Today instead, I want to highlight a couple of newer programs that showcase the power of combining regulome profiling with emerging AI models for drug discovery.

Most drug discovery approaches are reductionist by necessity.

If you’re screening millions or billions of compounds to find a hit, then you need to keep the assay simple. Usually that means measuring one thing at a time (does a compound bind its target? degrade a protein?). That also means engineering an artificial system, destroying the native biological state of the drug target. This narrow view works fine for simple mechanisms, the “druggables”. But most of these low-hanging fruits have been picked clean off the tree.

At Talus, we built a technology that profiles the entire native regulome.

Native means that each time that we test a compound, we see how it affects everything that controls the genome in an unmodified human cell. That includes transcription factors, chromatin regulators, remodelers, the RNA processing machinery, and the complex interplay between them all. Most technologies in biology have a Schrodinger’s Cat problem where measuring the system, by definition, requires disrupting it. What we get here is a unique opportunity to take high-resolution snapshots of what the regulome is doing without disturbing it, and see how it changes over time.

These snapshots open up a new kind of therapeutic opportunity.

Instead of screening compounds blindly in the lab with simple assays, we now have enough data to train AI to virtually screen billions of compounds, then bring the strongest hits into the lab to test, validate, and improve our models. Virtual hits are validated directly in native biological systems. This happens on the order of weeks.

The most important “undrugged” targets in complex disease are within the regulome and work through unpredictable, deeply interconnected mechanisms. Now, we can finally start to untangle how this system works, and I’ll share two recent examples.

A Protein-RNA Molecular Glue for Castration-Resistant Prostate Cancer

Six months ago, we initiated a second program with our regulome profiling engine targeting the androgen receptor (AR) in prostate cancer.

AR drives prostate cancer. The entire treatment paradigm is built around blocking it. First with hormone therapy, then with more potent AR inhibitors like enzalutamide and apalutamide. But prostate cancer is determined to reactivate AR any way it can. We see splice variants and mutations that block drug binding, amplification of the AR gene itself, even tumors that synthesize their own testosterone to boost AR signaling. The cancer works extraordinarily hard to keep AR signaling alive, and eventually it succeeds in most patients.

Our platform identified a way to cut off AR signaling before the protein is even made with a molecule that sticks to a regulome protein (NONO) and prevents the AR mRNA from maturing and ever being translated into AR protein, depleting AR protein from binding to the genome.

We started this program nine months ago, and regulome profiling has proven essential. Unlike a traditional screen where you’re measuring one binding event or one cellular readout, we can directly visualize:

• Compound binding directly to the NONO protein

• How that impacts levels of AR mRNA

• AR levels in the cell and directly on DNA, including all mutants or splice variants

• Which off-target proteins are being bound and altered

This technology lets us iterate incredibly quickly. Starting from our original hit compound, we’ve now tested over 350 versions and nominated TAL20231 as a lead compound that is currently in in vivo evaluation. Compared to our original hit, this compound has:

• 5-fold improvement in potency

• 10-fold enhanced selectivity with very few off-target binding events

• 95% target engagement at the NONO/AR mRNA complex

• A 10-fold improvement in metabolic stability

• Longer microsomal stability in both human (>145 min vs 26 min) and mouse (>49 min vs 16 min)

• Room to grow as we progress through lead optimization (MW of 416 g/mol)

The speed here is remarkable.

Six months from program start to a compound with this profile was only possible because regulome profiling gave us direct-to-biology mechanistic readouts at each optimization cycle.

Unlocking a Fusion Transcription Factor with AI regulome models

Beyond optimizing our programs in flight, our platform is unlocking new starting points across historically “undruggable” target space.

These targets can’t be solved with AlphaFold or structure-based design because many regulome proteins like transcription factors work through dynamic, context-dependent interactions across, not static binding pockets.

So we trained our own AI model.

This is possible now, with our regulome data atlas hitting an inflection with over 100M data points. The result is a multi-modal AI model combining our unique regulome data with state-of-the-art protein and chemical foundation models to run virtual screens for “undruggable” targets.

Already, our platform has delivered chemical starting points for 68% of all previously “undruggable” transcription factor families.

One example I’m particularly excited about is an oncogenic fusion transcription factor that drives a type of aggressive childhood tumor. It’s a perfect example of an “undruggable” target where two transcription factors are fused into one new mutant protein that activates genes that drive uncontrolled growth. These fusions have no enzymatic activity to inhibit, no clear binding pocket for a traditional small molecule, and they work by re-wiring the cell’s transcriptional machinery in complex ways.

