10.2.1 Automated Theorem Proving

Since the birth of computers, there has been a longstanding desire to make them reason at the level of human thought. This aspiration led to the emergence of automated theorem proving (ATP) systems in the early 1950s, which reached their peak during the 1960s (Russell & Norvig, Reference Russell and Norvig2016). These ATP systems are designed to assist, augment, or even independently carry out the process of proving mathematical theorems.

ATP systems operate by converting mathematical premises and proofs into a formal language, such as first-order logic or higher-order logic (Loveland, Reference Loveland2016), that computers can process. These languages provide precise syntax and semantics, allowing unambiguous representation of mathematical ideas. The core of ATP systems, inference engines, then applies logical rules to explore the space of possible proofs that can derive desired conclusions from these premises. Various strategies, such as resolution, term rewriting, and model checking, can be employed at this stage. To navigate through this vast space, sophisticated algorithms, often involving heuristic search techniques like depth-first search, breadth-first search, and A* search, are used. Once a proof is found, ATP systems verify its correctness by checking each logical step against the rules of the formal system, ensuring that the proof is error-free and adheres to mathematical standards.

The focus of ATP systems is on proving the existing theorems. These systems not only ensure rigor but also handle the complexity of lengthy proofs that might be error-prone if done manually. For instance, the proof of the Kepler Conjecture, a centuries-old problem about the densest arrangement of spheres, was formally verified using the HOL Light proof assistant (Harrison, Reference Harrison2013). This collaboration exemplifies how AI can handle the intricate details of mathematical proofs, allowing human mathematicians to focus on higher-level insights and strategies.

10.2.2 Discovery of New Mathematical Conjectures

With the advancement of AI technology, AI can not only provide the proof of existing mathematical theorems, but also propose new conjectures that require validation by human researchers.

Mathematical conjectures are propositions or hypotheses based on limited evidence but have not yet been proven or disproven. They hold significant meanings as they can lead to profound discoveries and advancements in mathematical theory. Famous examples of conjectures include the Goldbach Conjecture, which asserts that every even integer greater than two is the sum of two prime numbers, a conjecture that remains unresolved. Another well-known conjecture is Fermat’s Last Theorem, which claims that there are no three positive integers $a$, $b$, and $c$ that satisfy the equation $a^n+b^n=c^n$ for any integer value of $n$ greater than two; this conjecture remained unproven for over 350 years until it was finally resolved in 1994 (Wiles, Reference Wiles1995). Such conjectures not only challenge mathematicians but also drive the development of new methods and theories in mathematics.

Conjectures often arise from patterns observed in data and serve as steppingstones for further exploration and proof within the mathematical community. This data-driven aspect is where modern AI technology comes into play. In light of this, DeepMind, a leading AI research lab, has developed an AI-based framework for discovering new mathematical conjectures (Davies et al., Reference Davies, Veličković, Buesing, Blackwell, Zheng, Tomašev and Kohli2021). This framework aims to identify relationships between two mathematical objects, such as whether two properties $X\left(z\right)$ and $Y\left(z\right)$ associated with a given variable $z$ satisfy a relationship $f$, formally expressed as $f\left(X\left(z\right)\right)=Y\left(z\right)$.

To achieve this, data samples $\left(X\left(z\right),Y\left(z\right)\right)$ are computed and collected. AI models are then trained with $X\left(z\right)$ as input and $Y\left(z\right)$ as output to learn the relationship $f$. When the learned relationship $\widehat{f}$ is more accurate than what would be expected by chance, it suggests that a valid relationship may exist, warranting further exploration. This framework has successfully been applied to discover new connections in algebraic and geometric structures of knots in Knot Theory and to propose new resolutions for long-standing open conjectures in Representation Theory.

In these cases, AI acts as a creative partner, generating promising conjectures that mathematicians can rigorously test and prove. This collaboration effectively blends human intuition with machine computational power, enhancing the research process.

