Topos Bio: Drug Design for Biology in Motion

Intro

Today, most successful drug campaigns start with identifying the structure of the implicated protein and then designing drugs for that structure. This sequence → structure → function paradigm requires proteins that maintain a constant shape. However, roughly one-third of proteins – including the ones implicated in Alzheimer’s, Parkinson’s, and prostate cancer – are in constant motion. We believe that the key to solving these diseases is in understanding this motion.

Topos is building an AI-native drug discovery platform that takes a novel approach to modeling proteins focused on generating ensembles rather than single static outputs. Over the past year we’ve built out our first model, Topos-1, trained on our proprietary dataset of experimental and synthetic data. Today we’re excited to announce our initial set of results from Topos-1 technical report, and share for the first time how Topos is building the future of drug design.

A New Approach to Protein Modeling

AlphaFold-2 was undeniably a massive breakthrough for drug discovery. Instead of spending weeks (or months) on experiments to identify the folded structure of a protein based on its sequence, you could now get it nearly instantly. However, AlphaFold produces the same output as the experimental approach: a single, static structure for a given sequence.

However, living cells aren’t quite this simple. Proteins often shift shape. Roughly one-third of human proteins are referred to as intrinsically disordered proteins (IDPs) due to the vast number of shapes they can take on. IDPs are implicated in many major, currently incurable diseases, including neurodegenerative disorders such as Alzheimer’s and numerous aggressive cancers.

Topos is taking a fundamentally different approach to protein modeling by designing for the specific unique characteristics of these intrinsically disordered proteins. Our generative model is designed to learn distributions, rather than static structures, and generates a conformational ensemble of all of the shapes that the protein takes on.

Creating the Dataset

The training set required for a model to learn the movements of dynamic biology does not exist in any public dataset. In order to build our drug discovery platform we first had to generate a large proprietary dataset of both simulation data and experimental data specifically focused on IDPs.

Investing in building out experimental expertise has allowed us to use measurements from our lab to test and fine-tune the model, creating a flywheel accelerating us toward finding effective compounds for our targets.

Controlling Biology in Motion

We chose our initial neurodegenerative and oncology targets specifically because they are often considered “undruggable”, and we believe the hardest case forces the right ontology. If you can model and design against disorder, you build the primitives to unlock any target – static or dynamic.

The tools we are building at Topos have applications far beyond single proteins. Cells are molecular machines, and like all machines, their parts move. Modeling the full dynamics of this movement is essential for understanding how cells function in both health and disease. Topos is building the tools to understand, and ultimately solve, disease.