Helical
Run and fine-tune single-cell and genomics foundation models — Geneformer, scGPT, Evo 2 — via an open framework, hosted platform, and API.
Overview
Helical is an open-core framework and hosted platform for running and fine-tuning single-cell and genomics foundation models. It wraps a fragmented set of transcriptomics and DNA models behind one consistent interface, so you can load a model, generate cell and gene embeddings, and fine-tune on your own data without stitching together per-model codebases. This is the place to look when you want inference or fine-tuning for single-cell foundation models rather than protein structure prediction.
What you can run on Helical
Helical supports a curated set of single-cell transcriptomics and genomics models through a shared API. On the single-cell side it runs Geneformer, scGPT, UCE (Universal Cell Embeddings), TranscriptFormer, and Tahoe-x1 — models for cell-type annotation, batch integration, perturbation response, and generating embeddings from scRNA-seq data. For genomics and DNA sequence modeling it runs HyenaDNA and Evo 2, long-context genomic language models. Every supported model ships with a README and model card in the framework repository, and each can be run for inference or fine-tuned on labelled datasets. Weights for these models are published on Hugging Face by their original authors; Helical provides the unified runtime around them rather than hosting the weights itself.
Running and fine-tuning models on Helical
There are two ways in. The open-core Python framework lets you self-host: pull a model through the common interface and run or fine-tune it on your own infrastructure, which suits ML engineers and computational biologists who want full control. On top of that, a hosted "Model Factory" and managed fine-tuning let you train and serve models without provisioning GPUs, which suits teams that want a turnkey path from scRNA-seq or genomic data to a task-specific model. Because every model sits behind the same API, swapping backbones or benchmarking Geneformer against scGPT against TranscriptFormer is a small change rather than a rewrite — useful for anyone standardizing a single-cell modeling stack.
Run inference on Helical (7)
Generative pretrained transformer trained on 33 million human cells for single-cell annotation, batch correction, and perturbation prediction.
Single-cell foundation model pretrained on about 30 million human transcriptomes, using rank-value encoding for context-aware gene network inference.
Genomic foundation model built on the Hyena operator, processing DNA at single-nucleotide resolution with context windows up to 1 million tokens.
Genomic foundation model trained on 9.3 trillion DNA base pairs across all domains of life, with 40B parameters and a 1-million-token context.
Single-cell foundation model producing species-agnostic cell embeddings by representing genes through frozen ESM-2 protein language model embeddings.
Generative single-cell foundation model trained on 112 million cells from 12 species, autoregressively modeling gene identities and expression counts.
Perturbation-trained single-cell foundation models (up to 3B parameters) that jointly model genes, cells, and compounds for precision oncology tasks.
Fine-tune on Helical (7)
Generative pretrained transformer trained on 33 million human cells for single-cell annotation, batch correction, and perturbation prediction.
Single-cell foundation model pretrained on about 30 million human transcriptomes, using rank-value encoding for context-aware gene network inference.
Genomic foundation model built on the Hyena operator, processing DNA at single-nucleotide resolution with context windows up to 1 million tokens.
Genomic foundation model trained on 9.3 trillion DNA base pairs across all domains of life, with 40B parameters and a 1-million-token context.
Single-cell foundation model producing species-agnostic cell embeddings by representing genes through frozen ESM-2 protein language model embeddings.
Generative single-cell foundation model trained on 112 million cells from 12 species, autoregressively modeling gene identities and expression counts.
Perturbation-trained single-cell foundation models (up to 3B parameters) that jointly model genes, cells, and compounds for precision oncology tasks.