10.2.3 Optimization of Solutions

The optimization capabilities of AI are another crucial aspect of its collaboration with mathematicians. Problems in number theory, combinatorics, and other fields often require extensive computation to test hypotheses or explore large solution spaces. AI algorithms, particularly those in evolutionary computation and reinforcement learning, excel at optimizing solutions and searching through vast, complex landscapes.

FunSearch (short for “searching in the function space”) exemplifies this kind of innovative approaches (Romera-Paredes et al., Reference Romera-Paredes, Barekatain, Novikov, Balog, Kumar, Dupont and Fawzi2024). Designed for combinatorics problems, FunSearch addresses challenging problems that are difficult to optimize but easy to evaluate. For example, the famous cap set problem involves finding the largest possible set of $n$-dimensional lattice points such that no three points are collinear. For a large $n$, brute force methods become impractical due to the exponential growth in the number of possible combinations. However, evaluating whether a given set satisfies the cap set constraint is straightforward by checking the rank of a matrix composed of these points.

To solve the cap set problem, mathematicians have proposed heuristics to decide how to add points to a cap set without violating the constraints. Different heuristics can lead to different solutions, and there is no consensus on what the optimal heuristic should be.

To efficiently explore this vast heuristic space, FunSearch leverages LLMs to generate code that calculates the priority of adding a point to the set. The effectiveness of different priority functions is then evaluated based on the size of the resulting set. Evolutionary algorithms are employed to select the best programs and feed back to the LLMs, generating increasingly improved priority functions. By iterating this process, FunSearch can discover superior solutions that outperform those found by human researchers.

The aforementioned collaboration targets solving problems with unknown optimal solutions. Human experts pinpoint the critical aspects of the problem (i.e., the heuristics), while AI takes on the arduous task of generating new heuristics. Guided by human expertise, AI uses evaluation metrics and evolutionary algorithms to refine its output continuously. This dynamic partnership produces scalable and innovative solutions, pushing the boundaries of what can be achieved in mathematical problem-solving.

10.2.4 Summary

In this section, we review how AI assists mathematicians across various problem types: (1) proving existing theorems with known conclusions, (2) proposing new conjectures, and (3) optimizing solutions for problems with unknown optimality. AI serves not only as a computational and verification tool but also as a means to visualize mathematical structures and explore high-dimensional data, thereby enhancing the mathematical intuition for researchers. It can also offer new perspectives and suggests unconventional approaches, inspiring mathematicians to think beyond traditional paradigms. Currently, human intervention is needed to establish rules, generate training data, and provide guidance. However, as AI continues to advance, future systems may autonomously generate and prove theorems, potentially revolutionizing the field.

10.3.1 Efficient Data Analysis and Interpretation

One of the primary ways AI assists scientists is through advanced data analysis and interpretation. Physical sciences often involve complex datasets that are challenging to analyze manually. AI algorithms excel at sifting through these large datasets to uncover hidden patterns and correlations.

The discovery of Higgs boson (Chatrchyan et al., Reference Chatrchyan, Khachatryan, Sirunyan, Tumasyan, Adam, Aguilo and Damiao2012) is a notable example of how AI played a crucial role in physics research. The Higgs boson, also referred to as the “God particle,” is a fundamental particle associated with the Higgs field, which gives other particles their mass. Studying particles like the Higgs boson requires colliding them at extremely high energies to reveal the fundamental components of matter and the forces that govern their interactions. By analyzing the particles produced in these high-energy collisions, physicists can test and validate theoretical models, explore the properties of elementary particles, and uncover new physics beyond the Standard Model.

One of the most significant tools for this research is the Large Hadron Collider (LHC), the world’s largest and most powerful particle accelerator, located at European Organization for Nuclear Research (CERN). Within seconds of a collision at LHC, data from millions of sensors are recorded. The data rate – over sixty terabytes per second – is too great to be entirely written to disk. To manage this challenge, a trigger system is employed to keep only the data events that are interesting enough, targeting specific configurations of particles consistent with a physics process of interest. At the high-level trigger stage, RNNs are used to predict the physical quantity of the particles based on the spatial and temporal signals collected from sensors. If the prediction matches the predefined criteria for events of interest, the data is deemed interesting enough to keep. As such, the overwhelming majority of events produced during the collision can be rejected.

