appetite and psychomotor change. Fixing them in advance prevents searching for whichever label happens to fit the biology. Where research-grade PHQ-9 item scores exist, they serve both as a check on the extraction and as a competing label source.
Input side. Bulk blood expression is deconvolved into abundances of immune cell states: cell types at several levels of the Cell Ontology, plus maturation and activation programs within NK and T cells. Deconvolution uses POPPY-style ontology alignment and Count Bridges, each with a peripheral blood mononuclear cell (PBMC) reference built from Onek1k.
Predictor. The main model is a sparse additive predictor from cell-state abundances to each symptom dimension. It has one addition: a residual path from raw bulk expression, penalized more heavily in RAITR's "reluctant" manner. The size of the selected residual measures bottleneck leakage, meaning signal the cell-state layer fails to carry. This number is reported, not hidden.
Baselines
The naive reference point is a predictor that uses only age, sex and a standard bulk inflammatory summary. That summary includes neutrophil-to-lymphocyte ratio, platelet-to-lymphocyte ratio and the systemic immune-inflammation index, the markers the adolescent study found to move in the opposite direction from adult expectations. The reference model predicts both the diagnosis and the symptom dimensions and sets the floor.
The strongest conventional competitor is classical reference-based deconvolution (CIBERSORTx and MuSiC, the methods Count Bridges reports outperforming) followed by LASSO against the diagnosis. This is what a competent biomarker study would do today, and it corresponds to the $R(\text{cell},\text{dx})$ cell. We found no reported number on Lacuna for any of these methods on PBMC deconvolution accuracy or psychiatric prediction, so the comparison must be run fresh. All arms get the same patients, the same nested cross-validation folds, and the same hyperparameter search budget (an equal number of configurations per arm). Beating this baseline means a higher held-out area under the ROC curve (AUROC) for binary symptom dimensions and higher Spearman correlation for graded ones, with the bootstrap interval on the difference excluding zero.
The strongest unconstrained competitor is a gradient-boosted or small transformer model on the full bulk expression matrix, predicting symptom dimensions directly with no bottleneck. It shows the size of the interpretability cost. As a calibration point, the voice concept bottleneck lost about 3.5 accuracy points (91.33% for HuBERT versus 87.76% for the concept model, as reported in A Concept-based approach to Voice Disorder Detection). We will count the bottleneck as acceptable if it stays within a comparable margin of the black box on symptom dimensions.
Three ablations isolate the key ingredients. The label-granularity ablation swaps extracted symptom dimensions for extracted diagnosis codes while holding everything else fixed; this is the core of $\Delta$. The verification ablation removes the omission and grounding passes, to measure how much label noise from unverified extraction erases the interaction. The deconvolution ablation replaces POPPY and Count Bridges with classical deconvolution, to test whether the newer methods matter here at all.
Why not just…?
Why not just deconvolve and regress against diagnosis, as everyone does? Under our hypothesis, this design cannot see the effect. The equifinality result in Parsing clinical and neurobiological sources of heterogeneity predicts that diagnosis labels average away biology that differs between subtypes. The factorial design runs this approach as one of its four cells, so if it performs as well as the symptom-label arm, $\Delta$ will be near zero and the proposal fails on its own terms.
Why not just use PHQ-9 item scores instead of extracting symptoms from notes? This is a real threat. In cohorts with item-level research scales, those scores may be cleaner labels than anything a language model extracts. We therefore run PHQ-9 items as a third label condition wherever available. If they match or beat extracted labels, the conclusion is that label granularity matters and extraction is only a way to scale it to records without scales. That narrower claim is still useful, because routine clinical records usually hold narratives rather than item scores. Extraction also covers dimensions that PHQ-9 captures poorly, such as cognitive complaints, which the NK findings single out.
