Skip to content

DeepBD: A Grounded Agentic Workflow for Variant Prioritization and Diagnosis of Genetic Birth Defects

🕒 Published (v1): 2026-06-23 16:37 UTC · Source: Arxiv · link

Why this paper was selected

Grounded agentic workflow for clinical genetics; concrete harness design with domain-expert loop

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

DeepBD is a layered agentic workflow for post-sequencing variant prioritization in genetic birth defects that separates concerns across four interacting layers: LLM-assisted case structuring, a pretrained evidence engine, specialist evidence modules, and a grounded diagnostic agent. Trained on an in-house cohort of 18,622 fetal/infant cases, it achieves Recall@1/3/5/10 of 0.658/0.882/0.912/0.929, outperforming Exomiser, DeepRare, and prompted LLM rerankers. The design principle is evidence allocation: each workflow stage is assigned tasks matching its computational character.

Problem

Post-sequencing interpretation of genetic birth defects is a bottleneck: thousands of candidate variants must be ranked under incomplete, imaging-derived fetal phenotypes against heterogeneous evidence (population genetics, variant-effect predictors, gene-disease validity, phenotype ontologies, pathway/structure context). Standalone prioritizers (Exomiser, LIRICAL), molecular predictors, and general-purpose LLM agents each cover partial signal but lack a principled framework for weighting heterogeneous evidence patient-specifically for this fetal/infant disease space. No existing training substrate captures the specific distribution—fetal phenotypes, sequencing-derived competing variants, clinically curated causal labels—needed to learn a birth-defect ranking model.

Method

DeepBD implements an evidence-allocation principle across four layers:

  1. LLM-assisted case structuring: Maps free-text or HPO clinical phenotypes to computable HPO terms; normalizes VCF/TSV variant tables into per-variant records with provenance preservation. Lightweight by design—structures inputs without making causal decisions.

  2. Pretrained evidence engine (central ranking component): For each candidate variant, constructs three evidence streams:

  3. Rule evidence: allele frequency, consequence class, dosage sensitivity, gene-disease validity, curated pathogenicity (ClinVar/ClinGen/HGMD/OMIM/ACMG-AMP rules).
  4. Variant-intrinsic evidence: genomic foundation-model sequence representations and variant-effect predictors (PrimateAI, SpliceAI, AlphaMissense).
  5. Phenotype-conditioned biological context: links candidate genes to patient HPO profile through cellular/anatomical/pathway context via a biomedical knowledge graph with knowledge-guided attention.

The three streams are fused into a patient-specific variant score via learned fusion trained on the in-house cohort (pretraining on 18,622+ cases, supervised on solved cases).

  1. Specialist evidence modules: Callable, on-demand evidence from SHEPHERD, an independent Exomiser track, and structure-informed refinement (AlphaFold DB retrieval + PyRosetta mutation modeling computing RMSD and \(\Delta\Delta G\)). Reserved for candidates warranting deeper review; not universally applied to every variant.

  2. Grounded diagnostic agent: Operates on the ranked evidence workspace post-scoring. Performs constrained top-\(k\) reranking, reflection-style review, evidence audit, and diagnosis-oriented synthesis. The causal score remains anchored to the evidence engine; the agent does not override it with open-ended reasoning.

Evaluation uses case-level Recall@K (K ∈ {1,3,5,10}) on a held-out split of 549 proband cases with 1,307 curated positive variants.

Key Contributions

  • Grounded agentic workflow design: Formalizes the separation of learnable evidence integration (engine), callable specialist tools (modules), and language-facing synthesis (agent) as distinct roles—avoiding the failure mode of unconstrained LLM reasoning for variant ranking.
  • In-house fetal/infant cohort: 18,622 cases with paired phenotype-sequencing data and clinically curated causal labels, providing the training substrate specific to the birth-defect setting (not publicly releasable due to sensitivity).
  • Pretrained evidence engine: Jointly models rule evidence, variant-intrinsic features, and phenotype-conditioned graph context; learns from the actual competing-variant distribution clinicians face post-sequencing.
  • Complementary case recovery: DeepBD uniquely prioritizes 117/117/106/94 cases at top-1/3/5/10 that both Exomiser and DeepRare miss.
  • Ablation-supported evidence attribution: Demonstrates that rule evidence, mechanism context, and structure refinement each contribute non-redundant signal.

Results

  • DeepBD: Recall@1/3/5/10 = 0.658 / 0.882 / 0.912 / 0.929 (internal held-out benchmark, 549 cases, 1,307 positive variants)
  • Standalone Exomiser: 0.490 / 0.667 / 0.710 / 0.751
  • DeepRare (agentic baseline, reranking Exomiser top-20): 0.489 / 0.667 / 0.710 / 0.751
  • Prompted LLMs (reranking Exomiser top-20): Recall@1 = 0.387–0.444; Recall@10 = 0.532–0.546
  • Ablations (Recall@1 / Recall@10 when component removed):
  • No rule evidence: 0.474 / 0.639
  • No mechanism/graph context: 0.494 / 0.630
  • No structure-informed evidence: 0.521 / 0.741
  • Unique recoveries vs. Exomiser + DeepRare combined: 117 cases at rank 1, 117 at rank 3, 106 at rank 5, 94 at rank 10
  • Gains concentrate beyond rank 1, expanding the actionable shortlist for manual review

Limitations

  • Internal-only evaluation: The benchmark is entirely in-house; no public benchmark validation (e.g., ClinVar held-out, SimulConsult, or PhenoBench), limiting external reproducibility assessment.
  • Non-releasable dataset: The 18,622-case cohort cannot be shared, preventing community replication of the evidence engine.
  • Reranking setting for baselines: LLM and DeepRare baselines operate on Exomiser-derived top-20 candidates, not full VCF inputs—making the comparison a constrained reranking contest rather than an end-to-end head-to-head.
  • Diagnostic agent not quantitatively evaluated: The narrative synthesis output of the grounded diagnostic agent is explicitly excluded from the quantitative benchmark; only variant-level ranking is measured.
  • Single disease domain: Trained and evaluated exclusively on fetal/infant genetic birth defects; generalizability to adult rare disease or broader Mendelian settings is untested.
  • Structure module scope: AlphaFold/PyRosetta refinement is applied selectively (missense-relevant cases), so its contribution to overall Recall is partially masked by its narrow applicability.

Relevance to Harnesses / Meta-Harnesses

DeepBD is a domain-specific meta-harness that formalizes evidence allocation as its core architectural principle: distinct subsystems (trained engine, specialist tool modules, LLM agent) handle distinct computational tasks rather than routing everything through a single agent loop. This is directly relevant to harness design because it demonstrates that hybrid pipelines—where a trainable scoring model anchors ranking and an LLM agent handles synthesis/routing without overriding the score—can outperform pure agentic loops in structured prediction tasks. The layered design (structuring → scoring → specialist modules → diagnostic review) mirrors the phase decomposition common in research meta-harnesses, where different stages have different computational characters and failure modes. For practitioners building harnesses over heterogeneous evidence sources, the ablation framework (per-module Recall@K) provides a methodology for attributing contribution across pipeline components.