Despite this filtering, analyzing the remaining data manually would still be an overwhelming task due to its sheer volume and complexity (Andreassen et al., Reference Andreassen, Komiske, Metodiev, Nachman and Thaler2020). Detecting the Higgs boson involved identifying a very specific set of collision events among the remaining immense amount of data. To achieve this, a special clustering algorithm, named particle flow, was designed to sift through the vast quantities of collision data and identify patterns and events that might indicate the presence of new particles or unusual phenomena.

Furthermore, GNNs have emerged as a powerful tool for further data interpretation (DeZoort et al., Reference DeZoort, Battaglia, Biscarat and Vlimant2023). Events recorded by the sensors of LHC can be naturally represented as graphs, where nodes correspond to individual particles or detector hits, and edges represent the spatial or temporal relationships between them. GNNs leverage this graph structure to effectively capture the complex dependencies and interactions inherent in collision data. By using GNNs, researchers can enhance the accuracy of particle tracking and identification, improve the resolution of reconstructed events, and distinguish between signals of interests and background noises with greater precision. For example, GNNs can be applied to track reconstruction tasks, where they help to connect discrete hits into coherent particle trajectories, even in densely populated environments.

As exemplified by the applications of AI at the LHC, its strong computational power and high efficiency enable the effective analysis and interpretation of large-scale experimental data. This allows researchers to focus on innovative tasks, such as discovering new particles, thereby driving scientific innovations and breakthroughs.

10.3.2 Optimized Experimental Design

In addition to the labor-intensive aspects of data analysis, AI can further significantly enhance other intelligent tasks, such as experimental design for scientific discoveries in physics. Take nuclear fusion as an example. Nuclear fusion is a process where two light atomic nuclei combine to form a heavier nucleus, releasing a tremendous amount of energy. It is the same reaction that powers the sun and stars, making it a potentially limitless and clean source of energy. Fusion offers significant advantages over nuclear fission, which involves splitting heavy atomic nuclei and is currently used in nuclear power plants. Unlike fission, fusion produces minimal radioactive waste, poses no risk of catastrophic meltdowns, and uses abundant fuel sources like hydrogen isotopes.

In the context of nuclear fusion, plasma – the fourth state of matter consisting of charged particles – plays a central role. Plasma must be confined and controlled within a fusion reactor to sustain the fusion reactions. It is crucial to control plasma shapes for optimizing the performance and stability of the fusion reaction. However, maintaining the high-temperature plasma within a tokamak – a device essential for confining plasma using magnetic fields – poses a significant challenge. Traditional methods involve controlling multiple time-varying, non-linear variables and precomputing a set of feedforward coil currents and voltages to manage plasma position, current, and shape. While effective, this method demands extensive design effort and expertise, especially when being applied to new plasma configurations.

Recent advancements have shown that AI, specifically deep reinforcement learning (RL), can revolutionize this process (Degrave et al., Reference Degrave, Felici, Buchli, Neunert, Tracey, Carpanese and Riedmiller2022). The RL-based approach uses an NN to represent the controlling policy, which is essentially a strategy for managing the magnetic coils in response to the state of the plasma. The training is guided by a reward system designed to maintain the desired plasma shapes, achieving stable confinement and optimizing performance. Through a process known as trial and error, the NN interacts with the simulation environment. It takes actions, observes results, receives rewards or penalties, and adjusts its policy to improve performance. Once the NN has been sufficiently trained by the simulator, its control policies are transferred to an actual tokamak. Experiments conducted on the actual tokamak showcased the efficacy of the RL controller in achieving accurate plasma control across a variety of configurations.

Such applications of AI not only optimize experimental design but also push the boundaries and accelerate the long journey of physical discoveries. As one plasma physicist noted, “AI would enable us to explore things that we wouldn’t explore otherwise, because we can take risks with this kind of control system we wouldn’t dare take otherwise.”

