The world is in the middle of an Industrial Revolution in case you haven’t realized yet. For Healthcare industry specifically, which has been so resistant in the past to technological advances, the emergence of Foundation Model or Generative AI provides a huge opportunity to leapfrog over how things have been done historically.
From a data-centric AI perspective, medical data, healthcare data often contains highly valuable information related to the causality of real-life events. For example, a clinical note for a patient with Nystagmus records the full thought-process, in natural language, i.e., text data, of a neuro-ophthalmologist explaining why the diagnosis was made, what conditions the doctor has considered and ruled out, what the doctor believes the problem is, and outline their general assessment of the situation. At the same time, most likely a lot of correlated medical image data such Magnetic Resonance Imaging (MRI), Computed Tomography (CT), Electronystagmography (ENG) or Videonystagmography (VNG), retinal images, eye motion patterns data are available for this particular patient ordered by the doctor to test for his or her hypothesis. These are no ordinary data, these are the new “gold”. For a professional who works in AI, Machine Learning space, it’s a perfect setting to train Gen AI to learn a high-level general representation (latent embedding), that we can denote as “medical coding”, of a particular disease, a patient, a drug, a medicine, a treatment plan, so on and so forth. Downstream applications of these coding will be discussed later in this article in more detail.
Pre-training and Representation of Data
Pre-training methods have revolutionized the field of Natural Language Processing (NLP) over the past few years. The whole concept of pre-training leverages the fact that there is a much broader source of supervision in the raw data (text, image, video, medical note, etc.), and it’s generated at web-scale everyday! Most of these raw data are pre-processed by human brains such as language, or a clinical note. This is a much more scalable way in terms of training data and learning.
Pre-training itself can be task-agnostic and frequently used in Gen AI enabling models with an impressive capacity to understand and generate human-like content, chat-bot conversation, AI generated images, videos, or even 3D models.
A deep neural network (DNN), often with a transformer-based architecture, is used to learn a general understanding or representation of the data. Sequence-to-sequence modeling, sequence generation are the typical use case in Chat GPT which establishes the baseline methodology as shown below.
Typical Sequence-to-sequence or Sequence Transduction Architecture.
Further development of the methodology involve pairing the text data with associated images such as CLIP, DALL-E [1,2], 3D models such as ShapE [3], acoustics waves, etc. Gen AI has achieved great performance to learn the general latent embedding of data, therefore generate creative images, 3D models with controllability through techniques of Prompt Engineering, Fine-tuning to guide the generated output.
Generation of other formats of data including images, 3D models. Pairing text data with images and 3D models.
Perfect Use-case of Generating Compound-recipe
“Recipe-style” generation can be one of the best use cases for Gen AI. Here the recipe of a compound can be for a new drug [4], vaccine[5], mRNA information molecule [6], material [7,8], protein [9], and so on and so forth.
The main challenge of this problem usually comes down to the fact that it’s extremely hard to reason, rationalize the causality or mapping between input of composition of ingredients and the corresponding output performance. There can be so many combinatoric possibilities for ingredients, and it is very difficult to predict the performance of each recipe due to the subtle interactions among ingredients, how each ingredient influence the test results, confounding factors, etc.
The other bottleneck is the lack of creativity of the generated recipes. Many of those conventional approaches start with a baseline “recipe” and then follow a “trial-and-error” style incrementally changing the compositions. This approach cannot capture latent representation of each ingredient therefore difficult to discover new recipes.
Typical Design of the Model
A typical Gen AI based model for receipt generations can have an architecture shown below, where a transformer-based encoder can be used to learn the representation of either a molecule, or a ingredient, etc, and a simulator or renderer can be used to generate associated test events, test results. It is commonly referred to as Transmitter for the encoder and simulator. A generative diffusion model then can be used to sample the latent space, and generate new recipes!
A baseline design of the recipe generation model where recipe of a particular compound such as a drug, vaccine, protein, mRNA can be generated with its associated test results, performances.
Drug Discoveries
One immediate thought of applying such recipe generation model is in pharmaceutical development, which is a very serial process. Most of the time it starts with some kind of initial concept or idea, and then test starts in Petri dishes, move on to preclinical testing. And if all of that looks good, the process finally moves off to human testing with several different phases of clinical trials. The whole process tends to follow a “trial-and-error” approach as mentioned above, which can be immensely expensive, cost billions of dollars to take, up to a decade to finish. However, in the end, most of these trials fail. It is estimated that 86% of drug candidates developed between 2000 and 2015 did not meet their stated endpoints. The conventional approach does not really discover anything new unfortunately. There are already startup companies, such as Insilico Medicine (Pharma AI), that has been investigating the possibilities of using Gen AI to discover new drugs. A most recent example of AI generated drug is shown below.
