For over a decade, we’ve heard about AI’s transformative potential in therapeutics (Tx), but where’s the evidence? Has it all been hype? Can AI still revolutionize Tx? Or is there something the AI experts aren’t telling us?
First, some facts. The process of Tx development—from early clinical trials to FDA approval—is excruciatingly slow (over 10 years), expensive (over $2B, especially for immunotherapy drugs), and comes with a low probability of success (5% in Immune-Oncology, and similarly poor across the board). The process is only getting worse. The return on investment (ROI) keeps declining, a phenomenon known as Eroom’s law—the reverse of Moore’s law in engineering. So, is AI just ineffective at solving this problem?
What isn’t discussed is that most AI-platform companies in Tx focus on therapeutic discovery, not development. Discovery is the early, exciting phase where a new drug target is identified or a new therapeutic modality is engineered. This phase is thrilling for scientists, researchers, biotechs, and pharma giants alike. It’s where the glory is—groundbreaking discoveries get the headlines.
In contrast, therapeutic development is seen as less glamorous, more operational, and laden with scrutiny and regulation. It’s the stuff that sends scientists running. But the grim reality remains: the same “Eroom’s law” that applies to the development of traditionally-discovered drug candidates also applies to all these newly AI-discovered candidates. And the result is that AIis not yet tackling the extremely complex problems of drug development.
The most significant AI contribution to drug discovery is likely DeepMind’s AlphaFold, a breakthrough that revolutionized our understanding of protein folding and protein-protein or protein-small molecule interactions in 3D space. This chemistry breakthrough could have a significant impact on therapeutic discovery, especially in small molecules, and it will continue to improve our discovery capabilities in the coming years.
But chemistry isn’t biology and to produce an effective drug, one must also understand biology. The study of living organisms is far more complex than modeling defined chemical interactions. Systems biology involves interactions on a cellular level between trillions of cells, with outcomes depending on a web of events that is orders of magnitude more complex than static biochemical entities like proteins, genes, or cells. Understanding systems that evolve over time is a higher-dimensional problem, without any shortcuts—revolutionizing this with AI will take longer and cost much more.
So, why is drug development so slow, expensive, and prone to failure? The problem lies in the data used to push discovered drugs into clinical trials. This data comes from preclinical experiments, mainly in mice (in vivo models), and from lab experiments with human tissue fragments (in vitro or ex vivo models). The issue is stark: we can cure every type of cancer—in mice. But this preclinical success translates poorly to humans. When drugs that work wonders in mice fail in human patients, it becomes clear: “The only method we know doesn’t work well, but it’s the only method we know, so we keep using it.”
This painful reality is well-known and recognized by all drug developers, big and small. Expensive Phase 2 and Phase 3 trials of “promising assets” fail repeatedly, and companies often don’t understand why, leading to a cycle of failure.
Are drug developers just stupid? No. They’re actually quite smart, individually. But the drug development process itself is flawed, and we’re still relying on data we don’t trust to make decisions that shape the biotech industry and, more importantly, affect hundreds of millions of patients worldwide.
So, can AI really help the drug development industry? And if so, how?
The key is building the right data foundation—sequencing a large number of clinical samples from patients across various indications, treated longitudinally. This foundational dataset provides insights others lack, but this is only Part A. Part B involves generating massive amounts of ex vivo, in vitro, and in vivo perturbation data from experiments with many drugs. The more we harmonize preclinical and clinical datasets for the same molecules, the better we can fill data gaps and allow AI to learn the true mapping from preclinical to clinical data, and vice versa.
A crucial technical point that’s often overlooked: We aren’t just predicting whether a specific target will modulate a cell type in a dish or a mouse model. We’re interested in predicting what will happen to the multi-organ system of human patients who receive the drug. Preclinical data provides the context for AI to make this prediction accurate.
Some companies believe AI can revolutionize drug development by focusing on the complexity of human biology. For example, efforts to map the entire immune system and build predictive AI models are underway, aiming to optimize clinical development outcomes. This approach focuses on understanding why patients respond differently to drugs, due to the complexity and heterogeneity of our immune systems. However, there are significant challenges:
- Mapping the entire immune system is seen as an insane task—so no one tried.
- Even if you could map the immune system, how do you predict patients’ responses to new drugs?
- Investors want an amazing platform, but they expect ROI in five years.
Addressing these challenges requires significant investment and commitment. For instance, building the right data foundation involves sequencing a vast number of clinical samples and generating massive amounts of perturbation data. Harmonizing preclinical and clinical datasets for the same molecules allows AI to learn the true mapping from preclinical to clinical data, and vice versa. The aim is not just predicting whether a target will work in a controlled environment, but understanding its impact on the immune system of human patients. This is where preclinical data provides context for AI to improve its predictions.
To build a successful biotech platform, it’s essential to resist shortcuts and focus on genuine development. Unfortunately, many companies focus on pushing one or two assets into clinical trials, which fall victim to the same “Eroom’s law” dynamics, leading to failures regardless of AI involvement. Leveraging advanced multi-omics sequencing technologies that continually evolve is critical to capturing the most comprehensive cellular information possible. Biotech companies must stay on the cutting edge, using the highest-resolution mapping available.
To advance the potential of AI in drug development, We are building an AI model of the immune system that can address the complex challenge of prioritizing preclinical features that can accurately predict clinical outcomes. Achieving this will require significant time and investment, but a thorough and innovative approach can help AI unlock deeper insights into immune system dynamics, serving as a powerful tool for more precise drug development. By overcoming these challenges, the biotech industry can break free from “Eroom’s law” constraints, paving the way for a future with more effective treatments and better patient outcomes.
About Noam Solomon
Noam Solomon is the co-founder and CEO of Immunai, a biotech startup using single-cell genomics and machine learning to discover and develop novel therapeutics that reprogram the immune system. Noam, who began his studies at Tel Aviv University at the age of 14 and earned his bachelor’s degree in computer science by 19, continued his academic journey earning two PhDs at MIT and Harvard. While at MIT, Noam met his co-founder, Luis Voloch, and launched Immunai in 2018. Under Noam’s leadership, Immunai has partnered with top pharmaceutical companies and academic institutions to improve clinical trial success rates and therapy effectiveness.