10.3.3 Summary

In summary, AI has woven itself into every stage of physical discovery, from the intricate design of experiments to the analysis of vast datasets, and the interpretation of results. Its remarkable advancements have transformed the landscape of scientific research, dramatically enhancing the precision and efficiency with which we handle and interpret complex data. Yet, despite these technological strides, AI cannot replace the ingenuity of physicists, particularly in the creative and intuitive process of hypothesis formulation. This stage remains deeply rooted in human insight and imagination, areas where AI still has limited capabilities. The synergy between AI and physicists not only accelerates scientific progress and uncovers new frontiers but also highlights the irreplaceable value of human expertise in steering and shaping research.

10.4.1 Automated Data Collection

Datasets are the cornerstone of chemistry research, providing the essential information needed to understand chemical reactions, properties, and behaviors. Comprehensive and accurate datasets enable chemists to identify patterns, validate hypotheses, and predict outcomes, driving scientific discovery and innovation.

Chemical data can originate from a variety of sources, including laboratory experiments and computational simulations. These data are ultimately recorded in the literature, such as scientific articles and reports. Collecting and curating the valuable information from dispersed sources is time-consuming and resource-intensive. Furthermore, human errors in manual data entry, measurement inconsistencies, and variations in experimental conditions can result in unreliable datasets.

AI attacks many of these challenges by automating the data collection process and enhancing the quality and comprehensiveness of datasets. One notable example is the use of NLP techniques to extract chemical synthesis procedures from millions of materials science papers (Kononova et al., Reference Kononova, Huo, He, Rong, Botari, Sun and Ceder2019). In this text-mining pipeline, paragraphs related to solid-state synthesis are first identified from a vast number of articles using a random forest classifier, which is trained on topics extracted by unsupervised learning. Once these relevant paragraphs are pinpointed, a bi-directional long-short term memory (BiLSTM) NN, designed for processing text sequences, is adopted to recognize the material entities within them. Next, the synthesis operations mentioned in these paragraphs are classified into six categories – such as mixing, heating, and drying – using NNs and sentence dependency tree analysis. Operation conditions like temperature, time, and atmosphere are then extracted by using a regular expression matching approach. Finally, all these components are synthesized into a coherent chemical formula.

This automated pipeline has generated a dataset containing about 20,000 synthesis records from over 4 million papers, a task that would take humans decades to accomplish manually. It, in turn, illustrates the transformative power of AI in streamlining data collection. Under the meticulous guidance of human-designed processes, AI can handle the labor-intensive work of information extraction efficiently and accurately.

10.4.2 Enhanced Predictive Modeling

Molecular prediction in chemistry refers to the process of forecasting the properties and behaviors of molecules before they are physically synthesized and tested. Accurate molecular prediction is essential for guiding experimental work, reducing the time and cost associated with chemical research, and facilitating the discovery of new compounds with desired properties. By predicting how molecules will behave, chemists can make informed decisions and design more effective and efficient experiments.

Chemists often rely on empirical rules and heuristic models for molecular prediction, drawing on accumulated knowledge and observed patterns from previous experiments. For instance, group contribution methods (Constantinou & Gani, Reference Constantinou and Gani1994) predict properties by summing the contributions of individual molecular fragments. While these methods are straightforward and quick to apply, they may lack the precision and flexibility needed for more complex molecules and reactions. To address this challenge, computational chemistry is introduced, employing computer algorithms to solve complex equations that describe molecular interactions and reactions. Techniques such as quantum mechanics, including density functional theory (DFT) (Vignale & Rasolt, Reference Vignale and Rasolt1987), allow chemists to calculate electronic structures and predict reactivity with high accuracy. Although these methods can be highly precise, they are often computationally expensive and time-consuming.

The advancement of AI technology has opened new avenues for accurate and efficient molecular predictions. Early studies utilized non-NN AI technologies to predict molecular properties based on manually crafted features. For example, a kriging method (Fletcher et al., Reference Fletcher, Davie and Popelier2014), a type of Gaussian process regression, is introduced to predict the electrostatic energies and polarization effects of aromatic amino acids. This approach involves training kriging models with geometries distorted via normal modes of vibration, which significantly reduces computational cost while maintaining accuracy. However, a major limitation of these non-NN AI methods lies in their reliance on the quality of manually extracted features, which requires extensive domain expertise.