Why not just measure NK cells with flow cytometry and skip deconvolution? Flow cytometry gives direct measurements but covers a small preselected panel and is rarely stored in biobanks, whereas bulk blood expression is common. We do not argue against flow cytometry. We use it as the external check in the intervention test below. If deconvolved abundances never approach flow-measured ones, deconvolution fails, and the experiment will show it.
Why not just train a black box on bulk expression? It may be more accurate. The unconstrained baseline measures that gap directly, and the reluctant residual path measures how much of the gap comes from biology outside the bottleneck. If the residual carries most of the signal, the cell-state layer is decorative, and we will report it that way.
Proposed Experiments
Controlled simulation. Before touching patient data, we build pseudo-bulk blood samples by resampling Onek1k PBMCs. We plant known causal links from specific cell states, such as a CD57-low NK fraction, to simulated symptom dimensions, then define "diagnosis" as a many-to-one union of symptom profiles. Varying the effect size, the number of subtypes merged into one diagnosis, and the label noise (by corrupting symptom labels at rates matched to measured extraction error) gives a map of when $\Delta$ is detectable at realistic cohort sizes. Holding a planted cell state out of the reference atlas reproduces the atlas-dependency failure that POPPY describes and measures its effect on bottleneck leakage. The simulation also supplies a power analysis for the clinical cohort.
Validating each component. Extraction is validated on the benchmarks in the file before being trusted on psychiatric notes. On MIMIC-IV ICD-10 extraction, self-verification reportedly lifts F1 from 0.487 to 0.533, and on EBM-NLP trial-arm extraction from 0.419 to 0.530, with 93% overlap with expert evidence spans (Self-Verification Improves Few-Shot Clinical Information Extraction). We treat those numbers as the bar our reimplementation must reproduce. We then annotate a few hundred psychiatric note excerpts for symptom dimensions and report extraction F1 and span overlap on that set.
Clinical cohort. The decisive test needs patients with bulk blood RNA-seq, clinical notes, and ideally cognitive testing and flow cytometry in a subset. No such public cohort appears in the file. The adolescent MDD cohort has flow cytometry and MATRICS cognitive scores but no expression data or notes, and UK Biobank as used in the depression-heterogeneity paper has structured tables rather than notes. We therefore plan an institutional biobank collection sized by the simulation's power estimate. Before unblinding, we pre-register one hypothesis taken from the adolescent NK study: deconvolved NK maturation state predicts the extracted cognitive-complaint dimension better than it predicts diagnosis.
Intervention check on meaning. Concept bottleneck models allow an unusual test. At prediction time, the deconvolved concept value is replaced with the directly measured one, as in the voice paper's roughly 90.5% "perfect concept" upper bound. In the flow-cytometry subset, we replace deconvolved NK and T cell abundances with measured values. If the bottleneck is meaningful, symptom prediction should improve or hold steady. If performance collapses, the model was exploiting deconvolution artifacts, and that result on its own would argue against using the bottleneck as an explanation.
Clinician audit. Finally, a small audit extends the verification behavior described in AICare. Psychiatrists are shown cases with the evidence spans behind each symptom label and the cell-state contributions behind each prediction. They flag explanations they judge biologically or clinically implausible. We record agreement between their flags and cases where the intervention check reveals artifacts. This checks whether the explanation layer lets experts catch the model's errors, which is the only reason to pay for interpretability.
References
- Parsing clinical and neurobiological sources of heterogeneity in depression
- Altered NK Cell Receptor Profiles and Immune-Inflammatory Markers in Adolescent Major Depressive Disorder: Associations with Cognitive Impairment
- A Concept-based approach to Voice Disorder Detection
- BRI-MH: Behavioral Risk Index for Mental Health - An Interpretable Multimodal LLM-Augmented Framework (Student Abstract)
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Self-verification improves few-shot clinical information extraction
- A reluctant additive model framework for interpretable nonlinear individualized treatment rules
- Augmenting Clinical Decision-Making with an Interactive and Interpretable AI Copilot: A Real-World User Study with Clinicians in Nephrology and Obstetrics