Digitalisation and Machine Learning in Microbiology

Yonas
June 18, 2023

Microbiologist Prof. Adrian Egli from the University of Zurich presented a visionary talk on AI's potential to advance microbiology at the ECCMID2023 conference in Copenhagen. The audience was inspired to explore AI's transformative role in understanding microorganisms and their impact.

Share This Post

Key Takeaways

The era of AI has begun. ChatGPT’s rapid adoption showcases its success, while the medical sector faces challenges like data complexity and unique signals, hindering AI application.
Powerful machine learning systems like GPT-4 require substantial computational power and vast data with efficient learning strategies.
AI’s future in microbiological labs includes bimodal and multimodal AI, generative and autonomous AI agents, and AI integration with knowledge graphs. Addressing biases, privacy, and data security is vital for trustworthy AI.

Title and Copyright Information

Attention: all answers provided below are solely based on the content presented in this talk. Please be mindful of any potential biases.

What is AI in Microbiology?

AI in microbiology refers to the application of artificial intelligence (AI) techniques, such as machine learning and data analytics, to enhance research and understanding in the field of microbiology. AI can help analyze vast amounts of data from microbiomes, genomes, and metagenomes, enabling researchers to identify patterns, predict microbial behavior, and uncover new insights into microbial communities and their functions.

This technology has the potential to significantly accelerate research and discovery in microbiology, contributing to advancements in healthcare, agriculture, environmental sciences, and more.

Recent large language models like ChatGPT can aid in various aspects. For example:

Literature and data analysis: ChatGPT can quickly process and summarize vast amounts of scientific literature, helping researchers stay updated on the latest findings and trends in microbiology.
Data interpretation: ChatGPT can assist in interpreting complex datasets, making it easier for researchers to analyze and draw meaningful conclusions from microbiological data.
Assistance in experimental design: Researchers can use ChatGPT to generate hypotheses, suggest experimental setups, or brainstorm ideas for designing microbiology experiments.
Education and communication: ChatGPT can be employed as a tool for disseminating knowledge about microbiology to students, healthcare professionals, and the general public in a more accessible and engaging manner.

While ChatGPT cannot replace domain-specific expertise or conduct experiments itself, it can serve as a valuable tool in supporting microbiologists and other researchers by augmenting their capabilities and enhancing the overall research process.

Q&A

Q1: How successful has ChatGPT been in integrating into our daily lives?

A successful technology is one that truly disrupts the way we operate, significantly improving our work processes. It becomes highly valuable when it is affordable, user-friendly, and easily accessible. The time it takes to reach 100 million users is often used as a measure of success. Notably, it took 75 years for the phone to achieve this milestone, seven years for the World Wide Web, and only two months for ChatGPT. However, the ultimate winner in this race was the Pokemon Go game.

Q2: Why is the application of AI in medical sector not as fast as we wished?

The application of AI in medicine lags behind despite the tremendous transformations observed in language translation, speech recognition, and handling text and images, exemplified by technologies like ChatGPT. There are five significant challenges that need to be addressed for AI to advance in healthcare: the complexity and dimensionality of healthcare data, the identification of unique health-related signals, regulatory aspects, the need for expertise in AI, and the development of viable business cases. While AI has shown promise in research, its implementation in medical settings faces these hurdles.

Q3: What is the relevance of digitalization in the diagnostic workflow?

Digitalization is now relevant in the diagnostic workflow to improve efficiency and accuracy. With the increasing use of electronic ordering systems, test ordering is becoming more streamlined. Large language models like CHAT-GPT can play a significant role in pre-analytics, aiding in communication and understanding between clinical microbiologists and physicians. Additionally, artificial intelligence can be utilized in areas such as microscopy and plate reading, optimizing the entire diagnostic process from sample to culture identification and resistance testing. Emphasizing pre-analytics and post-analytics alongside analytics is crucial to ensure effective and actionable results are delivered in a timely manner, reducing the brain-to-brain time in the diagnostic workflow.

Q4: What are the key components and requirements for powerful machine learning systems like GPT-4?

Large language models like GPT-4 consist of billions or even trillions of parameters, representing the connections between neurons in the neural network. For instance, a modern state-of-the-art GPU, which is crucial for processing power, can commercially be bought for around $2,000 and provides an astounding 15 teraflops (10^12 calculations per second).

In addition to computational power, vast amounts of data are crucial for training and fine-tuning these language models effectively. Furthermore, efficient learning strategies are vital to manage the immense computational demands and optimize the performance of these models. Overall, the key question revolves around understanding the intricacies of building and training powerful machine learning systems, like GPT-4, that can process substantial data volumes and perform complex tasks efficiently.

Q5: Why is it only recently that we start to see breakthrough AI products like GPT-4?

Data storage capacity, computer system advancements, and powerful GPUs have a profound impact on developing large language models like ChatGPT. The massive availability of 175 zettabytes of data on the global data sphere and the significant increase in computational power provided by GPUs enable the training of language models with billions of parameters and tokens. For instance, ChatGPT with 175 billion parameters and 500 billion tokens required approximately 1,000 teraflops, a task that would take around 300 years on a regular laptop but can be accomplished within a year with about 10^4 GPUs. These advancements make it practical and efficient to train large language models, contributing to their remarkable capabilities in natural language processing and generating human-like text.

Q6: What are the key steps to take to make our labs AI ready?