To address this challenge, NNs, particularly GNNs, are now widely used to represent molecules more effectively. In GNNs, atoms are treated as nodes and chemical bonds as edges, enabling the training of models to predict various molecular properties without the need for manually crafted numerical descriptors. These models are applied to predict a range of molecular properties, including coordinates (Mansimov et al., Reference Mansimov, Mahmood, Kang and Cho2019), mass spectra (Park et al., Reference Park, Jo and Yoon2024), toxicity (Cremer et al., Reference Cremer, Medrano Sandonas, Tkatchenko, Clevert and De Fabritiis2023), etc. A compact review can be found in the literature (Wieder et al., Reference Wieder, Kohlbacher, Kuenemann, Garon, Ducrot, Seidel and Langer2020).

These AI-based prediction models reduce the reliance on trial-and-error methods, allowing chemists to focus on the most promising compounds. This accelerates the discovery of new materials, pharmaceuticals, and chemicals with specific functionalities.

10.4.3 Automated Synthesis and Experimentation

Building on the advancements in automated data collection and predictive modeling, AI-powered robots are transforming the way chemists approach experimental tasks. These robots are designed to free chemists from repetitive and laborious experimental activities across different scenarios, including materials handling, synthesis, and characterization (Coley et al., Reference Coley, Thomas, Lummiss, Jaworski, Breen, Schultz and Jensen2019, Reference Coley, Eyke and Jensen2020).

For example, solubility screening is a crucial step in understanding whether molecular compounds dissolve in a particular solvent. Traditionally, this process is time-consuming and labor-intensive, requiring periodic measurements throughout the experiment. Inspired by how human chemists visually assess solutions, CNN can be employed to automate this process (Pizzuto et al., Reference Pizzuto, De Berardinis, Longley, Fakhruldeen and Cooper2022). An autonomous robot takes photos of the samples of interest, and the CNN segments the sample from the images to determine whether it has fully dissolved in the solvent. This AI-driven approach not only automates the solubility screening process but also ensures consistent and accurate assessments.

AI algorithms are also used to optimize experimental configurations during autonomous experimentation. The outcome of chemical experiments depends on various variables, including time, temperature, and atmosphere. As the number of variables increases, the experimental complexity scales exponentially. To efficiently search for the optimal configuration, a regression model is often fitted based on historical experimental data (Burger et al., Reference Burger, Maffettone, Gusev, Aitchison, Bai, Wang and Cooper2020). This model predicts the portfolio of acquisition functions for different values of variables, narrowing the search space to configurations that are most likely to yield successful outcomes. This approach accelerates the identification of optimal experimental conditions, reducing the time and resources needed for experimentation.

Beyond individual tasks, integrated AI systems can manage entire materials discovery workflows. Traditionally, this process involves slow, laborious, iterative cycles of design, synthesis, testing, and analysis. An AI-driven robotic system can revolutionize this approach by starting with the generation of candidate materials through screening a vast array of possibilities for desirable properties. It then uses computer-aided synthesis planning to propose reaction pathways and finally synthesizes and characterizes the most promising candidates using a chemical robot assistant (Koscher et al., Reference Koscher, Canty, McDonald, Greenman, McGill, Bilodeau and Jensen2023). This integration streamlines the discovery process, enabling faster and more efficient development of new materials.

10.4.4 Summary

In this section, we briefly discuss the transformative trends in collaboration between humans and AI for chemical research, specifically in data collection, predictive modeling, and automated experimentation. The integration of AI enhances accuracy and efficiency, reduces the time and effort required by human researchers, and enables rapid synthesis and analysis of new materials.

