Richard L. Garwin (1928-2025): Force of Nature

There are physicists, and then there are physicists. There are engineers, and then there are engineers. There are government advisors, and then there are government advisors.

And then there’s Dick Garwin.

Richard L. Garwin, who his friends and colleagues called Dick, has died at the age of 97. He was a man whose soul imbibed technical brilliance and whose life threaded the narrow corridor between Promethean power and principled restraint. A scientist of prodigious intellect and unyielding moral seriousness, his career spanned the detonations of the Cold War and the dimming of the Enlightenment spirit in American public life. He was, without fanfare or affectation, the quintessential citizen‑scientist—at once a master of equations and a steward of consequence. When you needed objective scientific advice on virtually any technological or defense‑related question, you asked Dick Garwin, even when you did not like the advice. Especially when you did not like it. And yet he was described as “the most influential scientist you have never heard of”, legendary in the world of physics and national security but virtually unknown outside it.

He was born in Cleveland in 1928 to Jewish immigrants from Eastern Europe, and quickly distinguished himself as a student whose mind moved with the inexorable clarity of first principles. His father was an electronics technician and high school science teacher who moonlighted as a movie projectionist. As a young child Garwin was already taking things apart, with the promise of reassembling them. By the age of 21 he had earned his Ph.D. under Enrico Fermi, who—legend has it—once remarked that Garwin was the only true genius he had ever met. This was not idle flattery. After Fermi, Dick Garwin might be the closest thing we have had to a universal scientist who understood the applied workings of every branch of physics and technology. There was no system whose principles he did not comprehend, whether mechanical, electrical or thermodynamic, no machine that he could not fix, no calculation that fazed him. Just two years after getting his Ph.D., Garwin would design the first working hydrogen bomb, a device of unprecedented and appalling potency, whose test, dubbed “Ivy Mike,” would usher in a new and even graver chapter of the nuclear age.

Yet Garwin was never intoxicated by power. Because he thought the creation of the hydrogen bomb was driven by the exigencies of the age, he did not regret his work, but neither did he revel in it, seeking instead to control its worst excesses. For decades his contribution remained unknown, and he never bothered to update the record, saying that he could either get something done or get credit for it, but not both. Unlike many of his contemporaries, who leapt eagerly into the vortex of Cold War prestige and developed the braggadocio to go along with it, Garwin remained a paradoxical figure: at once the architect of supreme destructive force and one of its most persistent critics. He understood—perhaps more than any other scientist of his generation—that technical brilliance without ethical deliberation is merely efficiency in the service of catastrophe. He understood that his main job was to shoot down boondoggles and billion‑dollar bad ideas which were a defense department specialty. As perhaps the best example of how he saw his main function as offering stone‑cold sober advice based on the best technical analysis, he along with Hans Bethe pointed out fatal flaws with the Johnson administration’s ABM system and the Reagan administration’s SDI (Strategic Defense Initiative, derisively called Star Wars) system. As a member of Nixon’s science advisory council, he killed the proposal for a supersonic jetliner after citing noise pollution and structural concerns. In return the vindictive Nixon killed the science advisory council.

Garwin would advise every president from Truman to Obama, Republican and Democrat alike, offering a rare and vanishing example of a mind that could not be bought, flattered, or ideologically corralled. George W. Bush awarded him the National Medal of Science; Obama the nation’s highest civilian honor, the Presidential Medal of Freedom. He advised Kennedy during the Cuban Missile Crisis, pushed for the Anti‑Ballistic Missile Treaty under Nixon, helped Obama deal with the Deep Horizon Oil Spill and warned, presciently, of the folly of weaponizing space. From submarines to space weapons, from missile defense to nuclear disarmament, from oil spills to solar power, there was no field in which Garwin was not an expert, often the preeminent expert. Long after his hydrogen bomb design became a grim historical footnote, his advocacy for arms control, transparency, and rational defense policy continued unabated—earnest, unfashionable, and, above all, indispensable. As the physicist Freeman Dyson who was one of his fellow technical experts put it, Garwin was a prime example of how one man with a backpack could beat an army of government bureaucrats, if only that man’s name was Dick Garwin.

