💻 bioinformatics

Population-scale drug target and variant prioritization with Varformer

Varformer is a novel multimodal transformer model that leverages positive-unlabeled learning and cross-modal attention to integrate variant and gene-level features, achieving state-of-the-art performance in prioritizing both drug targets and pathogenic variants across diverse genetic ancestry groups.

Original authors: Wenteler, A., Neduva, V., Genes & Health Research Team,, Cabrera, C. P., Wei, W., Barnes, M. R.

Published 2026-10-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Wenteler, A., Neduva, V., Genes & Health Research Team,, Cabrera, C. P., Wei, W., Barnes, M. R.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Finding the right target for a new medicine is one of the hardest parts of drug discovery. Scientists must identify a specific protein in the body that, if blocked or boosted, will treat a disease without causing harm. For decades, researchers have relied on human genetics to guide this search, looking for natural variations in our DNA that protect people from illness or make them susceptible to it. However, most of these genetic studies have focused heavily on people of European ancestry. This creates a blind spot, because the genetic makeup of a person from South Asia, Africa, or the Americas can differ significantly, meaning a drug that works for one group might fail or cause side effects in another. Furthermore, existing computer models often struggle to handle the sheer complexity of these genetic differences, usually collapsing thousands of individual DNA changes into a single, simplified summary for each gene. This approach risks missing the subtle, varied patterns that actually determine whether a gene is a safe and effective target for a new therapy.

A team of researchers has developed a new artificial intelligence system called Varformer to solve these problems. Instead of simplifying the data, this system learns directly from the raw, complex patterns of genetic variation found across diverse human populations. The researchers trained the model using genetic data from four major groups: people of South Asian, African, Admixed American, and Non-Finnish European ancestry. By feeding the computer millions of specific DNA changes, known as missense variants, along with information about how genes function and where they are active in the body, Varformer learned to recognize the signatures of successful drug targets. The system treats the problem as a search for the "good" targets among a vast pool of unknowns, using a learning strategy that acknowledges that just because a gene hasn't been tested yet doesn't mean it is a bad candidate.

The results show that Varformer can distinguish between genes that have already been validated as drug targets and those that have not, with remarkable accuracy. When tested on three different sets of challenges, ranging from well-known protein families to more obscure candidates, the system consistently ranked true drug targets near the top of its list. On the most difficult test, which involved targets that had not yet been approved by regulators, Varformer outperformed previous methods by a significant margin, identifying nearly two additional correct targets for every twenty genes it reviewed. Crucially, the system did not just memorize the data; it learned to prioritize targets that remained strong candidates across all four ancestry groups. This suggests that the genetic support for these targets is robust and not dependent on the specific genetic background of a single population, offering a way to de-risk drug development for a global audience.

Beyond identifying whole genes, the system also scores individual DNA changes. In a test involving a cohort of British Bangladeshi and Pakistani participants, the model successfully ranked known disease-causing variants within the top twenty candidates for the majority of individuals carrying them. This ability to pinpoint specific harmful mutations is vital for understanding how genetic differences contribute to disease in individuals. The researchers found that combining the model's scores with existing methods improved the ability to identify harmful variants, suggesting that the new system captures unique information that other tools miss. While the model is not a crystal ball that guarantees a drug will succeed in clinical trials—since trials depend on many factors beyond genetics, such as trial design and commercial viability—it provides a powerful, unified foundation for deciding which genes and variants are worth pursuing first.

The work represents a shift toward a more inclusive and precise approach to drug discovery. By training on diverse populations and learning directly from the complex landscape of genetic variation, Varformer offers a way to find therapeutic targets that are likely to work for people of all backgrounds. The researchers have made the system available as an open-source tool, allowing other scientists to use it to prioritize targets and variants across the entire human genome. This approach does not just improve the odds of finding a new medicine; it helps ensure that the medicines found are relevant and safe for the entire human population, not just a fraction of it.

Technical Summary: Varformer for Population-Scale Drug Target and Variant Prioritization

Problem Statement
Identifying safe and efficacious therapeutic targets remains a primary bottleneck in drug discovery, contributing significantly to clinical trial failures. While human genetics offers a powerful source of evidence for target discovery, current prioritization frameworks face two critical limitations:

Handling Variable-Length Data: Existing models often collapse variable numbers of missense variants per gene into handcrafted summary statistics. This aggregation loses the heterogeneous functional landscape of naturally occurring variation.
Population Bias and Scarcity: Most genetic evidence is derived from cohorts of European ancestry, risking therapeutic inequities. Furthermore, target prioritization is a "positive-unlabeled" (PU) problem: few successful targets are known, most candidates are unlabeled, and clinical trial failures do not necessarily imply an ineffective target (due to design or commercial factors). Current methods struggle to learn from this sparse, biased data while accounting for diverse genetic architectures.