Current AI technology still requires meticulous guidance from human experts to accommodate specific data formats, molecular mechanisms, and experimental specifications. However, the future of AI in chemistry is bright, with potential for even greater breakthroughs as AI technologies continue to evolve. As AI advances, its role in chemistry will undoubtedly expand, driving further innovation and enabling scientists to tackle increasingly complex challenges.

10.5.1 Sequencing and Analyzing Genetic Information

The advent of genomic sequencing has revolutionized our understanding of biology, unlocking the code of life written in DNA. However, the sheer volume and complexity of genetic data present significant challenges. Sequencing technologies, like next-generation sequencing (NGS), generate massive datasets that require sophisticated analysis to extract meaningful insights. This is where AI steps in, offering powerful tools to decode genetic information, identify genetic variants, and understand their implications for health and disease (Dias & Torkamani, Reference Dias and Torkamani2019).

One of the critical challenges in genomic sequencing is to ensure the accuracy of large-scale DNA sequence data. Errors can occur during the sequencing process, leading to incorrect base calls, which can skew downstream analyses. AI algorithms, particularly DL models, have been developed to improve the accuracy of sequencing reads. By training on large datasets of correctly sequenced DNA, these AI models can learn to distinguish between true genetic variants and sequencing errors, enhancing the reliability of the data produced.

For instance, DeepVariant, developed by Google, is a deep learning tool that has been widely recognized for its ability to call genetic variants with higher accuracy than traditional methods (Yun et al., Reference Yun, Li, Chang, Lin, Carroll and McLean2020). It uses CNNs, similar to those used in image recognition, to analyze raw sequencing data and identify variants, reducing the error rate and improving the quality of genomic data.

Once the raw DNA sequence is obtained, the next step is variant calling – identifying differences between the sequenced genome and a reference genome. These variants can range from single nucleotide polymorphisms (SNPs) to larger structural variations. AI plays a pivotal role in this process by automating the detection and classification of genetic variants.

AI-driven tools like GATK (Genome Analysis Toolkit) use machine learning algorithms to improve the accuracy and speed of variant calling (Lin et al., Reference Lin, Chang, Hsu, Hung, Chien, Hwu and Lee2022). These tools can handle a vast amount of data generated by NGS and identify variants with high precision. Moreover, AI algorithms can prioritize variants based on their potential impact on gene function, guiding researchers toward the most biologically relevant differences.

Beyond variant calling, AI is also essential in interpreting the functional implications of these genetic variants. Machine learning models can analyze vast databases of known gene functions, protein structures, and clinical data to predict how a particular variant might affect an individual’s health. For example, AI tools can predict whether a variant is likely to be pathogenic, helping clinicians make informed decisions about diagnosis and treatment.

A groundbreaking example is the collaboration between the Broad Institute and Google, where AI was used to improve the accuracy and efficiency of genomic data processing (Genome Analysis Toolkit). The partnership led to the development of deep learning models that could analyze terabytes of sequencing data in a fraction of the time required by traditional methods, paving the way for faster and more accurate genomic studies.

Additionally, AI has been instrumental in large-scale projects such as the UK Biobank (Bycroft et al., Reference Bycroft, Freeman, Petkova, Band, Elliott, Sharp and Marchini2018), where AI-driven analyses are helping researchers understand the genetic determinants of health and disease in a population of over 500,000 participants. The insights gained from such projects are expected to lead to new diagnostics, treatments, and preventive strategies tailored to individual genetic profiles.

10.5.2 Drug Discovery and Development

The process of drug discovery and development has traditionally been a lengthy, expensive, and complex journey, often taking over a decade and costing billions of dollars to bring a new drug to market. This journey involves identifying potential drug candidates, testing their safety and efficacy, and navigating through a rigorous regulatory approval process. The challenges of this process are compounded by the high rate of failure, with many promising compounds falling short during clinical trials. However, the advent of AI has begun to revolutionize this field, offering new tools and approaches that significantly enhance the efficiency and effectiveness of drug discovery (Mak et al., 2023).