But to speak only of his role in the nuclear age is to risk caricature. Garwin’s genius was dazzlingly catholic. His parallel career at IBM mirrored some of the great technological innovations of our modern age. IBM prized him so highly that they allowed him to spend a third of his time doing classified government work on their dime. And their astute hiring decision paid off in spades. Garwin’s fingerprints are found in the design of gravitational wave detectors, in the refinement of MRI technology, in the development of GPS and satellite reconnaissance systems. He contributed to the understanding of superconductivity, to the design of the compact disc and the touch screen, to cryptography and to high‑energy physics – he narrowly missed the Nobel Prize for an experiment he conducted in his 20s to demonstrate what’s called parity violation in beta decay. And he did all this with a kind of quiet, omnidirectional intensity that shunned limelight but advanced civilization in tangible ways.

Garwin authored over 500 scientific papers, many of them marked not just by precision but by restraint—proposing solutions rather than pronouncements, elevating fact over flourish. He felt it his public responsibility to put almost the entire corpus of his unclassified work online at the Garwin Archive – more than 1600 pages of presentations, talks and technical analysis covering almost every defense‑related topic. If the academy in recent decades has become a theater of self‑display, Garwin belonged to an older order: one in which brilliance did not demand amplification, and integrity required no branding. As he said in a documentary made about him a few years ago, the preservation of our democracy was his lodestar; nothing else mattered if that goal could not be achieved.

Garwin kept on working, traveling, advising, relentlessly, until almost his last day. When he was 88 I had the privilege of having an email exchange with him about a project I was working on. My one‑paragraph question was answered with four paragraphs less than an hour after my message; characteristic of Dick, he also pointed me to technical tools that could aid my query. He sent me a touching obituary he had written of his wife of more than sixty‑five years, Lois, who had just died and who had been his pillar in supporting him and raising their three children.

It is difficult not to view Garwin’s passing as something more than a personal loss. It is the departure of a type—perhaps the last of a type. A scientist who not only understood the arcana of physics, but who grasped the frailty of human institutions and the perils of unmediated power. A man who could, with equal ease, work out a thermonuclear cascade and dismantle a delusion in a Senate hearing room. That such a figure would exit the stage at a moment not just when the world sorely needs him but when the machinery of government not only disregards scientific counsel but actively spurns it – preferring instead the solipsistic comfort of invented facts fueled by ideological biases – renders his absence all the more acute. Garwin represented the moral and intellectual ballast that keeps a civilization grounded, and without which it begins to drift. There were a select few handful like him – Hans Bethe, Freeman Dyson, Sidney Drell. But he was the last. That we are now adrift makes his departure feel less like the close of a chapter than the dimming of the last light on the horizon. His life, to those who knew it, offers a retort to cynicism. In an age when expertise is increasingly sneered at, and when the distinction between persuasion and knowledge is recklessly blurred, Richard Garwin stood for the proposition that facts still matter, that thought must precede action, and that it is possible—however rarely—for brilliance to be in the service of restraint rather than abandon.

Preventing misuse of chemistry cloud labs by bad actors

Cloud labs promise to significantly accelerate synthesis and testing by lowering cost and technology barriers - for both good and bad actors. Potential strategies for preventing misuse are explored.

Automated chemical synthesis platforms, especially cloud-based labs which represent an exciting emerging frontier, represent a potentially revolutionary step forward in scientific research. They offer the potential to accelerate discovery in fields such as pharmaceuticals, materials science, and environmental chemistry by setting up and running experiments more intelligently. Most importantly, by abstracting away technical details like specific instrumentation and consumable types and dimensions, these cloud labs can truly democratize science by enabling scientists and students who lack capital and access to sophisticated equipment to run experiments remotely from their laptops.

However, as with all powerful technologies, these platforms also present significant risks. A particular concern is the misuse of dual-use chemicals—chemical precursors that can be used for both legitimate scientific purposes and nefarious activities, such as creating illegal drugs, explosives, or chemical weapons. The lowering of barriers to the knowledge and capabilities needed to synthesize a chemical compound using a cloud lab concomitantly comes with the increased risk of relatively unsophisticated actors, both state and non-state, using such a system to make dangerous chemical agents.