Methodology: Varformer Architecture
The authors present Varformer, a multimodal transformer designed to integrate population-scale missense variation with gene-level biological features. The architecture operates on four distinct ancestral cohorts: South Asian (SAS), Non-Finnish European (NFE), Admixed American (AMR), and African/African-American (AFR).

Input Modalities:
- Gene Characterization (GC): Integrates global gene-level features including tractability (ligand binding, small molecule binders, pockets, chemical probes), safety (safety events), functional annotations (Gene Ontology terms), subcellular localization, and tissue-specific expression.
- Population Variant Characterization (PVC): Encodes ancestral missense variation as a sequence of variant tokens. Each token includes the AlphaMissense pathogenicity score, a model-position index (derived from amino acid coordinates), and the specific mutation identity. Variants are sorted by pathogenicity and truncated/padded to a fixed length ( $N_{max} = 1,024$ ) per gene.
Model Architecture:
- Shared Variant Encoder: A Transformer encoder processes the variable-length sequence of variant tokens to produce contextual embeddings for each variant.
- Cross-Modal Attention: A gene-to-variant cross-attention mechanism aggregates information from the variant tokens into a gene-level representation, conditioned on the gene's biological features (tractability, safety, etc.).
- Dual Output Heads:
  1. Gene Prioritization Head: Takes the fused gene-variant representation to predict a therapeutic target score ( $S_\theta \in [0, 1]$ ).
  2. Variant Prioritization Head: A shared linear layer scores individual variant tokens directly, independent of the cross-attention step, to prioritize specific alleles.
Learning Objective:
- Gene Level: Trained using non-negative positive-unlabeled (nnPU) learning. The model treats unlabeled genes as a mixture of viable-but-untested targets and true negatives, estimating classification risk without explicit negative labels. This addresses the selection bias inherent in clinical trial data.
- Variant Level: Trained via binary cross-entropy using allele frequency as a proxy for selective constraint (low frequency implies pathogenicity/constraint).
- Joint Training: The model minimizes a combined loss function ( $L_{joint} = L_{gene} + \lambda_v L_{variant}$ ), updating the variant encoder jointly for both tasks.

Key Results

Representation Learning: UMAP projections demonstrate that Varformer embeddings concentrate known therapeutic targets into distinct regions of the latent space across all four populations, achieving a 17- to 19-fold enrichment in target density compared to input summary statistics.
Benchmark Performance: Evaluated on three benchmarks of increasing difficulty (Pfam Druggable Targets, Recently Approved Targets, and Chemically Putative Targets/Pharos):
- Precision@20: Varformer achieved 0.95 (Pfam), 0.86 (Recently Approved), and 0.48 (Pharos).
- Comparisons: On the hardest benchmark (Pharos), Varformer surpassed a logistic regression ablation (Varformer-LR) by 20% and the state-of-the-art DrugnomeAI by 23%.
- Robustness: Cross-population rank normalization identified targets with consistent genetic support across ancestries (e.g., HCN3, INSRR, TIE1), demonstrating that prioritization can be de-risked against population-level genetic variation.
Prospective Validation: In a time-stratified evaluation (training cutoff June 2023, outcome June 2026), 10 of 13 targets that progressed to Phase III or approval were ranked above the model's threshold.
Variant Prioritization: In the Genes & Health (G&H) cohort, the variant encoder ranked known pathogenic variants within a carrier's top 20 candidates in 57% of 1,164 carrier-variant pairs. The variant score ( $\bar{s}_v$ ) achieved an AUROC of 0.890 for rare variants, outperforming gene-level scores and showing complementary value when combined with AlphaMissense.

Significance and Claims
The paper claims that Varformer shifts drug target prioritization from gene-level aggregates and GWAS-led signals to individual variants modeled directly at the population scale. By combining a multimodal transformer with positive-unlabeled learning, the authors demonstrate that:

Therapeutic potential can be estimated from genetic variation, multi-omic descriptors, and functional annotations without requiring explicit negative labels.
Population-specific de-risking is possible: Cross-population analysis identifies targets whose prioritization is robust to inter-population genetic variation, offering a computational method to reduce therapeutic inequities.
Unified Target and Variant Prioritization: The model provides a single framework for deciding which targets and which specific variants to pursue, linking the gene and the variant in one representation.
Disease-Agnostic Utility: As a disease-agnostic model, Varformer can prioritize across the entire exome, though the authors note it does not account for therapeutic-area-specific hurdles (e.g., blood-brain barrier permeability).

The authors emphasize that while the model provides strong computational evidence for target viability, it does not establish population-independent efficacy or safety, which require experimental and clinical validation. The variant scores are presented as prioritization tools rather than calibrated clinical diagnostic evidence.

Technical Summary: Varformer for Population-Scale Drug Target and Variant Prioritization

More like this