AI drug discovery HACARUS / Kenma Fujiwara CEO to challenge with high accuracy with a small amount of data | Venture tour

A venture that enhances its presence as a player in the pharmaceutical industry.We will visit the manager of the hot venture, talk about the start of the business, the desire for the business, and the future prospects.

System development with phenotype screening

―― We are developing AI (artificial intelligence) using “sparse modeling”. Please tell us about the whole picture of the business.

Originally, the company called Hacarus began with "weighing".At first, I was doing a light healthcare service that automated the nutritional record of the diet by weighing the ingredients with a smartphone.However, healthcare services for healthy people were difficult to monetize.Therefore, it is the flow that has been trying to enter the medical field of diagnosis and drug discovery.

However, if you look at the Hacarus website, I think it will be visible to companies that are developing business in the medical field.As a feeling, I want to do it with a single medical care, but the medical field is strictly restricted, and it takes time to proceed with the business.Currently, while doing medical care as a long -term initiative, we are developing services that utilize AI for the non -regulated field.

――In the medical field, we work on image diagnosis and drug discovery.

The first thing I did in the medical field was the development of image diagnosis using AI.Here, Dai Nippon Sumitomo Pharmaceutical AI, DS Pharmacoprota DS Pharma Animal Health and AI to assist the diagnosis of animal heart disease, and analyze Bayer drugs and MRI images with AI to assist hepatocytic cancer diagnosis.I've been developing an AI to do.

With the point of contact with the pharmaceutical company, the Hacarus DD, a drug discovery support AI platform, started development to see if you could enter the pharmaceutical business Honmaru.As an area that can develop the knowledge obtained by the development of image diagnosis AI, we first develop AI platforms used for phenotype (expression type) screening as specific products.

――What kind of technology is the sparse modeling used in HACARUS DD?

Currently, a variety of companies, from major pharmaceutical companies to venture companies, are working on AI drug discovery, but I think most are used there.

Compared to deep learning, sparse modeling has three main benefits.The first is that you can make high -performance AI with less data.Deep learning requires a lot of data, but sparse modeling can predict high accuracy with tens of minutes, hundreds of hundredths.Second, you can understand why you came to the conclusion.It is often referred to as "black box", but in deep learning, it is not clear why AI made some predictions, but why did that make such a decision.Using sparse modeling can enhance the interpretation of the conclusion, for example, in the case of drug discovery, you can present predictions, including some reasons why such toxicity can occur.The third is that it can be operated with high speed and low power consumption.It can be used with a small amount of data, but even if it is incorporated into a medical device, the power consumption can be reduced.

As the words sparse (meaning "sparse" and "sparse" in Japanese), sparse modeling is a technology that extracts characteristics from less data, and is not originally developed for AI.In the medical field, it has already been implemented as a compressed sensing of MRI, and has recently attracted attention as a technology that supported the successful black hole shooting.

The analysis time that took 15 to 40 minutes is 16 seconds

――In May this year, we announced that sparse modeling has succeeded in reducing the time required for screening in collaboration with Mitsubishi Tanabe Pharmaceutical.

Tanabe Mitsubishi Pharmaceutical was also working on the development of screening systems using deep learning, but when we talked to us, we were facing two issues.One is that although we have a system that can be evaluated to some extent, we could not move at a very realistic time and cost.The other is that we did not know what the compounds were affecting the experimental results.

The reason why it took a lot of time was that deep learning had to learn different AI models for each compound.The use of sparse modeling can reduce the time it takes to learn, and succeeded in reducing the analysis time that took 15 to 40 minutes per compound to 16 seconds in deep learning.

The problem that Tanabe Mitsubishi Pharmaceutical was facing is actually the problem of many pharmaceutical companies working on AI drug discovery, and there are one one after another.

―― Why is sparse modeling not spread like deep learning despite its excellent technology?Is there any barrier to use for AI?

There are two walls.First of all, there are not enough technicians because of the technology that is not very recognized.Deep learning is developing human resources, including universities, and the number of parameters is increasing, but some people who are doing AI drugs at pharmaceutical companies have heard "for the first time".We are trying to create a human resource development program in -house and clear the problem, but I think the problem is great.

The other is that it takes time to make it.Because it can be used with a small amount of data, it is necessary to drop the know -how of researchers into the algorithm.Deep learning is the mainstream of throwing data, but sparse modeling cannot be highly performed unless human know -how and wisdom are incorporated to some extent.We have thoroughly interviewed people who have domain knowledge, such as a reader for diagnosis, a drug discovery, and a researcher, and puts the tacit knowledge of why we made such a prediction into the algorithm.However, I think that there is a aspect that such troubles are a bottleneck.

There is a win depending on how you fight

――What kind of prospects do HACARUS DD draw in the future?

Now it is developed in the Fenotype screening area, but we plan to expand it to molecular design, bonding proteins and proteins, and dragli posts.There are many projects that have already been received, and multiple projects are running at the POC level.

Looking globally, I think that AI drug discovery has as many star players, and Japan is already delayed.We are currently at the stage of system development, but some overseas players are doing and creating compounds in collaboration with pharmaceutical companies, like bioventures.Some companies find their own compounds and have a large license agreement with a pharmaceutical company.

I think it will take time after IPO, but we want to proceed to the point where we make things and license them to pharmaceutical companies.I'm trying to catch up with overseas players, but we also have our own technology, and there are areas without big data, so if we target such places, we think that there is still a victory.I am.

――If you choose a way of fighting, there is a chance.

In fact, there are many rare diseases in the area where we speak, and there are many companies who want to do new drug discovery where there are few data.In such an area, I think there is a place where players like us come to life.There are many diseases among Japanese people, and even though it is a rare disease, there are some diseases that are quite data in Japan.In that respect, I think it makes sense to do it from Japan, and I think that there is such a way of fighting because our AI can connect a small amount of data to drug discovery.

(Listle, Yuki Maeda, is provided by Hacarus)

Answersnews editorial department reports pharmaceutical companies, Tanabe Mitsubishi Pharmaceutical, Dainippon Sumitomo Pharmaceutical

→ See other companies