Through lab-in-the-loop active learning, we converged on a series of molecules that directly disrupted the fusion and shut off its downstream gene targets. Here’s an example of 3 members in the series modulating 3 downstream gene targets in the expected directions.

We’re still early here, focused on hit expansion and establishing SAR. But it demonstrates the platform can generate chemical starting points even for targets most drug hunters have thrown in the towel for.

What’s Next

We’re energized every day by peering into biology other platforms can’t see, and translating directly into unlocking and accelerating new therapeutic opportunities.

Join our tech access program, or join our team.

Or see what else we’ve been up to

Subscribe now

Share

This blog post is a love-letter to my fellow proteomicists, not just the mass spectrometrists but the protein folders and the NMR spectroscopists, the CUT&RUNners and yes even the Western blotters. It’s also a rallying cry for us to get out of our echo chambers and collaborate to push the greater field of biology forward.

I’m headed home from my twelfth ASMS conference, on a flight to Baltimore as I write this out.

My first, the 2013 conference in Minneapolis, was both overwhelming and awe-inspiring. Although I’d done a bit of undergraduate research in a mass spectrometry proteomics lab, I didn’t get a chance to really use the instruments until I joined Dr. Steve Carr’s Proteomics Platform at the Broad Institute as a research associate several years later. There, my mentor Dr. Sue Abbatiello not only gave me access to the instruments but actively encouraged me to dive into the hardware, working collaboratively with a big instrument vendor (Thermo) on a hardware accessory for mass spectrometers (the early prototype of the FAIMS).

This is where the spark finally caught for me, and I was inspired by the potential of proteomics so much that I decided to go all-in and get my PhD, focusing on mass spectrometry proteomics.

Over the next few years, I managed to land (partially intentionally, but mostly serendipitously) among the front-lines of researchers developing and applying the “big data” method du jour for quantitative proteomics: data independent acquisition mass spectrometry (DIA-MS). Training at the University of Washington’s Genome Sciences department, my grad school cohort mostly consisted of genomicists and computational biologists, with us proteomicists frequently the outliers, so despite DIA-MS being a Big Deal in my mass spec circles, my departmental research reports mostly got glazed-over eyes and scattered polite claps. Amongst my fellow trainees, the joke was that proteins weren’t real – they’re “gene products” and therefore just a consequence of genomics – and to an extent, it’s true. Proteomics has definitely lagged behind genomics, for a variety of reasons.

I think the biggest is the lack of a shared space.

Genomics and transcriptomics had a commonality: the technology. DNA sequencing brought everyone together under a common readout, even if the sample prep and the data interpretation diverged. Proteomics is different, with dozens of instruments with hundreds of data outputs and interpretations. This difference became stark in the Genome Sciences department, where my fellow trainees could easily converse in the same “language” with each other, whether they were working on chromatin structure with ATAC-seq or mapping evolutionary trajectories with ancient DNA. They may have been using different sample preps or different interpretations, but it always seemed to me that they were united under a big “DNA Sequencing” umbrella.

Although I dearly love my fellow mass spec nerds at ASMS, I have to admit that we’re not the most welcoming field to outsiders.

Dr. Ben Orsburn (News in Proteomics Research, www.proteomics.rocks) often jokes about mass spectrometrists being some “guy in the basement”, with whom other scientists have to beg and bargain to get their samples run.

While mostly hyperbolic, there’s a grain of truth there. Mass spec proteomics data is not easily acquired, nor analyzed, and God forbid you have to actually make a biological interpretation from it. Mass spectrometry doesn’t even have a common file format, broken into vendor-specific proprietary file formats that require special converters, let alone even simple post-processing tools like batch correction. We mostly just borrow software from transcriptomics or even protein microarrays. Besides, just generating the data involves purchasing an LC-MS, which is going to cost $1M a pop, so nobody is casually getting one to tinker with like they might a minION for sequencing. And we haven’t even talked about sample prep, which historically has been home-brewed recipes and protocols that changed lab-to-lab, and only recently has started to conform into standardized kits.

But despite being low throughput and expensive and computationally frustrating, proteomics is still important, and the Free Market has noticed. Alternative technologies for protein detection and measurement have made quick work of expanding the market, making proteomics more accessible. Biofluid proteomics has been the beachhead for these technologies, representing a sample type with great therapeutic potential but also technically challenging for mass spectrometry with its large dynamic range. (Dynamic range refers to the distribution of protein abundances. Plasma has some proteins that are incredibly abundant, like albumin, and other proteins like metabolic enzymes that are very low abundance.) Understanding the plasma proteome is of such high commercial value that the Pharma Proteomics Project, a pharmaceutical consortium composed of 14 companies, is funding the measurement of 300,000 samples in the UK Biobank. 300,000 samples is unheard of for this type of quantitative proteomics work, but its enabled by these new technologies.

