vHAMSTeR Models v1.4.0

Model weights for vHAMSTeR (Virus Host Assignment Model using Sequence Transformers and Reading-frames), a tool for predicting the host of a virus from its genomic sequence.

Model description

vHAMSTeR predicts viral host as one of five classes: 'Prokaryote', 'Animal', 'Plant', 'Fungi' or 'Protists, Algae & Oomycetes'.

It is designed for viral sequences up to 10 kbp. Longer sequences are split into 10 kbp chunks, scored independently, and aggregated to produce a genome-level consensus.

Architecture

Each fold in the ensemble consists of:

PEFT adapter on InstaDeepAI/NTv3_650M_pre — a 650M-parameter nucleotide language model fine-tuned on viral sequences
XGBoost stacking layer — incorporates geNomad marker features alongside the transformer embeddings
GLM gate — learned weighting between the transformer and marker feature streams
Classifier head — produces the final 5-class probability distribution

Predictions are calibrated using length-aware vector scaling (length_aware_vector_scaling_anchors_5.json).

The full ensemble runs 5 cross-validation folds and aggregates their outputs.

Repository structure

  vhamster_models_v1.4.0/
  ├── length_aware_vector_scaling_anchors_5.json   # calibration parameters
  ├── fold_0/
  │   ├── config.json
  │   ├── best_macro_auprc_model/
  │   │   ├── adapter_config.json
  │   │   ├── adapter_model.safetensors
  │   │   └── classifier_head.pt
  │   └── xgb_stacking_artifacts/
  │       ├── xgb_stacking_artifacts.json
  │       ├── xgb1_binary_model.json
  │       ├── xgb2_euk_model.json
  │       ├── marker_classification.pkl
  │       └── marker_cutoffs.pkl
  ├── fold_1/ … fold_4/   # same structure

Note: The NTv3 tokenizer and base model weights are no longer bundled here. They are downloaded separately from InstaDeepAI/NTv3_650M_pre by running vhamster-install-models --cache-base-model, which will display theInstaDeepAI licence terms and prompt for HuggingFace authentication.

Usage

These weights are intended to be used via the vHAMSTeR command-line tool. See the vHAMSTeR repository for full installation instructions.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for DOEJGI/vhamster-models

Base model

InstaDeepAI/NTv3_650M_pre

Finetuned

(2)

this model