The discovery of compound 8g with superior potency against SIK2 compared with the reported inhibitor and demonstrated excellent selectivity. It is generated via AI-generative models (Chemistry42) https://arxiv.org/pdf/2101.09050.pdf
Vaccine Development
Vaccine development can be another use case for generative recipe modeling. The genetic sequence of the COVID-19 virus was first published in January 2020. It kicked off an international sprint to develop a vaccine to battle the virus. Generative AI models can assist in predicting potential vaccine antigens or epitopes. These models can analyze large datasets of viral or bacterial genetic sequences to identify regions that are likely to trigger an immune response. A collaborative effort among Baidu’s AI team, Oregon State University (StemiRNA Therapeutics) and the University of Rochester Medical Center developed an algorithm that is called “Linear Design” to encoding to a wider range of therapeutic proteins. For COVID-19 mRNA vaccine sequences, the algorithm achieved up to a 5-fold increase in stability (mRNA half-life), a 3-fold increase in protein expression levels (within 48 hours), and an incredible 128-fold increase in antibody response as shown below.
Overview of mRNA coding region design (LinearDesign) for two well-established objectives, stability and codon optimality, using SARS-CoV-2 Spike protein as an example. [6]
Opportunities in Automation of Clinical Operations
On the ever-evolving clinical trial side, the integration Gen AI can enhance clinical data management by providing automated data cleaning, data integrity checks, integrating data from multiple modalities. It can also streamline patient recruiting process by analyzing patient profiles, finding the best match for participants, selection of best trial site, patient journey mapping, best intervention timing, adverse events prediction, so on and so forth. Gen-AI driven assessments can also be used to determine the readiness of trial data and documents for regulatory submission such as Institutional Review Board (IRB) approval, reducing the risk of rejections and delays.
Barriers of Data
The sheer volume of data required for training generative AI models can be immense. Gathering high-quality, diverse, and representative datasets is a significant challenge. Ensuring that the training data is free from biases, inaccuracies, or noise is crucial for the model’s performance and ethical considerations. For healthcare data, in particular, which contains patient information, including medical histories, diagnoses, treatment plans, medications, laboratory results, and clinical notes, challenges include ensuring data privacy and security (compliance with regulations like HIPAA), data quality, interoperability, and overcoming data silos. Collaborations between healthcare institutions, data scientists, and AI researchers are essential for accessing and using medical data for AI model development while adhering to ethical and legal considerations. Furthermore, responsible AI practices must be followed to ensure that AI models trained on medical data are accurate, unbiased, and safe for clinical use.
Regulatory Consideration
Generative AI applications in clinical trials must adhere to regulatory guidelines set by authorities like the FDA and EMA. Compliance with regulatory requirements ensures that AI-generated results can be used confidently in submissions and decision-making.
Conclusion
A lot of the Gen AI needs manifested in Healthcare industry such as the need for accelerated cycles of drug discovery is not localized to the healthcare domain only. They can be applied to various other industries and domains such material science, new energy, etc. This concept highlights the broader significance of innovations and technologies developed in one field, which can often be adapted and utilized in others. Gen AI models are usually resource-intensive and require availability of high-quality training data and powerful computing resources. Data stacks, computing infrastructure technologies (NVIDIA’s GPU, Xilinx’s FPGA, etc.) both need to caught up with Generative AI enthusiasm. The combination of data, the hybrid of Cloud AI (super computer centers) and Edge AI (Remote patient monitoring devices), can open up new frontiers in breaking through long-standing bottlenecks. This convergence will revolutionize many industries, re-define what’s possible, ultimately reshape the future of technology and business.
1. Alec Radford et al., Learning Transferable Visual Models From Natural Language Supervision, http://proceedings.mlr.press/v139/radford21a/radford21a.pdf2. Aditya Ramesh, et al., Zero-Shot Text-to-Image Generation, arXiv:2102.12092, https://arxiv.org/abs/2102.120923. Heewoo Jun, Alex Nichol, Shap-E: Generating Conditional 3D Implicit Functions, https://arxiv.org/abs/2305.024634. Arnold, C. Inside the nascent industry of AI-designed drugs. Nat Med 29, 1292 — 1295 (2023). https://doi.org/10.1038/s41591-023-02361-05. Alka Bali et al., Role of artificial intelligence in fast-track drug discovery and vaccine development for COVID-19, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9069021/6. Zhang, H., Zhang, L., Lin, A. et al. Algorithm for optimized mRNA design improves stability and immunogenicity. Nature 621, 396–403 (2023). https://doi.org/10.1038/s41586-023-06127-z7. Pyzer-Knapp, E.O., Pitera, J.W., Staar, P.W.J. et al. Accelerating materials discovery using artificial intelligence, high performance computing and robotics. npj Comput Mater 8, 84 (2022). https://doi.org/10.1038/s41524-022-00765-z8. Addis Fuhr et al., Deep Generative Models for Materials Discovery and Machine Learning-Accelerated Innovation, https://www.frontiersin.org/articles/10.3389/fmats.2022.865270/full9. Andrew W Senior et al., Improved protein structure prediction using potentials from deep learning, https://pubmed.ncbi.nlm.nih.gov/31942072/