To mitigate these risks while enabling legitimate research, it is essential to implement a robust framework combining technological controls, human oversight, regulatory compliance, and ethical principles. On the technical side, there are two levels at which such control can be exerted - the supply chain level and the software layer level. Chemical vendors already have restrictions and controls in place for ordering legitimately hazardous or controlled substances.

However, since many chemical precursors are benign by themselves and because the sheer scale enabled by cloud labs can make it unrealistic for vendors to have a foolproof system of control, there need to be checks and balances at the level of the software level as well. A particularly attractive feature of these checks and balances is that they involve identifying compounds by their SMILES strings or a variety of other cheminformatics formats - the kinds of chemical recognition that are now standard in every chemical database or modeling software.

Cloud Labs in a Nutshell

A cloud lab is a software system for programmatically setting up and running experiments in a centralized or distributed remote facility, typically one that is heavily automated and enabled by robotic experimentation. Cloud labs are highly attractive for two reasons: first, because they can enable anyone without access to infrastructure to run experiments from their web browser at low cost; and second, because they can lower barriers to running experiments by only needing users to specify high-level experimental protocols while abstracting away low-level details like consumables (test tubes, vials etc.), specific instrumentation and other laboratory equipment. As such, cloud labs promise to revolutionize the practice of science by democratizing it.

The Problem of Dual-Use Chemicals

Dual-use chemicals are substances that have legitimate scientific or industrial applications but can also be used to create harmful products. For example, chemicals used in pharmaceutical research might also serve as precursors for illegal narcotics, while industrial compounds could be weaponized to produce explosives or toxins. In fact, except for obvious examples like sarin whose illegal use is well-known, the very fact that almost any chemical reagent can be dual-use illustrates the nature of the problem.

The key challenge in managing dual-use chemicals is balancing accessibility and security. Scientists require access to these chemicals for critical research, yet misuse by malicious actors could have devastating consequences and could also put them out of reach of legitimate applications. A comprehensive strategy is required to prevent misuse while ensuring that legitimate researchers can continue their work unhindered. Several such strategies can be imagined.

1. Tier-Based Access:

To address the dual-use dilemma, one of the most effective strategies is implementing tier-based access based on the risk level associated with a particular chemical and the qualifications of the user. Chemicals can be classified into different tiers based on their potential for misuse.

Low-Risk Chemicals: Basic laboratory reagents and chemicals with low misuse potential can be made accessible to a wide range of users, such as students or early-career researchers. These could include, for instance, common solvents like methylene chloride, hexane and acetone.

Moderate-Risk Chemicals: These substances may have some dual-use potential, but they are primarily used for industrial or pharmaceutical research. Access can be restricted to verified researchers who demonstrate legitimate scientific purposes. Examples of such compounds include ephedrine (which is regulated as Sudafed in pharmacies), toluene (a common lab solvent but also a precursor for TNT) and acetic anhydride (a common compound in the synthesis of drugs like aspirin but also used in heroin production).

High-Risk Chemicals: Chemicals with significant misuse potential, such as explosives or nerve gas precursors, should be accessible only to highly regulated users, such as defense contractors or government agencies, under strict oversight. Examples of these would be potassium chlorate (used as an oxidizer but also as an explosive additive), hydrogen cyanide (used as a reagent in synthesis but is very toxic) and chlorine gas (used in water purification and as a disinfectant but can be - and was - used as a chemical warfare agent).

This tiered system allows for granular control, ensuring that only qualified individuals with appropriate credentials can access high-risk chemicals. For instance, while a defense or aerospace contractor might have a legitimate use for potassium chlorate or red fuming nitric acid (RFNA), a pharmaceutical company or small startup asking for large quantities of the same substances in synthesis might raise red flags. At the software level, this granular control would translate into the right permissions model, something that modern software systems already do quite well.

2. Quantity Monitoring:

A key strategy for preventing the misuse of dual-use chemicals is implementing strict quantity limitations; a chemical that might be fine in small quantities would raise questions if purchased in larger quantities. For instance, using a few hundred milliliters of toluene in a chemical synthesis might be legitimate, but if a synthesis starts requiring several liters of the material, it would be reasonable for the software to trigger a request for information (RFI), asking the users to justify their purchase. The quantities of chemicals used would also need to be tracked across time, since bad actors would try to get around the limitations by adding up small quantities purchased over long periods (for instance, illegal drug users pursued this strategy by stockpiling small quantities of Sudafed over time until that loophole was closed). By restricting the amount of a chemical that can be synthesized at any given time, the platform can prevent users from producing large quantities of precursors that could be used for harmful purposes. Batch tracking is already a feature embedded in vendor databases, so that feature can be easily extended to cloud labs.

