GenePheno: Interpretable Gene Knockout-Induced Phenotype Abnormality Prediction from Gene Sequences

AI summary

Overview

Research area: Machine learning for genomics — specifically, deep learning that links DNA/gene sequence to organism-level phenotypic outcomes (computational biology, functional genomics, protein/gene function prediction).

Technical level: Advanced. The paper combines large-scale sequence encoders, contrastive multi-label learning, cross-attention modality fusion, ontology-structured supervision, and formal generalization/conflict-probability analysis.

Scope in one sentence: The paper introduces GenePheno, the first interpretable end-to-end framework that predicts multi-label knockout-induced phenotype abnormalities directly from gene sequences, supported by four curated benchmark datasets.

What This Paper Is About

Predicting which phenotypic abnormalities appear when a gene is knocked out is normally done by feeding models curated information such as protein–protein interaction networks or Gene Ontology annotations, which only exist for well-studied genes. Variant-effect methods instead take sequences as input but only estimate how specific variants change a limited set of phenotypes. GenePheno targets the gap between these two lines of work: predicting the presence or absence of many phenotype abnormalities under gene knockout using only canonical gene sequences, while also exposing the functional mechanisms behind the prediction.

Key Contributions

1. Task formulation: The authors state they are the first to formally formulate the deep learning task of predicting gene knockout-induced phenotype abnormality directly from gene sequences.
2. Four curated benchmark datasets: Four datasets are curated with canonical gene sequences as input and multi-label phenotypic abnormalities induced by gene knockouts as targets, using a stratified split — MPO, HPO, GWAS, and CAFA2 (wPPI).
3. GenePheno framework and objective: An interpretable end-to-end framework mapping gene sequences to multi-label phenotype abnormalities, with a contrastive multi-label objective that captures inter-phenotype correlations plus an exclusivity regularization that enforces biological logic consistency (exclusive pairs identified through an LLM pipeline).
4. Interpretable functional bottleneck: A gene function bottleneck layer using coarse-grained GO categories (depth d = 2) as supervision, giving human-interpretable concepts; case studies show the resulting interpretations align with biological understanding of phenotypes. Comprehensive experiments report state-of-the-art gene-centric F_max and phenotype-centric AUC, with ablations validating each component.

Main Findings

• Best overall performance on all four datasets (Table 1, "All" setting): GenePheno reaches 31.14 F_max / 67.86 AUC on MPO, 43.17 / 71.44 on HPO, 40.54 / 56.34 on GWAS, and 37.94 / 62.53 on CAFA2 (wPPI).
• Margins over the second-best methods: On MPO the paper reports an improvement of +1.48 in F_max and +7.97 in AUC over the second-best method; on HPO the margins are +2.72 in F_max and +13.27 in AUC.
• Smaller dataset behavior: On GWAS, GenePheno improves over the best baseline result of 37.83 (F_max) / 53.53 (AUC) to 40.54 / 56.34, and the paper reports a 0.8–2.4-point advantage in the high-frequency (≥ 301) subsets.
• Gains attributed to the contrastive objective: The authors attribute the large improvements on larger-scale datasets (MPO and HPO) to the contrastive learning objective capturing inter-label relationships in high-dimensional phenotype spaces.
• Ablation results (Table 2): Removing the contrastive loss causes the largest drop in performance; removing GO inputs notably reduces AUC; the mutual exclusivity and bottleneck losses consistently improve results; removing the sequence input also leads to a performance decline.
• Biological plausibility of interpreted links (case studies): In MPO, "maintenance of location" (GO:0051235) is linked to "abnormal aorta tunica media morphology" (MP:0009873), and "isomerase activity" (GO:0016853) to "increased energy expenditure" (MP:0004889). In HPO, "molecular function inhibitor" (GO:0140678) is associated with "decreased circulating complement C4 concentration" (HP:0045042) and "postsynapse" (GO:0098794) with "axial muscle weakness" (HP:0003327). In GWAS, "growth cone membrane" (GO:0032580) is associated with "compulsive behaviors" (HP:0000722) and "glioblastoma multiforme" (HP:0012174), and "cytokine production" (GO:0001816) with "inflammation of the large intestine" (HP:0002037).
• Theoretical results: Proposition 1 (stationary analysis of the exclusive regularization) states that every first-order stationary point satisfies s_i(x) ≤ 0 or s_j(x) ≤ 0 for any exclusive pair. Theorem 1 gives (a) a generalization gap bound and (b) a conflict-probability guarantee that decays at least as fast as the generalization gap.
• Not reported in the provided content: dataset sizes, and any explicit limitations section.

Methodology in Plain English

GenePheno treats a gene's sequence as the primary input and a set of phenotype abnormalities as multi-label targets.

