Accelerating drug discovery

Systematic identification of new drug targets using Machine Learning

HotSpot Therapeutics Inc.

Early stage drug-discovery startup backed by top-tier biotech VCs including Atlas Venture, Sofinnova Partners, MRL Ventures, SR One, Tekla Capital Management. Approximately $100MM raised with Series B financing announced in April 2020.


HotSpot Therapeutics is targeting nature’s regulatory mechanisms to create new allosteric medicines that exhibit high precision and potency. The company leverages its proprietary SpotFinder™ technology, the first and only platform designed to identify and target “regulatory hotspots,” a unique family of pockets that sit remote from the active site on a protein and are used by nature to control protein function.

Using bespoke chemistry approaches, HotSpot is developing a pipeline of first-in-class small molecules for the treatment of autoimmune and rare diseases as well as cancer. The company has identified regulatory hotspots across multiple target classes of importance to the pharmaceutical industry. HotSpot’s lead compounds include the first and only allosteric inhibitors to target PKC-theta for autoimmune diseases and S6 kinase, a critical signaling node involved in the regulation of mitochondrial and metabolic function


As of April 2020, our Machine Learning algorithm helped to identify two lead targets that were drawn from our ranked shortlist of 900 targets, generated from a comprehensive scan of the entire space of human proteins.


A key bottleneck in drug discovery is the limited number of proteins that can be effectively targeted with drug molecules.  HotSpot’s addresses this problem by identifying natural regulatory mechanisms at work in proteins that can be targeted with drug candidates.

To uncover natural allosteric mechanisms systematically, we have developed a Machine Learning-based triangulation approach for HotSpot’s proprietary SpotFinder™ platform. Our search engine integrates evidence from three different perspectives, adding up weak signals to a clear fingerprint:

  • Amino acid sequences of proteins
  • 3D protein structure information
  • Mechanism descriptors from biomedical publications.

Connecting data points at different levels not only combines complementary third-party evidence, but also provides hooks to incorporate experience and judgement (soft evidence) of HotSpot’s scientific team. Using an iterative hybrid approach, SpotFinder’s PREDICT engine continuously improves based on expert feedback and experimental data.

What clients say

What made it successful?

Put your engine where your heart is.

If you want to reap the competitive advantage that accrues from the scale of industrialization, you need to build a truly mighty machine. Mighty not primarily in terms of sheer computing power, but in terms of brains: capturing as much as possible of the precious knowledge of your team in silico is where the magic lies. This is a never-ending effort, which only over time will turn into a self-perpetuating flywheel. At the beginning, you need to build trust and understanding in your organisation, and the right processes to make sure your algorithms can learn and grow.

Systematic search needs a different map.

Going after a few select opportunities is a different game than scanning an ocean, mile-by-mile. As such, it needs different instruments for navigation which ideally, can be understood by everybody, from scientist to Board members. To go from an overwhelming number of possible directions to a defined set of strategic options, we have found that overlaying the search space with information connected to strategic considerations (e.g. disease indications, potential partner companies, patient need etc.) leads to a map that is much more useful. If done well your map scales seamlessly from big picture to detail, aligning your company along the way.

Tooling matters, or: you can’t do this with Excel.

Machine Learning first and foremost means: taking data seriously. The most precious resource you have is the data you have generated yourself, i.e. structure-function insights, expert feedback on predictions, and soft data points encoding decades of your team’s experience. Using files scattered across scientist’s work benches is a remnant of the artisanal data age that is unfit for the systematic industrial generation and exploitation of data. At the same time, you do not need to become a software company, embarking on full-scale infrastructure development. There is a pragmatic middle ground to start with: a stable database back-end with a user interface built from existing components, with a few semi-manual steps as glue.