3. Automated Pathway Detection:

A particularly sophisticated strategy for managing dual-use chemicals involves the use of synthesis pathway screening. Platforms can deploy algorithms that analyze the proposed chemical reactions submitted by users. These algorithms would flag any reaction pathways that could lead to the synthesis of dangerous compounds, even if the individual chemicals requested are themselves benign.

Some of the recent advances in software for retrosynthesis and reaction prediction would be particularly helpful in this context. If a user provides a target compound structure, retrosynthesis algorithms can break the compound down into its individual precursors and corresponding reagents. If more than a certain percentage of the precursors or reagents are questionable, the user request would trigger a warning flag. The same applies, albeit in reverse, for forward reaction prediction. If a user submits a list of building blocks to make, the algorithm can predict the variety of target compounds that can be made. If more than a certain percentage, including byproducts, appear questionable, the system can trigger a warning.

4. Behavior Anomaly Detection:

Any system can be fooled with the right strategy, and one of the strategies for gaming a chemistry cloud lab would be to confuse the system by making suspicious requests rare or sophisticated enough. To further enhance security, automated chemical synthesis platforms can leverage user-specific data analytics to track user behavior.

For example, consider the case noted above: A researcher who typically works on pharmaceutical compounds suddenly requests a precursor associated with explosives. This shift in behavior could trigger an alert, prompting further investigation. Users who display unpredictable or risky behavioral patterns could be subject to tighter restrictions at the organizational level. Aggregated user behavior would also be very helpful in this case. For instance, a user who orders X quantities of chemical A and Y quantities of chemical B with a frequency of Z would be subject to more scrutiny, even if any one of these individual actions would not be a cause for concern. Ultimately, only holistic user and organizational profiles would provide the most useful predictors of potentially suspicious behavior, but that kind of profiling - essentially tracking behavior through space and time - is one that current data analytics systems can handle quite well.

In all these cases, one will naturally have to balance privacy with security. The system can be blinded to personal identifying information, identifying users only with cryptic labels. Serious violations would need the users’ organizations, not the cloud lab provider, to unblind their identity and take necessary action. But user restrictions can certainly be implemented through mutual agreement between the cloud lab provider and the organization as well.
By continuously monitoring user behavior, platforms can proactively detect misuse, even before harmful chemicals are synthesized.

5. Batch and supply chain tracking:

The feasibility of batch tracking has already been mentioned above, but it deserves some elaboration. To further enhance accountability, cloud labs should implement features that track both source and usage. This ensures that every chemical can be traced back to its origin, whether it was purchased from a supplier or synthesized within the cloud lab.

This transparency creates a clear chain of custody, ensuring that if a dual-use chemical is misused, the platform can quickly identify where it came from and who was responsible for synthesizing it. Real-time monitoring and sharing with regulatory and other concerned organizations would reduce incentives to cheat and close potential loopholes related to the sourcing inherent in complicated supply chains.

Balancing Creativity and Security

Managing the risks associated with dual-use chemicals on automated synthesis platforms is a complex but vital task. By implementing a multi-layered approach—tiered access control, real-time and aggregated behavior monitoring, and automated pathway detection —these platforms can allow scientists to innovate while minimizing the risk of misuse. This approach won’t make the system foolproof, but it will minimize risk by implementing multiple checkpoints. Other human-level controls can be combined with this kind of pure algorithmic control. These controls could include expert panel reviews and collaboration with regulatory agencies. These checks are important as well, but their use must be balanced with potential impacts on the ease of use and lowering of barriers which are the most attractive features of cloud labs.

As cloud-based labs continue to play an increasingly important role in scientific research, it is essential to ensure that these powerful tools are used responsibly. With the right safeguards in place, we can unlock their potential to do good science in fields like medicine, agriculture, materials science and energetic materials while preventing misuse by bad actors.

