Cherimoya: A Small, Fast Genomic S2F Model Reaching SOTA

On July 15, genomics researcher Jacob Schreiber posted a series introducing Cherimoya, a lightweight sequence-to-function (S2F) model for local regulatory function prediction in genomic data, noting that he had presented it the day before at the ISCB RegSys conference. Its core pitch is being "smaller, faster, and cheaper" while still reaching state-of-the-art (SOTA) on related tasks, which makes it especially attractive for design and screening workloads that need heavy iteration.

Key Details

According to Schreiber, Cherimoya is built on the Cheri block, a ConvNeXT architecture adapted for genomic data, and the team implemented two custom GPU kernels to speed it up. At a configuration of 128 filters it has fewer than 1 million parameters, training typically takes only about 5–20 minutes, throughput can exceed 40,000 samples per second, and run-to-run variance is low with good stability. Anshul Kundaje of the same lab also endorsed it, emphasizing that the author did thorough hyperparameter exploration and efficiency optimization, making it suitable for study and reproduction.

Performance and Comparison

Schreiber says these tiny models, with under 1M parameters and only minutes of training, still reach SOTA on related tasks. In comparisons such as ENCODE, the gap between Cherimoya and much larger models is markedly narrowed. He notes that models like Borzoi and AlphaGenome beating ChromBPNet is unsurprising given their more parameters, larger receptive fields, and more training data; the more notable point is that Cherimoya, without using more information, approaches these systems' performance at a far smaller scale.

Availability and Significance

The model has been added to the Programmable Genomics package. The work's significance is demonstrating a "small models can be strong too" path for genomic tasks: with training costs of just a few minutes and high inference speed, such models become especially practical for design and screening scenarios that demand large amounts of iteration.

2026-07-15 ~ 2026-07-15 · 6 related posts

• [source] Efficient Small Model for Genomic Function Prediction — anshulkundaje · 2026-07-15
• Lightweight S2F Model Cherimoya Achieves SOTA — jmschreiber91 · 2026-07-15
• [source] Cherimoya: A Model Tailored for Genomic Data — jmschreiber91 · 2026-07-15
• [source] Cherimoya: Fast and Stable Small Model — jmschreiber91 · 2026-07-15
• Tiny Models Can Achieve SOTA — jmschreiber91 · 2026-07-15
• Small Models Approach SOTA on Genomic Tasks — jmschreiber91 · 2026-07-15