Mendelspod Podcast

This is a free preview of a paid episode. To hear more, visit www.mendelspod.com

Is biology headed back into the genome?
While much of biology has been expanding outward into single cells, spatial biology and ever more biological context, Žiga Avsec and his team at Google DeepMind are making the case that there is still an enormous amount to learn by going deeper into DNA itself.
Their new AlphaGenome Atlas uses AlphaGenome to predict the molecular effects of every possible single letter change in the human genome. That’s some nine billion variants. Now researchers can easily explore how a variant might affect gene expression, or splicing, or other layers of gene regulation across different cell types. The Atlas also introduces a variant impact score (VIS) designed to help researchers quickly identify which variants deserve a closer look.
Avsec argues that the scale itself opens new possibilities. Researchers can use the Atlas to prioritize rare variants across enormous cohorts, but they can also work backward from the predictions to investigate something more fundamental, the regulatory grammar of the genome. The grammar of the genome—long a provocative idea, but still not cracked. Short DNA motifs could act something like words, and AlphaGenome may help reveal how those words work together to activate or repress genes across different biological contexts.
There are important limitations. AlphaGenome predicts molecular consequences rather than phenotype, and Avsec says the distance between genotype and phenotype remains long and complex. The model also has more difficulty with regulatory elements that far from genes—a proximity issue—and has so far been trained largely on bulk tissue data rather than the much richer universe of individual cell types and cell states.
For now, Avsec just wants researchers to use the Atlas. The portal is free for noncommercial research, and he explicitly invites scientists to tell the DeepMind team what works and what does not. That feedback matters because the Atlas is not being presented as a finished map. It is part of a cycle in which better models suggest better experiments, those experiments generate better data, and better data produce the next generation of models. Bioinformaticians always want two things. Better data and more data.
This may be an argument for where genomics is headed. Rather than AI replacing experiments, Avsec hopes it will give researchers greater confidence about which experiments are worth doing and ultimately lead to more experiments with more positive findings.
The AlphaGenome Atlas is available free for noncommercial research here:
https://deepmind.google/science/alphagenome/

Mass spectrometry has been one of the most powerful technologies in clinical testing for decades. So why is it still largely confined to specialty labs?
Don Mason has spent more than 25 years in clinical mass spectrometry. Now Senior Marketing Manager for Mass Spectrometry at Roche Diagnostics, he says the technology’s strength has also been its weakness: “Part of its power and part of its challenges are linked.” Mass spec is extraordinarily sensitive, selective and flexible, but traditionally requires specialized operators, complicated workflows and batch processing.
That may finally be changing. Last year Roche launched their new cobas Mass Spec solution designed to automate the process from sample preparation through result reporting. It also brings random-access mass spectrometry into the routine clinical lab. Mason says a sample can now produce a numerical result in as little as 34 minutes, compared with turnaround times that can stretch into days with traditional workflows. He points to transplant drug monitoring and antibiotic and antifungal monitoring in critically ill patients as examples where that difference could matter clinically.
But automation isn’t simply about replacing expertise. Mason argues that moving established assays onto standardized systems could free mass spectrometrists to develop the next generation of tests: “This in turn becomes the innovation engine for tomorrow’s routine mass spectrometry-based tests.”
Roche currently has seven assays available in the U.S., with more expected beginning in 2027. Looking further ahead, Mason sees a much bigger transition. Just as MALDI-TOF transformed microbial identification, he predicts LC-MS will become standard equipment in more and more clinical laboratories over the coming decade.


