Scigantic
Fine-tune open protein and DNA language models — ESM-2, ESMC, DNABERT-2 — from a CSV of sequence-to-label pairs, no GPU setup.
Overview
Scigantic is a turnkey fine-tuning service for open protein and DNA sequence language models. The workflow is deliberately minimal: upload a CSV that maps each sequence to a label, choose a backbone, and train — the platform handles the GPUs, training loop, and serving. It is built for researchers who have labelled sequence data and want a task-specific model without writing training code or managing infrastructure, whether the target is protein function, stability, expression, or a genomic property.
What you can run on Scigantic
Scigantic fine-tunes and serves a focused set of protein and DNA/genomics backbones. For proteins it supports ESM-2 (with ESMFold for structure) and ESMC — the ESM Cambrian protein language model — covering sequence-to-property prediction across function, stability, and binding tasks. For DNA and genomics it supports DNABERT-2, a tokenized genomic language model. Two further genomic backbones, the Nucleotide Transformer and GENA-LM, are forthcoming and will extend coverage of long-range DNA sequence tasks. Each available backbone has its own fine-tuning page, and every model is offered for both inference and label-based fine-tuning. The underlying weights live on Hugging Face; Scigantic provides the hosted training and inference layer on top of them.
Fine-tuning models on Scigantic
The core mode of access is CSV-driven fine-tuning: you supply a spreadsheet of sequence-to-label pairs, pick ESM-2, ESMC, or DNABERT-2 as the backbone, and receive a trained model you can run for predictions. There is no cluster to provision and no PyTorch to write, which makes it a fit for wet-lab scientists, bioinformaticians, and protein engineers who want to turn assay results or annotated sequences into a predictive model quickly. The forthcoming Nucleotide Transformer and GENA-LM backbones will bring the same upload-and-train flow to additional genomic modeling problems.
Run inference on Scigantic (5)
Meta AI's family of protein language models scaled to 15B parameters, paired with ESMFold for fast, alignment-free atomic-level structure prediction.
Multi-species genomic foundation model swapping k-mer tokenization for byte pair encoding, matching Nucleotide Transformer with 21x fewer parameters.
DNA foundation models from 500M to 2.5B parameters, trained on 3,200+ human genomes and 850 species for variant effect prediction.
Protein language model trained on roughly 2.8 billion sequences, forming the representation core of Biohub's world model of protein biology.
Family of transformer-based DNA language models using BPE tokenization and BigBird sparse attention to reach context lengths up to 36,000 base pairs.
Fine-tune on Scigantic (5)
Meta AI's family of protein language models scaled to 15B parameters, paired with ESMFold for fast, alignment-free atomic-level structure prediction.
Multi-species genomic foundation model swapping k-mer tokenization for byte pair encoding, matching Nucleotide Transformer with 21x fewer parameters.
DNA foundation models from 500M to 2.5B parameters, trained on 3,200+ human genomes and 850 species for variant effect prediction.
Protein language model trained on roughly 2.8 billion sequences, forming the representation core of Biohub's world model of protein biology.
Family of transformer-based DNA language models using BPE tokenization and BigBird sparse attention to reach context lengths up to 36,000 base pairs.