Minority Report Meets Drug Discovery: Intelligent Gestural Interfaces and the Future of Medicine

In 2002, Steven Spielberg’s Minority Report introduced one of the most iconic visions of the future: a world where data is accessed, manipulated, and visualized through an immersive, gestural interface. The scene where Tom Cruise’s character, Police Chief John Anderton, swiftly navigates vast amounts of visual information by simply swiping his hands through thin air is not just aesthetically captivating but also hints at the profound potential of such interfaces in real-world applications—particularly in fields as complex as drug discovery. Just like detective work involves combining and coordinating data from disparate sources such as GPS, real-time tracking, historical case studies, image recognition and witness reports, drug discovery involves integrating data from disparate sources like protein-ligand interactions, patent literature, genomics and clinical trials. Today, advancements in augmented reality (AR), virtual reality (VR), and high-performance computing (HPC) offer the tantalizing possibility of a similar interface revolutionizing the way scientists interact with multifactorial biology and chemistry datasets.

This post explores what a Minority Report-style interface for drug design would look like, how the seeds of such a system already exist in current technology, and the exciting potential this kind of interface holds for the future of drug discovery.

The Haptic, Gestural Future of Drug Discovery

Perhaps one of the most memorable aspects of Minority Report is the graceful, fluid way in which Tom Cruise’s character interacts with a futuristic interface using only his hands. With a series of quick, intuitive gestures, he navigates through complex data sets, zooming in on images, isolating key pieces of information, and piecing together the puzzle at the center of the plot. The thrill of this interface comes from its speed, accessibility, and above all, its elegance. Unlike the clunky, keyboard-and-mouse-driven systems we’re used to today, this interface allows data to be accessed and manipulated as effortlessly as waving a hand.

In drug discovery, such fluid navigation would be game-changing. As mentioned above, the modern scientist deals with a staggering amount of information: genomics data, chemical structures, protein-ligand interactions, toxicity reports, and clinical trial results, all coming from different sources. The ability to sweep through these datasets with a flick of the wrist, pulling in relevant data and discarding irrelevant noise in real-time, would make the process of drug design not only more efficient but more dynamic and creative. Imagine pulling together protein folding simulations, molecular docking results, and clinical trial metadata into a single, interactive, 3D workspace—all by making precise, intuitive hand movements like Tom Cruise.

The core of the Minority Report interface is its gestural and haptic nature, which would be crucial for translating such a UI into the realm of drug design. By introducing haptic feedback into the system—using vibrations or resistance in the air to simulate touch—a researcher could "feel" molecular structures, turning abstract chemical properties into tactile sensations. Imagine "grabbing" a molecule and feeling the properties of its surface—areas of hydrophobicity, polarity, or charge density—all while rotating the structure in mid-air with a flick of your wrist. Like an octopus sensing multiple inputs simultaneously, a researcher would be the active purveyor of a live datastream of multilayered data. This tactile feedback could become a new form of data visualization, where chemists and biologists no longer rely solely on charts and numbers but also on physical sensations to understand molecular behavior. The experience would translate to an entirely new dimension of interacting with molecular data and models, making it possible to “sense” molecular conformations in ways that are impossible with current 2D screens.

Such a haptic interface would also make the process more accessible. Students and new researchers in drug discovery would quickly learn how to navigate and manipulate datasets through a gestural UI. The muscle memory developed through these natural, human movements would make the learning curve less steep, transforming the learning experience into something more akin to a hands-on laboratory session rather than an abstract, numbers-on-a-screen challenge. Drug discovery and molecular design would be democratized.

Swiping Through Multifactorial Datasets

One of the most exciting possibilities of a Minority Report-style UI in drug discovery is its ability to merge multifactorial datasets, making complex biology and chemistry data "talk" to each other. In drug discovery, researchers deal with data from various domains, — genomics, proteomics, cheminformatics, clinical data etc. — each of which exists in its own silo; any researcher in the area would relate to the pain of integrating these very different databases, an endeavor that requires a significant amount of effort and specialized software. Currently, entire IT departments are employed to these ends. A futuristic UI could change that entirely.