That’s not to say that there haven’t been other ambitious projects in proteomics before these next gen technologies. As a natural consequence of the completion of the Human Genome Project, the Human Proteome Organization (HUPO) launched the Human Proteome Project with the goal of assigning “a function for every protein” by mapping all the predicted proteins encoded by the human genome (). There’s also spatiotemporal aspects of proteomics to consider, with projects like the Human Protein Atlas.

We can’t ignore the elephant in the room either.

The 2024 chemistry Nobel Prize was awarded for proteomics – specifically, protein folding. Here’s where maybe I’ll get controversial, but I would consider protein folding to be under the broader umbrella of proteomics. After all, AlphaFold and Rosetta were only made possible by their underlying training data: decades and decades of painstakingly crystallized and measured protein structures, curated and deposited in a central database. The triumph of AlphaFold is really an homage to the PDB, a warehouse dedicated to structural proteomics no matter what the readout, whether X-ray crystallography or nuclear magnetic resonance (NMR) or electron microscopy (EM) techniques.

So as I’m getting ready to commune with my fellow mass spec-nerds at ASMS, I’m also thinking about how we – speaking for all protein-focused scientists here – could do a better job of reaching out to our fellow proteomicists that happen to use other readouts. The imaging microscopist, the protein folder, and yes even you, Western blotter, you’re all measuring proteins and I want to learn more about why and how you’re interested in proteins.

I want to learn how to link our collective data together in a way that will enable the next big advance, on the scale of AlphaFold or even bigger, maybe model a virtual cell? To do something like that will require trustworthy data, and to have trustworthy data means it will be validated and reproduced with orthogonal technologies. Sure there may be a dozen people all studying the same protein in the same disease space, but if they’re all using different approaches, that’s not a risk of competition, that’s a potential for collaboration. The existence of competing tech doesn’t shrink the individual slices of the Proteomics Market Size Pie, it expands the size of the field overall. Instead of building up taller and stronger fences, I’d love to see us open up the gates and collaborate across instrumentation, making bridges from protein catalogs to structure to function, and to scientific breakthroughs.

In the meantime, what can we do literally today?

We’re headed home now from an amazing ASMS conference, let’s keep the spirit of the conference going. Invite someone you don’t already know at your institute out for a “coffee chat” and talk shop. (Don’t forget those early career researchers, they’re the future of the field!) See a cool talk? Get in touch now with the presenter and propose a collaboration. Peruse the conference proceedings and check out the posters that are way outside of your depth and email the presenter to ELI5. Watch a recording from an oral session you’ve never been to before and, even better, get in touch with the authors to ask a question. And beyond the ASMS hangover, longer term, attend a conference outside of your field. Reach out to a researcher using a different technique, but asking similar biological questions, and offer to share data or run a follow-up experiment. If you’re really ambitious, reach out beyond researchers to nonprofit organizations, to investors, to federal policymakers, and ask them what they think needs to happen for proteomics to make the next big breakthrough. Is it technology development? Is it training? Is it funding? Having a more holistic approach over the next decade will benefit the individual (greater impact, greater reach), the scientific community (virtual cell), and humanity (therapeutics, diagnostics, agriculture, forensics).

That’s what I’ll be doing on the plane back to Seattle, following up on the brilliant things I learned and the amazing new people I met.

We have such a unique community in mass spectrometry and the ASMS conference is such an overwhelming yet awesome experience. This year I got uncomfortably out of my depth, and talked to people about how mass spec proteomics can be doing a better job of making new connections beyond our own community. The future of proteomics is interactions. Interactions between techniques, interactions between data, and interactions between people. The more we connect, the more we discover.

Leave a comment

Subscribe now

Before founding Talus Bio, I trained with Jay Bradner, a champion of open source drug discovery, and this philosophy runs deeply into our work today.

Much like Meta’s open-sourcing of transformative AI tools such as Llama, we see open source software (OSS) as a way to amplify impact and accelerate breakthroughs globally. For Talus, the most substantial value we will create is in the life-saving medicines we aim to deliver, not the software we create along the way. In fact, making our tools open-source will increase their value, enabling the broader scientific community to build on them, fostering innovation that benefits everyone.

This post, originally written by our Head of Data Science and Engineering, Will Fondrie, explores the power of open-source software in proteomics, strategies for sustainable development, and why openness is essential for advancing science.