One of the first steps in drug discovery is to identify a biological target – typically a protein or gene associated with a disease – that a drug can modulate. AI technologies have proven to be invaluable in this phase (Pun et al., Reference Pun, Ozerov and Zhavoronkov2023). By analyzing large datasets from genomics, proteomics, and other omics technologies, AI can identify potential drug targets that may not be obvious through traditional methods. For example, AI algorithms can sift through a vast amount of genetic data to pinpoint mutations or expressions linked to specific diseases, helping researchers identify new targets for drug development.

Once a target has been identified, the next step is to find chemical compounds that can interact with it effectively. Traditionally, this involved screening large libraries of compounds in the lab – a time-consuming and costly process (Han et al., Reference Han, Yoon, Kim, Lee and Lee2023). AI has dramatically accelerated this process through virtual screening, where machine learning models predict how different compounds will interact with the target. These models can rapidly filter out compounds that are unlikely to be effective, allowing scientists to focus on the most promising candidates.

In addition to screening, AI is also used in lead optimization, where the chemical properties of a compound are refined to improve its efficacy, reduce toxicity, and enhance its pharmacokinetic properties. AI-driven models can predict how modifications to a compound’s structure might impact its behavior in the body, helping scientists design more effective drugs (Vora et al., Reference Vora, Gholap, Jetha, Thakur, Solanki and Chavda2023). For instance, companies like Insilico Medicine and Atomwise use AI to predict the biological activity of molecules, optimizing them for better performance before they even reach the laboratory.

The success of a drug in clinical trials is often the make-or-break point of drug development (Urbina et al., Reference Urbina, Lentzos, Invernizzi and Ekins2022). AI is increasingly being used to improve the design and execution of these trials. Predictive modeling, powered by AI, can analyze patient data to identify which populations are most likely to respond to a new treatment, enabling more targeted and efficient trials. This approach not only increases the likelihood of success but also reduces costs by minimizing the number of participants needed and shortening the trial duration.

Moreover, AI can monitor and analyze data in real-time during clinical trials, identifying potential issues or trends that might otherwise go unnoticed. This capability allows for adaptive trial designs, where protocols can be adjusted based on interim results, further increasing the chances of success.

A striking example of AI for drug discovery is the development of IBM Watson for Drug Discovery, a platform that uses AI to analyze a vast amount of scientific literature and data (Visan & Negut, Reference Visan and Negut2024). In one notable instance, Watson helped researchers at Barrow Neurological Institute identify five new genes linked to amyotrophic lateral sclerosis (ALS), a breakthrough that could lead to new treatment avenues. Similarly, the British AI company BenevolentAI used its technology to identify an existing drug, baricitinib, as a potential treatment for COVID-19, which subsequently received emergency use authorization during the pandemic.

10.5.3 Summary

The integration of AI into life sciences has revolutionized the way research is conducted, enhancing both the speed and accuracy of scientific discoveries. By integrating AI into the research process, scientists can enhance their ability to uncover new insights, accelerate discoveries, and address questions that were previously too complex to tackle. AI not only augments human expertise but also opens up new avenues for exploring the intricacies of life, making it an invaluable partner in the pursuit of scientific discovery in the life sciences.

10.6.1 Impacts of AI on Scientific Discovery

Scientific discovery generally unfolds in three key steps: (i) hypothesis generation and selection, (ii) the design of methods to validate or disprove the hypothesis, and (iii) the derivation of insights from interpreting the results. AI influences each of these components to varying degrees.

The generation of hypotheses is a critical first step in scientific enquiry, traditionally driven by human intuition, experience, and creativity. This intuition is often rooted in a deep understanding of a specific scientific domain and is sparked by human intellect. AI, with its unparalleled ability to process vast amounts of information, can be harnessed to explore extensive hypothesis spaces, pinpointing those that align with existing knowledge. However, it is widely recognized that while AI excels at synthesizing and organizing existing knowledge, it still struggles to generate truly novel insights.

Once a hypothesis is formed, the next step involves developing methods to test it. This task typically falls to human experts who specialize in the relevant scientific field. AI systems can enhance this process by optimizing experimental designs, simulating various scenarios, and reducing reliance on trial and error. This not only minimizes resource use but also accelerates the path to discovery.