Imagine a scientist swiping through an assay dataset with one hand, while simultaneously bringing in chemical structure data and purification data on stereoisomers with the other. Perhaps throw in a key blocking patent and gene expression data. These diverse datasets could then be overlaid in real time, with machine learning algorithms providing instant insights into correlations and potential drug candidates. For instance, one swipe could summon a heat map of gene expression related to a disease, while another flick could display how a particular small molecule binds to a target protein implicated in that disease. A few more gestures could allow the scientist to access historical drug trials and toxicity data as well as patent data, immediately seeing if any patterns emerge. The potential here is enormous: combining these multifactorial datasets in such a seamless, visual way would enable researchers to generate hypotheses on the fly, test molecular interactions in real-time, and identify the most promising drug candidates faster than ever before.

The Seeds Are Already Here: AR, VR, and High-Performance Computing

While this vision seems futuristic, the seeds of this interface already exist in today's technology. Augmented reality (AR) and virtual reality (VR) platforms are rapidly advancing, providing immersive environments that allow users to interact with data in three dimensions. AR devices like Microsoft's HoloLens and VR systems like the Oculus Rift already provide glimpses of what a 3D drug discovery workspace might look like. For example, AR could be used to visualize molecular structures in real space, allowing researchers to walk around a protein or zoom in on a ligand-binding site as if it were floating right in front of them.

At the same time, high-performance computing (HPC) is already pushing the limits of what we can do with drug discovery. Cloud-based platforms provide immense computing power that can process large datasets, while AI-driven software accelerates the pace of molecular docking simulations and virtual screening processes. Combining these technologies with a Minority Report-style interface could be the key to fully realizing the potential of this future workspace.

LLMs as Intelligent Assistants

While the immersive interface and tactile data manipulation are powerful, the addition of large language models (LLMs) brings an entirely new layer of intelligence to the equation. In this vision of drug discovery, LLMs would serve as intelligent research assistants, capable of understanding complex natural language queries and providing context-sensitive insights. Instead of manually pulling in data or running simulations, researchers could ask questions in natural language, and the LLM would retrieve relevant datasets, compute compound properties, run analyses, and even suggest possible next steps. Even if a researcher could summon up multiple datasets by swiping in an interactive display, they would still need an LLM to answer questions pertaining to cross-correlations between these datasets.

Imagine a researcher standing in front of an immersive display, surrounded by 3D visualizations of molecular structures and genomic data. With a simple voice command or text prompt, they could ask the LLM, “Which compounds have shown the most promise in targeting this specific binding site?” or “What genetic mutations are correlated with resistance to this drug?” or even fuzzier questions like “What is the probability that this compound would bind to the site and cause side effects?”. The LLM would then comb through millions of datasets, both existing and computed, and instantly provide answers, suggest hypotheses, or even propose new drug candidates based on historical data.

Moreover, LLMs could help interpret complex, multifactorial relationships between datasets. For example, if a researcher wanted to understand how a particular chemical compound might interact with a genetic mutation in cancer cells, they could ask the LLM to cross-reference all available data on drug resistance, molecular pathways, and previous clinical trials. The LLM could provide a detailed, synthesized response, saving the researcher countless hours of manual research and allowing them to focus on making creative, strategic decisions.

This kind of interaction would fundamentally change the way scientists approach drug discovery. No longer would they need to rely solely on their own ability to manually search for and interpret data. Instead, they could work in tandem with an intelligent, AI-driven system that helps them navigate the immense complexity of modern drug design. With the right interface, researchers could manipulate massive amounts of drug discovery data in real-time, powered by already existing HPC infrastructure.

Current challenges

While this vision of an all-in-one molecular design interface sounds promising, we would be remiss in not mentioning some familiar current challenges. Data is still highly siloed, even within organizations, and inter-organizational data sharing is still bound by significant legal, business and technological challenges. While AR and VR are now being democratized through increasingly cheap headsets and software, the experience is not as smooth as we would like, and bringing in disparate data sources into the user experience remains a problem. In the future, common API formats could become a game changer. Finally, LLMs still suffer from errors and hallucinations. Having a human in the loop would be imperative in overcoming these limitations, but there is little doubt that the sheer time-saving and consolidation they enable, along with the ability to query data in natural language, would make their use not just important but inevitable.

