Chapter 9 – Emerging Biophysics Techniques  431

It has, of course, had a significant impact on multiple fields, including the biophysics and

the physics of life:

i Data analysis and modeling: AI algorithms can process vast amounts of experi­

mental data, identify patterns, and extract valuable insights. In the physics of life,

AI is used to analyze complex datasets obtained from experiments such as protein

folding, molecular dynamics simulations, and gene expression studies. For example,

AlphaFold developed by Google DeepMind in 2018 and the improved version

AlphaFold2 released in 2020 (Jumper et al 2021) show potential for genuinely disrup­

tive impact in enabling the prediction of the 3D structure of proteins from sequence

data. Both versions utilize DL methods utilizing large-​scale supervised training

datasets from established experimental structural biology data, but AlphaFold2

incorporates an additional attention-​based neural network architecture called the

“transformer,” which allows for better modeling of longer-​range interactions within

proteins compared to the original version. The key potential is that it enables predic­

tion of protein structures, which are very challenging to achieve using existing experi­

mental techniques, such as membrane integrated proteins.

  AI models can also be developed to simulate and predict the behavior of complex

biological systems, aiding in understanding and predicting emergent biological phe­

nomena. AI can also be used to build accurate models for biophysical systems, such as

protein folding, molecular interactions, and drug design. An excellent recent example

is the application of AI methods to develop new, an entirely synthetic antibiotic (Liu

et al., 2023), which is effective against the Gram-​negative pathogen Acinetobacter

baumannii, a superbug prevalent in many hospitals that exhibits multidrug resistance.

Here, researchers first screened 7684 small molecules, consisting of 2341 off-​patent

drugs and 5343 synthetic compounds, to look for evidence of growth inhibition of

A. baumannii in vitro. These were used to train a neural network, which was then used

on a chemical library of 6680 molecules selected as proof-​of-​concept for its struc­

tural diversity and favorable antibacterial activities, to predict potential undiscovered

candidates that would inhibit A. baumannii growth. Of these, 240 molecules showed

promisingly high scores and were then subject to laboratory testing, with nine “pri­

ority molecules” ultimately whittled down to a top candidate of abaucin, an entirely

new antibiotic with narrow-​spectrum activity against A. baumannii. A particularly

telling finding from this study was that once the training dataset was developed, the

in silico screening of 6680 molecules top reduce down to 240 likely candidates took

only ~90 minutes.

ii Bioimage analysis and pattern recognition: AI excels in image analysis and pattern

recognition tasks, which, as can be seen from the range of biophysics tools and

techniques discussed in this book, are crucial in studying life and the physics of life.

They can aid in big data tasks (i.e. those in which the data sets are too large/​complex

to be analysis by traditional data processing tools) such as recognition techniques to

identify and cells and track them in real time, or to detect specific tissues or specific

molecular structures in biological images, enabling the study of a range of biological

processes at far more granular levels than was possible previously.

AI also shows clear promise in enabling general data-​driven discovery by revealing pre­

viously hidden relationships from large-​scale datasets and helping to identify non-​intuitive

correlations and patterns that may lead to new discoveries in the physics of life. This may serve

to catalyze the generation of new hypotheses and new experimental designs and optimized

protocols. As with all applications of AI, however, several challenges remain. Regarding bio­

physics and physics of life research, two of these are key. One is ethical, in that allowing

non-​human intervention in critical decision-​making concerning new drug designs and bio­

tech innovations could, for example, result in inventions which are potentially dangerous

for unforeseen reasons, and potentially difficult to control since the AI optimization process

is, at least at the time of writing, very largely a black box. The other is more nuanced but

important issue comes down to aspects of performance. Ultimately, the output accuracy of