The final stage of scientific discovery involves interpreting results and drawing meaningful insights. This is where AI truly shines in its capacity to validate or disprove hypotheses. AI excels at processing and analyzing large datasets, identifying trends, anomalies, and subtle patterns that might elude human analysts. For example, AI can analyze complex genetic data to uncover interactions that contribute to specific traits or diseases, offering insights that could lead to new therapeutic approaches.

10.6.2 Human Guidance to AI for Scientific Discovery

While AI is increasingly embedded in every facet of scientific research, significantly reducing human labor, the success of AI in this domain still hinges on crucial human guidance. For AI to reach its full potential in scientific discovery, expert human intervention is indispensable. This guidance is vital for navigating the complexities of three key areas: (i) acquiring and preparing datasets to train AI models, (ii) designing AI models that are precisely tailored to the chosen datasets, and (iii) developing effective training schemes that enable AI models to learn meaningful patterns from the data. Each of these elements demands a deep understanding of both the scientific domain and the unique capabilities of AI technology.

10.6.3 Ethical Considerations in Human–AI Collaboration for Scientific Discovery

Human–AI collaboration in scientific discovery holds immense potential, but it also raises significant ethical concerns that must be addressed to ensure that this partnership benefits society without unintended harm. One of the most pressing issues is the potential for bias in AI systems. AI models are trained on existing data and, if that data reflects historical biases – whether related to gender, race, or other factors – the AI could perpetuate and even amplify these biases in scientific research. This could lead to skewed results, misrepresentation of certain populations, or biased scientific conclusions, ultimately affecting the fairness and credibility of scientific discoveries.

Another ethical concern is the transparency and interpretability of AI-driven research. AI systems, particularly complex models like deep neural networks, often operate as “black boxes,” making decisions that are difficult for even experts to fully understand or explain. This lack of transparency can pose significant risks in scientific discovery, where understanding the rationale behind a finding is crucial for its validation and acceptance. If scientists rely too heavily on AI without demanding interpretability, there is a danger of accepting results that cannot be independently verified, leading to a potential erosion of trust in scientific outcomes.

Moreover, the power dynamics in human–AI collaboration also raises ethical questions. As AI systems become more integrated into scientific research, there is a risk that the role of human scientists may be diminished or that decision-making power may shift disproportionately toward those who control the AI systems. This could create inequalities within the scientific community, where certain groups or individuals have more influence over research outcomes due to their access to or control over AI technologies. Ensuring equitable access to AI tools and maintaining a balance in human–AI collaboration is essential to prevent the concentration of power and to foster inclusive scientific progress.

The ethical implications of AI-driven discoveries also extend to the potential consequences of these discoveries themselves. AI has the capability to accelerate the pace of research, leading to breakthroughs that could have profound societal impacts – both positive and negative. For instance, AI could uncover new drugs or therapies at an unprecedented speed, but it could also be used to develop harmful technologies or exacerbate existing inequalities. The scientific community must therefore engage in ongoing ethical reflection and dialogue to consider the broader implications of their work and to ensure that AI-driven discoveries are aligned with societal values and contribute to the common good.

Finally, the integration of AI into scientific research raises questions about accountability. When AI systems are involved in generating hypotheses, designing experiments, or interpreting data, it becomes more challenging to determine who is responsible for the outcomes, particularly if those outcomes are harmful or erroneous. Establishing clear guidelines for accountability in human–AI collaborations is crucial to address this issue. This includes not only holding developers and users of AI systems accountable but also ensuring that there are mechanisms in place to correct or mitigate any negative impacts resulting from AI-driven research.

In conclusion, while the collaboration between humans and AI in scientific discovery holds great promise, it also presents a complex landscape of ethical challenges. Addressing these issues requires a proactive and multidisciplinary approach, involving ethicists, scientists, policymakers, and the public in ongoing discussions to guide the responsible development and use of AI in science. By doing so, we can harness the full potential of AI to advance knowledge while safeguarding the ethical foundations of scientific inquiry.