This blog post was originally written and shared by Will Fondrie, our Head of Data Science and Engineering, on his personal blog. We are now sharing it here (with permission) to celebrate the release the our latest preprint, “Open-source and FAIR Research Software for Proteomics,” which was written as a collaboration between Will and other leaders in proteomics software.

In modern scientific research, computational methods play a pivotal role---and in many cases are the central focus of a scientific paper. Yet, despite the heavy reliance on algorithms and software, code is often not treated with the rigor and expectations of openness as data in many fields. Sparked by a recent Twitter/X controversy in the proteomics field, this post explores the importance of open-source software in scientific research and why it’s essential for driving scientific progress. It also discusses how folks can commercialize their open-source software to support its continued development. It will not discuss why for-profit companies should release open-source software; that'll be a future post.

But first, I'll offend some folks:

Computational Papers Without Code Are Just Advertisements

In 2024, both the Nobel Prizes in Physics and Chemistry were awarded for breakthroughs in deep learning and protein structure prediction. These accomplishments highlight how impactful computational research can be, yet they also stress the importance of transparency. Without open-source code, computational papers are little more than advertisements for an idea—claims that cannot be fully verified or reproduced.

A striking example is the controversy surrounding the 2024 AlphaFold3 paper which—unlike its predecessor—failed release the code. This move sparked frustration in the scientific community, because, without code, it's impossible to replicate or scrutinize the work. Publishing an algorithm without code is like withholding experimental data in traditional research—it prevents others from building on or challenging the findings.

Moreover, even the best-described algorithms often differ in their real-world implementations. Code provides the essential details needed to truly understand how a model works in practice. Open-source code fuels scientific progress by allowing others to reproduce results, identify bugs, and push the field forward. Without it, computational research risks becoming a black box, stifling innovation rather than promoting it.Indeed, I've seen this in my own work where users have found nuances not captured in the methods of the papers we've published, or have found bugs that we subsequently fixed.

One example of open-source impact from my own work has been with Casanovo. Since Casanovo is fully open-source, we've seen an explosion of new deep learning methods for *de novo* sequencing peptides from mass spectra. Some of these tools explicitly depend on Casanovo, others depend on the underlying open-source library I wrote (Depthcharge), while others literally copy and paste our codebase into their work as their starting point. All of this is awesome and what science is all about; exploring areas that have never been explored before, sharing what we find, and providing a stepping stone to propel all of humanity forward—not just ourselves.

Open-Source Promotes Scientific Transparency and Rigor

We hit on this in the previous section, but transparency is the cornerstone of science; and open-source software exemplifies this principle. Sharing code allows other researchers to not only reproduce the results but also scrutinize them. Open-source fosters an environment where peer review extends beyond the publication of a paper—it includes the software itself. Indeed, the software becomes alive when it picked up by users. Bugs or inefficiencies can be spotted by others, improving the overall quality of the science. For me, a prime example of this is mokapot, which is niche proteomics software, but now has 13 contributors!

Additionally, open-source code encourages rigorous testing. As others in the community work with the software, they often find edge cases that the original authors may have missed. This collaborative environment leads to better, more reliable tools, increasing the overall credibility of the research. It also allows others to replicate algorithms in their own code, and contribute back improvement that they may find.

Open-Source Propels Science Forward

And closed-source scientific software holds a field back.

When new algorithms are proposed for scientific fields, yet their implementation is locked behind proprietary barriers, it stifles innovation. Conversely, open-source software is the building block of progress. By allowing others to build on your work, you create opportunities for the field to evolve more rapidly. Programming languages like Python and R, and the many packages like PyTorch and Scikit-learn have become cornerstones of research, precisely because they’re open and free for all to use and improve.

Wait, so is propietary, closed-source, scientific software ever acceptable?

I think the answer is clearly yes—but not under the guise scientific research intended for public benefit. Rather, the process of incorporating new algorithms into propietary software is inherently a commercial activity. Even when built around open-source software, there is the opportunity to improve usability, scaling, and efficiency with commercial investment. Commercialized scientific software can be a huge boon to researches who are not primarily focused on the advancements of the computational methods and instead care most about their application.

Indeed, I think it follows that research funded by public grant mechanisms (like through the NIH), should always be open-source.1 Anything else would be contrary to the mission of publicly funded research.

You might find yourself thinking, "I don't want some company stealing my open-source software, making it part of their propietary code, and making money off of it." If that is you, I'd recommend researching common open-source software licenses. The strong copyleft licenses like the GPL-3.0 or the AGPL-3.0 licenses generally require any code derived from their source be open-source under a compatible license as well. Companies are generally risk-adverse on legal matters and respect these licenses.