A Future of Instant, Integrated Data at Your Fingertips

The promise of a Minority Report-style interface for drug discovery lies in its ability to make data instantly accessible, integrated, and actionable. By swiping and gesturing in mid-air, scientists would no longer be constrained by traditional input methods, unlocking new levels of creativity and efficiency. This kind of interface would enable instant access to everything from raw molecular data to advanced machine-learning models predicting the efficacy of new drug candidates.

We can image a future where a drug designer could pull up decades of research on a specific disease, instantly overlay that with genomic data, and compare it with molecular screening results—all in a 3D, immersive environment. The heightened experience would make it possible to come up with radically new hypotheses about target engagement, efficacy and toxicity in short order. Collaboration would also reach new heights, as teams across the world interacted in the same virtual workspace, manipulating the same data sets in real time, regardless of their physical location. The interface would enable instant brainstorming, rapid hypothesis generation and testing, and seamless sharing of insights. The excitement surrounding such a future is palpable. By blending AR, VR, HPC, and LLMs, we can transform drug discovery into an immersive, highly interactive, and profoundly intuitive process.

Let the symphony start playing.

Why Technology (and AI) Won’t Save Biology

There seems to be no end to biology's explosive progress. Genomes can now be read, edited, and rewritten with unprecedented scope. Individual neurons can now be studied in both space and time. Mathematical models help us understand the dynamics of virus spread and ecological populations, and vaccines for deadly diseases like HIV and Ebola seem to hold more promise than ever. Many say the twentieth century belonged to physics and the twenty-first belongs to biology, and indeed, this sentiment seems to be borne out by the astonishing advances in the biological sciences.

Six Major Revolutions in Biology

Over the past five hundred years, biology has undergone six major revolutions that transformed our understanding of life. The first was Linnaeus’s classification of organisms into binomial nomenclature. The second was the invention of the microscope by pioneers like Hooke and Leeuwenhoek. The third was the discovery of cells' composition in health and disease by Schwann and Schleiden, thanks to advancements in microscopy. The fourth was Darwin’s formulation of evolution by natural selection. The fifth was Mendel’s discovery of the laws of heredity. And the sixth was the discovery of the structure of DNA by Watson, Crick, and others. An ongoing seventh revolution centers around mapping genomes and understanding their implications for disease and ecology. This has been aided by breakthroughs in statistics and new imaging techniques like MRI and CT scans.

These revolutions were driven not just by new ideas, but also by new tools—a pattern that underscores the intertwined nature of scientific progress. Scientific revolutions are a two-pronged affair: breakthroughs come both from new paradigms of thinking and new technological advances.

The Role of Technology in Biological Discoveries

From the microscope to electron microscopy, x-ray diffraction, and modern imaging techniques, four of the six revolutions in biology hinged on technological advancements. The invention of tools that enabled deeper observation revealed new truths that theory could then catch up with. In genomics, for example, rapid sequencing methods paired with powerful computers and statistical techniques for identifying rare events have unlocked a flood of genetic data. However, this bounty of data brings its own challenges: it hides gems of understanding within mountains of information. Systems biology seeks to make sense of this deluge by piecing together the puzzle at different levels, but even this approach highlights the gaps between what we can observe and what we can truly understand.

The Promise and Peril of Artificial Intelligence in Biology

This is where artificial intelligence (AI) enters the conversation. AI promises to address biology’s complexity by analyzing vast datasets and finding patterns that elude human comprehension. Indeed, AI and machine learning have already shown promise in areas like genomics, protein folding, and drug discovery. For instance, AlphaFold, a machine learning algorithm, achieved a breakthrough in predicting protein structures, which had remained a tough challenge in biology for decades. AI models also excel at recognizing anomalies in medical imaging, predicting the spread of pandemics, and optimizing experimental designs.

However, while AI offers immense potential, it also amplifies existing challenges. The major limitation lies in the fundamental difference between recognizing patterns and understanding causation. AI systems excel at correlating data and identifying intricate relationships, but these correlations do not necessarily reveal underlying biological mechanisms. In fields like cancer genomics, sequencing technologies have revealed hundreds of mutated genes, and AI can help sift through these to find correlations between certain mutations and cancer progression. But beyond identifying these correlations, biologists must still build causal frameworks to explain why and how these mutations impact cancer growth. Without such theoretical underpinnings, AI runs the risk of creating a black box where patterns are identified without leading to genuine understanding. Even in terms of strict utility, these patterns may be inadequate since prediction without understanding can lead to blind spots; for instance witness "activity cliffs" in medicinal chemistry or missed interaction partners in cancer biology, both of which can thwart true prediction.

