Cognaisent was co-founded in 2018 by Jonathan Fisher and Carla Emmons after Jonathan completed 3 years of independent research culminating in the invention of the GoFish machine learning technology. David Rudel, Hugh O’Reilly, and Ted Pine joined shortly thereafter to round out the founding team.
In 2018, the team completed a reinforcement learning prototype in the Atari framework showing great promise. In January 2019, Cognaisent received a grant from the Maine Technology Institute and then placed at the Dartmouth Entrepreneurs Forum pitch contest in March. In 2020 the team has been working on generalizing and optimizing the engine.
Cognaisent is incorporated in Delaware.
Patent application history United States non-provisional patent application Serial No. 15/658,104 entitled "Methods and systems for identifying patterns in data using delimited feature-regions" filed in the United States Patent and Trademark Office on July 24, 2017. Status: Allowed (Nov 2020).
Improve people’s lives by applying breakthrough advances in machine-learning technology.
There is a growing sense of urgency throughout the domain of machine learning to solve three problems which currently hobble state-of-the-art neural networks:
- 1. Neural networks require large amounts of data to learn useful functionality, which can be infeasible or expensive to obtain.
- 2. Neural networks are black boxes which cannot provide the explainability required for people to trust their results and extend them with human insights.
- 3. Neural networks are not robust, and can be easily fooled by deliberate deception, or even by incidental deviation from the space they were trained on.
Cognaisent has developed a fundamentally new technology which learns the way people do, allowing it to sidestep all three of those problems. The core engine is called GoFish, for “Generation of Features in Structured Hierarchies”.
Cognaisent is applying GoFish to the analysis of novel biological compounds, where acquiring data can be expensive in terms of both monetary costs and calendar time. In some cases, as with rare diseases, it may be simply impossible to generate enough data to produce reliable results with neural networks. Cognaisent developing tools to predict properties of molecules in silico before experiments are run, streamlining the laboratory process and allowing users to focus their efforts on the highest-value candidate molecules.
The current target MVP use-case is in antibody engineering, where the goal is to evaluate candidate antibodies for a target antigen of interest. This evaluation is based on properties central to druggability, such as binding affinity, chemical stability, immunogenicity, oligospecificity/polyspecificity, etc. The task lab-techs face is to measure these properties in a huge array of initial antibody candidates, and narrow them down to a small pool of candidates to be tested in vivo. This tool will enable lab-techs to circumvent expensive laboratory processes by predicting these properties in silico.
Laboratory data for antibodies’ performance relative to specific target antigens is typically quite limited. GoFish’s data-efficient learning is expected to far outstrip the state-of-the-art on such datasets, giving us a strong competitive advantage.
GoFish technology can be applied to a wide range of use-cases and industries where only small data sets exist or explainability is required. They have therefore built the core functionality as a back-end engine which can be applied across use cases easily and efficiently.
Development of therapeutic antibodies is a time- and resource-intensive funneling process. The vast majority of the candidate antibodies fall out of the funnel because they fail to meet one or more criteria in terms of binding affinity, chemical stability, non-immunogenicity, etc. By using machine learning to predict many of those properties earlier in the development process, this product will conserve resources while focusing the search on higher-value candidates, and at the same time dramatically reducing the time to market.
Each antigen is unique, and the process of optimizing antibody candidates to a target antigen is highly dependent on its individual properties. Therefore, it is essential that a machine-learning property-prediction tool be tuned to each particular target being pursued. The GoFish machine-learning technology has been proven to learn effectively from far less data than the existing state of the art, giving it the ability to make more accurate predictions earlier in the antibody development process — thereby narrowing the funnel at its widest point, with concomitant savings in resources and time-to-market.
- For the lab scientist: Improve early stage pipeline productivity.
- For the R&D Innovation manager: Reduce the high costs of testing in vitro and especially in vivo. Increase the probability of positive hits leading to viable therapeutic options early in the drug discovery pipeline by identifying optimal candidates as quickly as possible.
- For the VP of R&D: Shorten the time to clinic, reducing the cost of capital and freeing up resources across the organization.
The Cognaisent team has extensive experience across industries ranging from software and media to pharmaceuticals and edu-tech.
Jonathan Fisher, CEO/Inventor Jonathan is a data scientist and machine-learning researcher as well as the inventor of the GoFish technology. Since 2014, he has devoted his efforts full-time to Cognaisent and the research & development that resulted in GoFish. Jonathan has bachelor’s and master’s degrees in Mathematics from Dartmouth College.
Carla Emmons, COO Carla is a technology leader with expertise in business development, operations and expansion. She has 20+ years experience in technology and software development, especially in rapid-growth scenarios. She’s the project manager on the team as well as the main English-to-English translator. Carla has a bachelor’s degree from Dartmouth College.
Hugh O’Reilly, CFO Hugh is a finance executive with a track record managing companies from start-up through rapid growth. He closed Cognaisent’s friends and family funding round. Hugh has a bachelor’s degree from Dartmouth College and a JD from Vanderbilt University.
Ted Pine, CRO (Revenue) Ted has extensive experience building sales & marketing teams at life sciences analytics start-ups, most recently at genome sequencing analysis software developer, Parabricks, which was acquired by NVIDIA in December, 2019. Ted has a bachelor’s degree from Yale University.
David Rudel, Architect David is a data scientist who specializes in AI and mathematics systems and modeling. David has a bachelor’s degree from Harvey Mudd College and a master’s degree from Dartmouth College, both in mathematics.
The field of AI in Lead Discovery is quite crowded, but it is a wide-open market where no competitor has attained a market share greater than 20%. Of the dozens of active competitors, none appear to be differentiated in terms of technology, relying instead on standard deep learning libraries and frameworks combined with specialized areas of clinical focus.
The most well-known of the market entrants operate either as AI-driven biopharma drug developers (e.g, Atomwise, Exscientia, and Recursion) that both partner with larger firms in a shared risk/reward model and/or develop their own inhouse portfolio of candidates; while others such as BenevolentAI and InSilico operate on a fee-basis as consultants or clinical research outsourcers.
With either model, AI expertise, know-how, and tool-sets remain the province of the partnering firm or service provider.
Cognaisent seeks to disrupt this landscape by offering SaaS based tools that can be accessed by bench scientists and integrated directly into existing drug discovery workflows. Under this model, Cognaisent AI is licensed on an annual, project, or per use model, and can be integrated in the data domain with other in-silico development tools.
|Goal Set||Goal||Target Date||Result|
|October 23, 2020||Validate prototype against data set formats and expectations. Demonstrate GoFish’s ability to learn actionable predictive ability on such data, for one or more of the properties necessary for successful candidate antibodies.||December 10, 2020||Complete - December 10, 2020|
|January 2, 2021||Build out the generalized framework to process images and achieve a successful pass of the ImageNet Large Scale Visual Recognition Challenge for object detection with a substantial and representative portion of categories.||March 31, 2021|