Making Molecules People Need — PostEra’s Journey to Y Combinator
PostEra — Medicinal chemistry as a service powered by machine learning
Just over six months ago, we were sitting in our office at University of Cambridge, wondering how to get our technology out to the world. We knew we had something. As computational drug discovery researchers, we had just published the first model to outperform trained human chemists in predicting the outcomes of chemical reactions.
A few months earlier, we had published an interpretable machine learning algorithm that achieved state-of-the-art performance in predicting active compounds against proteins implicated in neurological disease. We had validated both models with a large pharmaceutical company and showed they transferred well to their in-house data. We had also shown these models weren’t just “black boxes,” and that they could learn when to be uncertain about their own predictions.
Now all we needed was a way to get the technology out there for use on important, highly confidential real world projects, not just select academic collaborations.
Coincidentally, while still debating strategy at our desks, an email came from a partner at an American group called Y Combinator. He asked if we had ever considered applying to their program. We hadn’t really, some of us had never even heard of YC (for those of you, ignorant like some of us were, it’s a program for startups that you should definitely consider applying to).
But six months later, we’re halfway around the world as a new startup, PostEra, and we’ve found how we can best contribute to improving the long and costly drug discovery process.
What we do
We make molecules, and we teach computers how to make molecules. Specifically, we make small molecules for therapeutic development using machine learning.
Unlike most other companies in this space, we are just as concerned with how to make the molecules as we are with what molecules to make. It turns out that many “molecular designs” fail because they are either very hard to make or perhaps even unmakeable. If it can’t be made and tested, simply designing a molecule is useless.
Why we focus on synthesis
Overall, synthesizing (making) compounds is still a rate limiting step of drug discovery and often the most expensive part of preclinical R&D .
If you want to make a new anti-viral to hit a promising SARS-CoV-2 target, you need to make it — often by outsourcing to China (pharmaceutical companies are realizing the need for supply-chain diversification given recent events). Any modification to improve toxicity of an anti-cancer drug has to be synthetically feasible. Changing a single atom on a potential cardiovascular drug to improve metabolism has to be done in a lab. By the end of a R&D project, millions are spent on synthesis, which is needed at pretty much every step of the process. Good computational design still helps a lot, but it is too often limited to the steps of development where ample data is present.
Coming from London, a civil engineer friend once likened the problem to building the Shard, the tallest skyscraper in the UK. Going from the architect’s initial paper napkin sketch to the real thing took quite a few years and over a billion dollars. Along the way, the design was altered to something that could actually be built, such as controversially not having the facets meet at the top like a true shard.
In the end, far more time and money is spent on good engineering and construction than the initial design.
Our long journey studying synthesis
Before starting PostEra, we spent much of our time reading papers and writing code, as academic chemists and machine learning engineers. Our work resulted in many papers published in top journals, involving state-of-the-art deep learning techniques, and still a lot of chemistry.
Molecular Transformer: A Model for Uncertainty-Calibrated Chemical Reaction Prediction
Molecular Transformer unifies reaction prediction and retrosynthesis across pharma chemical space
Ligand biological activity predicted by cleaning positive and negative chemical correlations
(Email us if you can’t access! founders@postera.ai)
Synthesis is hard
The details of the science are mostly covered in those technical papers, but the crux of the problem is that making complex organic molecules is still hard and expensive — though we now have a way to help.
Complex organic molecules are made by joining together simpler chemical building blocks, but the building blocks often react in unexpected ways, and our understanding of this reactivity is incomplete. As humans, we are limited by memory and biases — often simply resorting to trial and error experimentation of the reactions we know well.
In contrast, machine learning algorithms can learn complex patterns of chemical reactivity from the millions of reactions published in patents. Furthermore, the computational power of modern-day cloud architectures allows us to quickly search through the myriad routes to a given molecule to find the optimal path. In effect, we have built a GPS for Chemistry that shows chemists the best directions to their desired chemical destination.
It’s similar to how computers can search through the combinatorially exploding number of configurations of a Go board to defeat humans. In fact, like professional Go players, we now examine our algorithms’ output to learn what we as humans are missing. As trained chemists, we are often surprised at the creativity of our technology in finding ways to a compound of interest. Our models have an understanding built on millions of patents and millions of purchasable building blocks, that would simply be impossible for our human brains to internalize.
Just as Lee Sedol sat amazed at the ability of his virtual Go opponent, we often sit surprised at our model’s clever solutions, which often defy the intuitions we painstakingly learned in organic chemistry classes years ago. We’re just glad someone, human or not, could think of it.
Where we’re heading
During the past few months, we’ve been lucky to have great mentors and colleagues help us start getting the technology into the hands of the people who want to work with us. And we are excited about working with many more biotech and pharmaceutical companies on advancing their initial hit compounds to the clinic by way of clever design of molecules that we know how to build.
Additionally, we’ve also started work on democratizing access to some of our tools that are useful for medicinal chemistry projects around the world. Small biotechs should have access to the same, innovative computational tools as the world’s largest pharma — and they should be easy to use.
Our soon-to-be-released cloud based platform at postera.ai, hopes to change the process of making complex molecules from one of pain, to a convenient, informative experience. If one can order a pizza with transparent pricing, communication, and delivery times — that experience should also be available for the amazing, complex molecules that may one day be life-saving therapeutics.
We definitely know we are at work on a hard problem that is central to a long and difficult drug development process. But we also know it’s worth it, and certainly worth all of the hours we sat reading papers and staring at code. Despite wearing more t-shirts now, every day, we all still wake up excited to keep working away, just the same, on this ever important problem.
Be the next great email
This journey all started with one email from a friendly person interested in our technology. If you are similarly interested in what we’re doing, please reach out. If you want to apply PostEra’s technology to your medicinal chemistry projects or join PostEra to help us work on this important problem, just email us at founders@postera.ai
