F43. NEW MODEL SYSTEMS FOR PSYCHIATRIC GENETICS: XENOGRAFTING POLYGENIC RISK INTO LARGE LANGUAGE MODELS (AND THEN INTERVIEWING THEM)

Jacob Michaelson

doi:10.1016/j.euroneuro.2026.113438

Background No model systems exist to study polygenic psychiatric traits under experimentally controlled conditions. Large language models (LLMs) could fill this gap, offering an interactive platform for heritable phenotypes that resist traditional characterization. But LLM generative processes lack biological grounding, limiting their value as psychiatric models. We steer LLM output along directions extracted from human polygenic data, test the approach on emotion reading from facial photographs, and report initial findings from psychiatric interviews under polygenic steering, observing consistent increases in psychopathology signs under mental illness vectors. Methods We used picture narration (n=4 prompts), COWAT word generation (n=5 letters), and the Achenbach Adult Self-Report (ASR) in N=33,691 SPARK adults, of whom N=19,924 (3,015 with autism, 16,909 without) had polygenic scores. Trial activations and ASR item embeddings were extracted from layer-11 residuals of Gemma 4 (26B-A4B; Google DeepMind). A decoder network reconstructed these task-specific embeddings from participants' polygenic scores, age, sex, and autism diagnosis. Each input contributes one row in the shared 2,816-dimensional hidden layer; those rows are the canonical phenotype-specific steering vectors. Steering was applied by adding the chosen vector (e.g., P-factor, autism diagnosis) to live layer-11 activations during generation, with a coefficient α controlling magnitude. Results Unsteered, Gemma classified seven naturalistic facial emotions at 37.5% accuracy (168 images of diverse human faces captured while viewing emotionally evocative videos, null accuracy = 14.3%). Sex-based steering left accuracy unchanged. Age-based steering produced a modest, uniform improvement of 1 to 2 percentage points. Autism-diagnosis steering at high magnitude dramatically degraded emotion recognition to 17.9%, a 19.6-point drop, while leaving the generated text fully coherent (suggesting a specific perceptual distortion). Polygenic-score steering for educational attainment and the P-factor (general liability to mental illness) each produced gains of up to ∼4 percentage points, accompanied by sensitivity to a broader range of emotional presentations. Under SCID-style psychiatric interviews and stressful real-world scenarios (e.g., breakup, loss of a loved one, work presentation, etc.), polygenically steered models reliably show increased signs of psychopathology, as judged by blinded third party LLMs, when steered with human-derived mental illness vectors. Discussion This pipeline (residual extraction during natural-language tasks, decoder-network derivation of consensus polygenic directions, controlled steering at inference) creates a tractable model system in which the link from heritable trait to behavior can be studied in a biologically grounded way. The autism-driven emotion-recognition collapse, without loss of linguistic coherence, supports a perceptual contribution to social-cognitive differences in autism. Gains under educational-attainment and P-factor steering point to polygenic axes that sharpen affective discrimination, an unexpected direction worth further investigation. The framework opens a route to redesigning psychiatric instruments: items most differentiable under polygenic steering identify the most diagnostically informative probes, and scenario-based interviews under steered conditions reveal patterns that conventional self-report might miss.

F43. NEW MODEL SYSTEMS FOR PSYCHIATRIC GENETICS: XENOGRAFTING POLYGENIC RISK INTO LARGE LANGUAGE MODELS (AND THEN INTERVIEWING THEM)

Files and links (1)

Abstract

Details

Metrics