BioWire Bytes 001 - AlphaGenome Predicts Gene Regulation

Byte-sized Biotech

To all of my fellow Americans, Happy Independence Day! We are living through the most important period in human history, at the epicenter of many significant advancements and innovations in biotechnology, health, and artificial intelligence. It is the framework of our country and the diverse people who inhabit it that made this progress possible. While you are enjoying time outdoors with friends and/or family, a pool (or other body of water), fireworks, and BBQ, remember how much we have to be grateful for and the gravity of this remarkable period. There is nothing we did to deserve being born here, but since we are here, let’s strive to leave the woodpile higher for those who come next.

To the rest of my readers outside of the US, continue being relentless! I am happy to celebrate your holidays with you as well (;

Some quick housekeeping:

I will be trying a new format for the blog. My intention with BioWire has been to provide one general update on biotech, AI, and healthcare every other week with 3-5 topics. Rather than cramming this all into a single newsletter, which are becoming quite long, I am going to experiment with a new format:

• BioWire Bytes: Quick-read newsletters 4–5 times per week (approximately 5-minute reads).

• BioWire Weekly: One weekly newsletter that recaps Bytes in a concise format.

Overall, I want this to be informative, timely, and, perhaps most importantly, sustainable for me as a hobbyist writer. I would greatly appreciate your feedback in the comments.

At that, let’s dive into our first BioWire Byte!

---

First, if you enjoy these updates, consider subscribing and becoming a part of our growing community!

---

DeepMind’s new AlphaGenome AI model is trained to read long stretches of DNA and predict how genes are regulated

I truly believe the most impactful applications of artificial intelligence will be through its impact on biotechnology and human health. This past week is another datapoint for that hypothesis with the launching of AlphaGenome, an ambitious artificial intelligence model that aims to decode the genome’s regulatory “instructions”. In simple terms, it takes in a raw DNA sequence and predicts what that sequence is doing inside a cell (Avsec et al., 2025). What sets AlphaGenome apart is its unifying approach: this single model can perform many genomic tasks from a single input, rather than requiring separate tools for each task. The model can handle a remarkably long DNA sequence – up to 1 million letters (base pairs) at a time – and output thousands of molecular readouts about gene activity. These readouts include things like where genes start and end, how genes are spliced (cut and rejoined in RNA), how much RNA is produced (gene expression levels), and which parts of the DNA are open or “accessible” in the cell or bound by regulatory proteins. Impressively, AlphaGenome can even score the effect of a single-letter DNA mutation by comparing the model’s predictions for the original vs. the mutated sequence. In other words, it can flag whether a tiny genetic change might turn a gene up, down, on, or off – a crucial ability for understanding genetic diseases.

So, why should we care about a model predicting gene regulation? Simply put, AlphaGenome could accelerate genetics research and deepen our understanding of how our DNA works. Here are a few potential benefits:

• Rare disease genetics: The model can help scientists pinpoint which genetic variants might cause disease by disrupting gene regulation. By more accurately predicting the effects of DNA mutations, especially rare ones, AlphaGenome could identify the culprit variants in unexplained genetic disorders. This might lead to new diagnoses or even new therapeutic targets for conditions that have puzzled doctors till now.

• Regulatory DNA mapping: AlphaGenome provides a way to chart the regulatory regions of the genome – the switches and controls that tell genes when to turn on or off. This could assist in mapping the genome’s functional elements and understanding the “instruction manual” that dictates different cell types’ behavior. In the long run, these insights help researchers identify which DNA sequences are truly vital for specific cellular functions (and which might be safely altered in gene therapies).

• Synthetic biology: The AI’s predictive power might guide bioengineers in designing custom DNA sequences with particular functions. For example, one could imagine using AlphaGenome to create a synthetic genetic switch that activates a gene only in certain cells (say, nerve cells, but not muscle cells). This kind of fine-tuned control is valuable for developing gene therapies.

These are fun futuristic applications, but is AlphaGenome doing anything now?

To illustrate its potential, the DeepMind team has already put AlphaGenome to work on a real scientific puzzle. In a recent cancer study, researchers noticed mutations in a non-coding region of DNA in patients with T-cell acute lymphoblastic leukemia (T-ALL). When the team fed this sequence to AlphaGenome, the model correctly predicted that those mutations would activate a nearby gene (called TAL1) by creating a new binding site for a transcription factor. This matched what was observed in the patients’ cells, essentially confirming the disease mechanism with the AI model’s output. It’s a powerful example of how AlphaGenome can connect the dots between a DNA variant and its downstream effect, linking an obscure non-coding mutation to a change in gene activity that contributes to cancer.

One exciting aspect of AlphaGenome’s launch is that DeepMind has made it accessible to the scientific community. The model is available via an API (currently in a preview release) for non-commercial research use. This means academic and non-profit researchers can start experimenting with AlphaGenome on their own genomic data and problems. DeepMind plans to fully release the model in the future, but for now the API allows scientists to query the model and get predictions for research purposes. It’s worth noting again that these predictions are for research only and the tool is not approved for clinical or direct diagnostic use. DeepMind has also set up a community forum to gather feedback and discuss use cases, signaling that they welcome collaboration to unlock AlphaGenome’s potential. By sharing this model, DeepMind hopes that other experts will build on it, apply it to new challenges, and help refine its capabilities. In the long run, AlphaGenome and models like it could become an integral part of the genomics toolbox, accelerating discoveries in genetics, aiding the interpretation of DNA in health and disease, and bringing us closer to understanding the complex language of our genome.

---

These newsletters take significant effort to put together and are totally for the reader's benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

• Subscribe: Never miss an update by subscribing to the Substack.

• Share: Help spread the word by sharing posts with friends directly or on social media.

---

References:

https://deepmind.google/discover/blog/alphagenome-ai-for-better-understanding-the-genome/

Avsec, Ž., Latysheva, N., Cheng, J., Novati, G., Taylor, K.R., Ward, T., Bycroft, C., Nicolaisen, L., Arvaniti, E., Pan, J. and Thomas, R., 2025. AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model. bioRxiv, pp.2025-06.