AnomalyModifier: Suppressor Modifier Discovery in Familial Hypercholesterolemia via One-Class Anomaly Detection

Inpyo Hong, Joohyun Han, Min Kang, Doyeon Ha

Keywords: One-Class Anomaly Detection, Deep SVDD, DNA Language Model, Suppressor Modifier Discovery, Drug Target Identification, Rare Monogenic Disease, Familial Hypercholesterolemia

TL;DR: We reformulate suppressor modifier discovery as one-class anomaly detection over patient variant pairs, retrieving all four FDA-approved familial hypercholesterolemia drug targets (PCSK9 rank 2 of 19,292) without any suppressor labels.

Abstract: Target identification is the principal bottleneck in developing therapies for rare monogenic diseases. Unaffected carriers, who harbor a known causal variant yet do not develop disease, offer a direct route: their resilience has been partly attributed to suppressor modifier variants that attenuate the causal effect, and pinpointing such modifiers directly nominates new therapeutic targets. We propose AnomalyModifier, a one-class anomaly detection framework that treats co-occurring (causal, non-suppressor) variant pairs from patient samples as in-distribution and flags out-of-distribution query pairs as suppressor modifier candidates. AnomalyModifier is trained on patient variant pairs with a hypersphere-regularized autoencoder (AE) objective; the hypersphere center and radius, inspired by Deep SVDD, are updated by deterministic rules decoupled from the encoder gradient. From a 1,121,951-pair deduplicated patient corpus (897,559 used for training) for familial hypercholesterolemia (FH) type 1 (OMIM 143890, LDLR causal gene), the model retrieves approved drug targets near the top: PCSK9 at rank 2, ANGPTL3 at rank 58, MTTP at rank 384, and APOB at rank 101 (4-gene mean rank 136, gene-level AUROC = 0.993) on a synthetic loss-of-function (LoF) benchmark over 19,292 protein-coding genes. On 3,130 ClinVar-curated clinically validated variants, the model achieves variant-level AUROC = 0.930 (Cohen's d= +1.94).

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 61