This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.mendelspod.com/subscribe

This is a free preview of a paid episode. To hear more, visit www.mendelspod.com

Jennifer Dionne, Stanford physicist and co-founder of Pumpkinseed, joins us to talk about a radically different way of reading proteins. Pumpkinseed’s deSIPHR technology combines nanophotonics with Raman spectroscopy to detect the molecular vibrations of individual amino acids. The goal is de novo protein sequencing that can read not only the 20 canonical amino acids, but potentially the enormous alphabet created by post-translational modifications and other forms of protein variation.
Raman spectroscopy is super cool even if nearly a century old. Shine a laser on a molecule and a tiny fraction of the photons change color as they interact with the molecule’s vibrations, producing a characteristic molecular signature. The problem has always been sensitivity. As Dionne explains, only about one photon in a million undergoes this Raman scattering. Her work uses nanophotonics for specially patterned materials to amplify that faint signal by orders of magnitude.
Why does that matter? Mass spectrometry has been the workhorse of proteomics, but it loses much of the sample during ionization and generally depends on existing catalogs for identification. Pumpkinseed wants to read what is actually there, including proteins and modifications we may never have seen before.
One early application is particularly timely with the recent news of cancer vaccines. Working with Genentech, Pumpkinseed is studying immunopeptides, the protein fragments displayed on the surface of cells that allow the immune system to distinguish healthy from diseased tissue. Direct sequencing of these peptides is a way to improve personalized cancer vaccines by identifying the mutations actually present in an individual patient’s tumor.
Ultimately, the ambition is much larger. “We want to be able to sequence all of the proteins that are in individual cells,” Dionne says. For biology and for AI models trying to learn biology, she argues, we first need to learn how to read much more of its language.

This is a free preview of a paid episode. To hear more, visit www.mendelspod.com

Digital technology has been promising to transform pathology for years. But the last year looks different. Roche paid roughly $1 billion for PathAI. Tempus acquired Paige. Other deals are adding to a sudden wave of consolidation. And AstraZeneca is developing a computational pathology algorithm for TROP2 that could become a companion diagnostic used to determine which patients receive a drug.
DeciBio partner Katie Maloney says these are signs that digital pathology may finally be reaching an inflection point.
The important shift is from digital to computational pathology. Until recently, much of the value proposition was about making an existing workflow more efficient. Now algorithms are beginning to extract clinical information that a pathologist could not simply determine by eye. Maloney points to tools that can predict prognosis, stratify patients and potentially predict drug response. Computational pathology is beginning to compete with, and increasingly complement, molecular diagnostics.
The transition is still early. Maloney estimates that only 20 to 30 percent of US labs have adopted even a slide scanner. Reimbursement remains a major obstacle, with labs generally not paid for scanning slides, using image management software or deploying computational algorithms. And some of the hardest problems are surprisingly basic. Different labs stain the same tissue differently, creating variability that algorithms must accommodate if they are going to work across thousands of clinical sites.
But pharma may change the equation. Maloney is watching to see whether AstraZeneca’s work proves to be an isolated example or the beginning of something much larger. If computational pathology becomes important across a significant share of new drugs, particularly antibody drug conjugates, pathology images become another rich source of patient data that can be layered with clinical and molecular information. As Maloney puts it, computational pathology is becoming “not just a tool for pathologists, but it’s a precision medicine tool.”

This is a free preview of a paid episode. To hear more, visit www.mendelspod.com

It was the heady days of the CRISPR revolution and of the gene therapy Casgevy. And when longtime genomics editor and author Kevin Davies went looking for a book that told the story of sickle cell disease and could not find one, he was perplexed and then inspired.
A few years later, the result is Curved Air, a biography of sickle cell anemia that traces one of the most remarkable arcs in modern biology. Davies begins with Victoria Gray, the first sickle cell patient treated with the CRISPR therapy that became Casgevy. Her transformation leads him backward through more than a century of discovery, from the first description of sickled blood cells to the identification of sickle cell as the first molecular disease. He writes of the extraordinary biology of fetal hemoglobin that made today’s therapy possible.
This is also a story about the gap between biology and medicine. Davies explores the neglect and discrimination endured by sickle cell patients and the difficulty of bringing a multimillion dollar therapy to those who need it. With a list price of $2.2 million for Casgevy, there is an enormous challenge of extending this advanced therapy to the millions of patients around the world who are in need.