1. Sequence encoding. Each gene's sequence is embedded with two encoders: GENERator for the gene sequence, and GoBERT for GO function information.
2. Dual-granularity GO information. Fine-grained GO terms (depth d > 2) are used as inputs and fused with the sequence representation through cross-attention. Coarse-grained GO categories (depth d = 2) are used as supervision targets at a bottleneck layer placed at the second-last network layer.
3. A contrastive loss instead of plain binary cross-entropy. The standard BCE approach assumes labels are independent and is hurt by class imbalance. The authors derive a contrastive multi-label loss from InfoNCE that pulls all positive phenotype labels together, keeps negatives clustered separately, and avoids the higher-order exponential terms that worsen imbalance. They note that the ZLPR loss used in InterLabelGO can be derived from their loss when τ = 1 with the denominator shift omitted.
4. Exclusivity regularization. An LLM pipeline with a structured query prompt scans the phenotype ontology (e.g. HPO, where hypotonia and hypertonia are semantically incompatible) to find mutually exclusive phenotype pairs among direct descendants of each non-leaf node. A soft Softplus regularizer on the logits of those pairs then discourages predicting both at once.
5. Interpretation at inference. The weights connecting GO bottleneck nodes to phenotype outputs quantify function–phenotype associations, producing a heatmap of which GO functions drive which phenotypes (Figure 3).
6. Evaluation protocol. CAFA-style preprocessing, stratification by phenotype label frequency into four bins (11–30, 31–100, 101–300, ≥ 301), with gene-centric F_max and phenotype-centric AUC as metrics. Baselines include max-BLAST, weighted BLAST, kmer2vec with logistic regression, and adapted sequence-to-GO models (DeepGOPlus, SPROF-GO, InterLabelGO), plus curated-modality phenotype methods on CAFA2 wPPI (DeepPheno, GraphPheno, HPOFiller, HPODNets, SSLPheno).

Why This Matters

Research impact. The work reframes phenotype abnormality prediction as a sequence-to-multi-label problem, removing the dependence on labor-intensive curated modalities that restrict existing methods to well-annotated genes. It also gives the multi-label loss an exclusivity-aware regularizer with formal guarantees, and shows how an ontology bottleneck can make a deep model's internal representation human-readable.

Real-world applications:

• Therapeutic discovery: predicting which phenotypes a gene knockout perturbs supports target identification and hypothesis-driven experimentation.
• Functional genomics: assigning functional consequences to newly discovered or poorly annotated genes that lack PPI or GO curation.
• Systems biology: mapping function–phenotype associations at scale to study how molecular mechanisms propagate to organism-level traits.
• Model-organism and clinical genetics research: the MPO case studies concern mouse knockouts and the HPO and GWAS case studies concern human gene–phenotype and SNP associations, illustrating how the same framework transfers across species and evidence types.

Industry relevance. Pharmaceutical and biotech groups that prioritize gene targets, and any organization building large-scale genomics pipelines, benefit from a model that runs on sequence alone and can justify its predictions with interpretable functional concepts.

Future Directions

• Extending the sequence-only approach to genes that currently have no GO annotations at all, since the framework still uses fine-grained GO embeddings as an input modality.
• Expanding the LLM-based exclusivity discovery pipeline beyond mutually exclusive pairs within direct descendants of each ontology node to broader logical constraints.
• Testing whether the GO bottleneck's function–phenotype weights can be used prospectively to guide new knockout experiments, given that the current validation is through case studies and cited literature.
• The paper reports no explicit limitations section in the provided content, so questions about scaling to the full space of genes and phenotypes, and about performance under extreme label sparsity, remain open.

Target Audience

Machine learning researchers working on multi-label and contrastive learning; computational biologists and bioinformaticians working on genotype–phenotype prediction, gene function annotation, and knockout studies; and practitioners in drug discovery or functional genomics who need scalable phenotype prediction for under-annotated genes and interpretable outputs they can inspect.

Authors’ abstract

Exploring how genetic sequences shape phenotypes is a fundamental challenge in biology and a key step toward scalable, hypothesis-driven experimentation. The task is complicated by the large modality gap between sequences and phenotypes, as well as the pleiotropic nature of gene-phenotype relationships. Existing sequence-based efforts focus on the degree to which variants of specific genes alter a limited set of phenotypes, while general gene knockout induced phenotype abnormality prediction methods heavily rely on curated genetic information as inputs, which limits scalability and generalizability. As a result, the task of broadly predicting the presence of multiple phenotype abnormalities under gene knockout directly from gene sequences remains underexplored. We introduce GenePheno, the first interpretable multi-label prediction framework that predicts knockout induced phenotypic abnormalities from gene sequences. GenePheno employs a contrastive multi-label learning objective that captures inter-phenotype correlations, complemented by an exclusive regularization that enforces biological consistency. It further incorporates a gene function bottleneck layer, offering human interpretable concepts that reflect functional mechanisms behind phenotype formation. To support progress in this area, we curate four datasets with canonical gene sequences as input and multi-label phenotypic abnormalities induced by gene knockouts as targets. Across these datasets, GenePheno achieves state-of-the-art gene-centric $F_{\text{max}}$ and phenotype-centric AUC, and case studies demonstrate its ability to reveal gene functional mechanisms.