Strategies to Commercialize Open-Source Software

While open-source software is freely available, that doesn’t mean that it lacks commercial potential. When commercializing open-source software, it's important to strike a balance—ensuring that core functionalities remain open and free for scientific exploration while building sustainable revenue models through value-added services or exclusive features.Here are just a few models for commercializing open-source projects without compromising the ethos of openness:

• Dual Licensing: Offer the software under an open-source license under a strong copyleft license, while providing a paid license with less strict terms for entities that may want to incorporate your software in a closed-source tool.

• Support and Services: Many companies, like Red Hat or RStudio, profit by offering support, custom development, or consulting for open-source tools.

• SaaS (Software as a Service): Offer a hosted version of the open-source software with added features, such as ease of use, maintenance, or cloud integration, similar to how GitLab or MongoDB monetize their platforms. This is also the route that open-source proteomics software like Sage is using to support development under Chaparral.

• Open core: Offer the core functionality of your software as fully open-source, but allow access to certain features only to paying customers. Often, this model uses a time delay for adding features to the open-source core, essentially allowing paying customers early access to new features. This model is used by projects like mkdocs-material.

These strategies allow developers to remain committed to open-source ideals while maintaining financial sustainability.

Open-source scientific software is not just a good idea; it’s a philosophy that embodies transparency, collaboration, and progress. As computational research becomes increasingly central to breakthroughs in every scientific field, open-source software must be part of the broader push for open science. Transparency, reproducibility, and collaboration are the cornerstones of scientific progress, and they rely on the free exchange of both data and code. Plus, it is way more fun to build in the open.

Welcome to Talus Bio’s Quarterly Update!

You’re receiving this newsletter because you expressed interest in staying informed about Talus Bio’s progress. We’ve had an exciting few months filled with groundbreaking results, new leadership, and significant progress towards our mission to bring forward a new wave transcription factor therapeutics.

Talus Bio is unlocking a new class of medicines — the 98% of cancer-associated transcription factors (TFs) long considered “undruggable.” Despite their potential, TF drugs have largely eluded the pharmaceutical industry, even though previous TF-drugs like Revlimid ($120B LTV) and enzalutamide ($50B LTV) have made a transformative impact.

We built a proprietary AI-driven discovery platform to identify and optimize TF modulators in their native cellular environment, overcoming the decades-old limitations that have hindered progress in this space.

$11.2M Financing Fuels Discovery

Talus Bio recently closed $11.2M in additional financing, led by Two Bear Capital and supported by a stellar group of new and existing investors including WRF Capital, NFX, YCombinator, BoxOne, and others. J. Seth Stratton, Managing Partner at Two Bear Capital will join the Talus Bio board in conjunction with this financing.

This capital empowers us to accelerate deployment of our AI-powered MARMOT platform, which integrates machine learning and next-gen proteomics to unlock drug targets previously considered “undruggable”. With these resources, we’re expanding our world-class team and broadening our pipeline of transcription factor therapeutics for new targets in oncology and beyond.

Gaelle Mercenne, PhD, Joins as Head of Biology to Drive Discovery

We’re excited to announce a key addition to the Talus Bio team — Dr. Gaelle Mercenne, a trailblazer at the intersection of AI and drug discovery. Coming from a leadership role at Sumitomo Pharma and previous experience as the first scientist hired at Recursion Pharmaceuticals, Gaelle will accelerate our previously “undruggable” transcription factor programs towards the clinic. Her deep expertise in pushing AI-driven discovery to unlock novel therapies makes her the perfect fit to lead the next phase of our development, expanding our preclinical efforts across high-impact targets.

Brachyury Inhibitor Shows First Activity in Lung Cancer

TAL-061 is a first-in-class inhibitor of the previously undraggable “Brachyury” transcription factor. Brachyury was identified over 50 years ago as the bona fide diver of a rare spinal cord sarcoma (chordoma). Earlier this year, TAL-061 was shown to block growth of patient-derived chordoma tumors grown in mice.

Recently, TAL-061 was tested in an in vivo mouse model of non-small cell lung cancer (NSCLC) with high brachyury expression (which occurs in 45% of all NSCLC patients). In this study, we observed robust single-agent activity for TAL-061, with stronger effects than standard-of-care first line therapy. This was the first demonstration of in vivo efficacy of a Brachyury inhibitor in NSCLC. Further studies will investigate Brachyury inhibition in combination with standard of care therapies in NSCLC.

These results represent a significant advancement, positioning TAL-061 as a potential breakthrough therapy for chordoma patients who have access to no FDA-approved drugs, as well as a new approach to address the most deadly form of cancer in NSCLC.

Subscribe to Alex's other writing

Subscribe now