The very nature of AI models may also reveal their limitations. For instance, LLMs and other leading AI models are trained and based on discrete, binary data, while biology is full of continuous data; something as simple as dose-response curves testifies to this fact. The difference between digital and analog systems comes to mind: computers are digital, but life is a hybrid of analog and digital, combining digital outputs built on analog substrates and vice versa. In a previous post, I argued that the relative thermodynamic inefficiency of the brain might point to its analog workings. If AI were to truly transform drug discovery and biology, its discrete digital models would have to learn to deal with continuous, analog data.

This challenge is not new. Sydney Brenner once pointed out that biology in the 1950s used to be "low input, low throughput, high output," whereas today it's often "low input, high throughput, no output." While AI has the potential to plug gaps in improving output, it can also exacerbate this trend by focusing on high throughput without significant output. Just as we once mistook data accumulation for understanding, we now risk mistaking AI-generated correlations for genuine scientific insight. AI-driven discoveries must therefore be complemented with theoretical and experimental models that go beyond pattern recognition.

Reductionism and Emergence in the Era of AI

One of the key reasons why technology—and now AI—hasn’t “saved” biology lies in the fundamental philosophy of reductionism. Reductionism, the idea of breaking down complex systems into simpler parts, has been the great legacy of twentieth-century science. However, as complexity theorists like Philip Anderson and Stuart Kauffman pointed out, complex systems often display emergent properties that can’t be deduced from their individual components. AI tools, which mostly operate on reductionist principles by breaking down biological data into discrete elements, are thus poorly equipped to handle emergent phenomena.

For example, while AI can analyze neuronal firing patterns and model brain networks, it struggles to explain higher-order cognitive functions and consciousness. Mirror neurons are a case in point: AI can track their activation across brain regions, but understanding their role in human empathy or social behavior remains elusive. Neuroscientists like John Krakauer argue that such limitations reflect the broader challenge of understanding emergent properties within biological systems.

The Future: An Integrated Approach

For AI to truly revolutionize biology, it must not merely amplify reductionist methods but enable a more integrated approach to biological understanding. This involves studying biological systems at multiple levels of organization - both digital and analog - and creating AI models that can reconcile these levels. For example, in neuroscience, we must bridge low-level recordings of single neurons with models of neuronal clusters and behavioral observations of entire organisms. Similarly, in fields like genomics, AI must work alongside biologists to develop causal models that connect molecular changes with physiological and ecological consequences. This is why organizations that center their entire product or discovery pipeline around AI need to be careful and avoid the proverbial pitfall of having everything look like a nail when they have a hammer.

The promise of AI in biology also depends on fostering closer interdisciplinary collaborations. Physicists, computer scientists, biologists, and psychologists need to work together to build models that combine reductionist and holistic perspectives. AI should serve as a tool that integrates data from different levels, facilitating not just more precise measurements, but more comprehensive theories.

Final thoughts

As the late biologist Carl Woese observed, living systems are not simply collections of molecular machines, but resilient patterns in a turbulent flow. In his elegant essay titled “A New Biology for A New Century,” Woese argued for a broader perspective that sees organisms as complex, dynamic organizations. AI must help us move toward this vision by integrating the patterns it finds with deeper theoretical insights, enabling a biology that goes beyond the sum of its parts.

The era of AI in biology holds great promise, but it also poses new risks of mistaking data accumulation for understanding. AI must be seen not as a savior of biological sciences, but as a vital partner in building bridges between different levels of biological understanding. By combining AI’s pattern recognition with theoretical frameworks that address emergence and historical contingency, we can move toward a more integrated, holistic view of life.

As technology advances, we must remember the limitations of reductionism and the dangers of technological solutionism. AI’s role should not be to replace human understanding, but to augment and expand it. Just as we harness the tools of the past revolutions in biology, we must now harness AI to not just process data, but to deepen our understanding of the flow of life itself.

Image link: https://scitechdaily.com/ai-reveals-previously-unknown-biology-we-might-not-know-half-of-whats-in-our-cells/