The key steps to get our lab AI ready are as follows:

Understand your objectives: Clearly define your goals and the questions you want to address with AI. Consider the available data and develop a data strategy that aligns with your objectives. Understand where the data is stored, how it is governed, and the technology and processes available in your lab.
Implement data standards: Follow FAIR principles (Findable, Accessible, Interoperable, and Reusable) for your data and metadata. Use quality controls like the ISO standard 8000 and consider using tools like ontologies (e.g., SNOMED CT and LOINC) to structure your data.
Develop a local data infrastructure: Establish a data warehouse, such as a laboratory data warehouse, to store and manage your data efficiently. Demonstrate practical use cases to convince stakeholders, including hospital administrators, of the business opportunities of investing in IT infrastructure.
Invest in people and education: With AI rapidly advancing, invest in educating your lab staff about digitalization and artificial intelligence. Emphasize change management to help them embrace new technologies and overcome any fear or resistance to adopting AI in the lab.

Q7: Where are the most advanced area of AI in clinical microbiology at the moment?

The most advanced area of AI in clinical microbiology is total lab automation, particularly in automating the recognition of growth on culture plates. This technology enables the detection of colonies on plates and the accurate differentiation between positive and negative plates with high sensitivity and specificity. These advancements have significantly improved the efficiency and accuracy of diagnostic processes in clinical microbiology laboratories.

Another significant application of AI in this field is the prediction of antibiotic resistance profiles directly from MALDI-TOF data. Researchers have utilized machine learning models such as logistic regression, random forest, and neural networks to analyze MALDI-TOF spectra and resistance information. By training these models on large datasets comprising hundreds of thousands of spectra, they have achieved promising results in classifying bacterial susceptibility to antibiotics. However, it is essential to consider the biological context of the data and select appropriate training algorithms to enhance the accuracy of predictions.

While these advancements in AI have shown great promise, challenges remain, including the need for standardization in data and metadata, developing local data infrastructures, and investing in educating laboratory staff about digitalization and artificial intelligence. Nonetheless, with continued research and integration into clinical practice, AI is poised to revolutionize the field of clinical microbiology, improving diagnostics and patient care.

Q8: What will happen next in the world of AI and digitalization in microbiological labs?

In the world of AI and digitalization in microbiological labs, we can expect significant advancements. Currently, AI tools are mainly single-modal, handling either text or images, but the emergence of bimodal and multimodal AI tools will allow them to combine different types of information and make more precise predictions about diseases. Combining AI with calculation programs, vector databases, and other AI tools will lead to generative AI and autonomous AI agents that can solve complex problems on their own.

In the future, the Internet of AI will enable seamless communication between different AI tools, enhancing their capabilities to access and assess broad medical data sources. This will enable voice ordering, dynamic feedback through chatbots, summarizing patient history, and interactive lab results, providing better support and efficiency for healthcare professionals.

Furthermore, the combination of AI with knowledge graphs will improve contextual understanding, allowing AI systems to perform with better accuracy and relevance. Though it may still be a long journey, the potential for a generalist medical AI for microbiology and infectious diseases is promising, revolutionizing the way we diagnose and treat patients.

Q9: What are the problems we need to be aware of when applying AI in Microbiology?

The problems with AI, particularly large language models like ChatGPT, include:

hallucination, where they completely fabricate references due to lack of knowledge of their training data, which hasn’t been disclosed.
Validating certain answers is challenging.
Biases in data sources, with 25% of data coming from China and limited representation from other regions, need to be addressed to promote open information and science.
Privacy and data security are critical, as demonstrated by Italy’s ban on ChatGPT.
There is a strong tech company dependence, with large language models requiring hundreds of millions of dollars and 10,000 GPUs for training.
A potential fear is AI singularity, where AI reaches human-level capacity and development becomes unpredictable and difficult to comprehend.

Q10: What would you particularly like to say to younger generation of microbiologists?

I really wish everyone to see AI as a lasting force in our lives. I emphasize following essential points that warrant thoughtful consideration and open dialogue. Firstly, I encourage each individual to make AI a personal priority in 2023 by engaging in extensive reading, including books and articles, to gain a comprehensive understanding of its potential. Secondly, a crucial step is to develop a robust data strategy and concept for laboratories, hospitals, or relevant units, recognizing the importance of data incentives. Drawing inspiration from successful initiatives like the Swiss sepsis network, it becomes evident that ensuring fair data availability is critical, especially for technologies like MALDI-TOF. However, we must also acknowledge that implementing AI entails significant investments. While AI promises to enhance healthcare information organization and accessibility, achieving reproducibility remains of paramount importance. To achieve this, adhering to standardized documentation practices will instill greater trust in these tools. Lastly, it is worth noting that AI’s practical applications are already visible in culture plates, and the future promises even more innovative solutions.

Subscribe To Our Newsletter

Get updates and learn from the best

More To Explore

Case Studies

Quarterly Campaign Featuring Sequencing Library Pooling

Our budget sequencing services are highly popular with our clients. We provide quarterly sequencing campaigns at unmatched discounts, significantly reducing costs through pooling various sequencing libraries.

Admin

Holidays 2024

This is a summary of official holidays of 2024 in Denmark, the UK, and China. We use this calendar to count the number of workdays spent on a project.

Technical documents

Microbial Genome Sequencing

Sequencing a microbial genome has never been as easy and affordable as now. We provide a complete sequencing package with the basic bioinformatics analysis included. Our professional services allow you to be hustle-free and focus on your critical tasks.