Weinan E on Biocomputing: Promoting the transformation of scientific research from a "small farmer workshop" to an "Android" model

The process of finding successful new drugs is the most difficult part of drug development, and the development of AI is breaking this dilemma.

Biology and computing are originally two independent and parallel development industries.
In recent years, more and more people have combined them to form a fusion word.
More and more companies have taken root in this field and jointly derived New industry.

Recently, at the "First China Conference on Biological Computing", Academician E Weinan from the Department of Mathematics and Institute of Applied Mathematics of Princeton University gave a speech entitled "Machine Learning and Scientific Computing", bringing his own understanding of biological computing .

Figure | Academician E Weinan (Source: Site)

The following is the transcript of the speech of Academician E Weinan, and Shenghui has made modifications without changing the original intention:

Before discussing biological computing, let me talk about scientific computing.
I have been using deep learning to analyze biological data since 2015.
At that time, I made a software called Define.
At that time, I saw the possibility.
Deep learning has brought us new opportunities from the perspective of data analysis and scientific models.

There are two major themes in scientific computing.
The first major theme is dealing with physical models.
The video just now emphasized mathematics, computers, and statistics, but missed the most important physics.
The main models of science are derived from physics, such as Newton's equations, aerodynamics, elasticity, electromagnetic field theory, quantum mechanics, etc.

The Schrodinger equation is the most basic equation of quantum mechanics, and our task is to solve this type of equation.
Many people say that nano and biological will bring new scientific models.
At least they have not seen them yet.
If there are new scientific models, they are derived from the original models.

Before the emergence of effective mathematical methods, the method for scientists to solve practical problems was to simplify the model.
Although physicists found the basic principles, they were actually used in a different way, and this is still the case today.

It was not until the 1950s that fundamental changes took place with electronic computers.
People like me developed a series of methods, such as difference methods, finite element methods, and spectral methods.
With these methods, for the first time mankind has realized the use of basic principles to solve practical problems, such as bridge design, building design, and aircraft design.
The impact on structural mechanics, aerospace, weather forecasting, oil exploration and exploitation is also great.

However, there are still many problems that have not been resolved, including drug design.
At present, biological design is a very empirical subject.
The result is that the three scenarios of the theoretical person, the experimenter, and the enterprise are very different.
Why does this happen? Kind of situation? The root cause is the disaster of dimensionality, that is, there are too many internal variables, and as the dimensionality increases, the complexity increases exponentially.

(Source: sdlcpartners.
com)

The second major theme is processing data.
There are many types of data, DNA, genomics, proteomics, these are the most concerned about here.

Another example is images.
We treat images as data.
There are three main tasks.
The first is imaging, which is inversion, which uses experimental instruments and data to invert the internal structure; the second is image processing, which is to remove image noise and process the image.
Segmentation, repair, etc.
; the third is image recognition, which means image recognition.
Image recognition is the problem of identifying cats and dogs in a bunch of images.

From my perspective, from the perspective of scientific calculations, the biggest problem is to combine physical and mathematical models, starting from the model, getting data, and getting more effective models from the data.

Let me give you an example.
If you are doing drug design, molecular dynamics is one of the most important tools, but from the basic principles of drug design, molecular dynamics is an indispensable tool, but the difficulty lies in the relationship between atoms.
Interaction.
Theoretically speaking, this interaction is realized by electrons and must follow the principles of quantum mechanics.

In 1985, there was an epoch-making work to calculate the interaction force between atoms through quantum mechanics, making molecular dynamics a reliable tool, but the effect was not good, and it could only handle a few hundred atoms.
This was an overly simple system.
According to the machine learning routine, data is provided from the quantum mechanical model, and on this basis, a new, more effective and equally reliable model is provided through machine learning, so that the new routine has been well realized.

In this way, machine learning, scientific computing, and high-performance computing are combined to bring about orders of magnitude change.

In addition to this, in addition to molecular dynamics, we have also developed a series of methods that are useful for everyone to make drugs, such as deep learning.

Finally, I would like to emphasize AI for science.

Applications include biopharmaceuticals, energy materials, and advanced manufacturing.

At the same time, it promotes the transformation of scientific research from the "small farmer workshop" model to the "Android" model.