New AI Model 'DeCAF-Pearl' Accelerates Drug Discovery by Making Million-Molecule Screening Practical

The intersection of artificial intelligence and pharmaceutical research has reached a critical new milestone with the introduction of DeCAF-Pearl, a biomolecular AI model that drastically reduces the computing power required to design new drugs. Developed by a collaborative team including researchers from Imperial College London, Carnegie Mellon University, and the Genesis Research Team, the framework addresses one of the most severe bottlenecks in modern computational biology: the immense time and hardware costs of simulating molecular interactions at the atomic level. By rethinking the underlying mathematics of how AI generates three-dimensional structures, the team has achieved a fivefold increase in processing speed without sacrificing the accuracy required for clinical research.[1][2]

At the heart of this breakthrough is a process known as "cofolding." In drug discovery, it is not enough to simply know the shape of a target protein; researchers must simultaneously predict the precise three-dimensional structure of the protein and how a small drug molecule will bind to it. This lock-and-key mechanism dictates whether a potential medication will effectively neutralize a disease or fail entirely. For years, the pharmaceutical industry relied on slow, physical laboratory testing to find these matches, a trial-and-error process that contributes to the decade-long timeline and billion-dollar price tag of bringing a new drug to market.[3]

Recently, the industry shifted toward virtual screening, utilizing advanced AI to simulate these interactions digitally. The current state-of-the-art tools—including AlphaFold 3, Boltz-2, and Genesis Therapeutics' own Pearl model—rely on "diffusion" architectures. Similar to the AI used to generate hyper-realistic images, these diffusion models start with a cloud of digital noise and slowly refine it into a highly accurate molecular structure. While this approach yields unprecedented fidelity, it requires the AI to take hundreds of tiny, incremental "denoising" steps for every single molecule it evaluates.[2]

This iterative process is computationally exhausting. When a pharmaceutical company wants to screen a library of millions of potential chemical compounds against a new cancer target, the computing costs of running full diffusion simulations become prohibitive. The sheer volume of Neural Function Evaluations (NFEs)—the number of times the AI must process data to complete a single task—creates a hard ceiling on how many drugs can be virtually tested in a reasonable timeframe. The industry needed a way to maintain the accuracy of diffusion models while bypassing the slow, step-by-step generation process.[3]

Enter the Denoiser Cofolding All-Atom Flowmap, or DeCAF. Instead of forcing the AI to walk step-by-step along the generation trajectory, the researchers built DeCAF on a mathematical framework called "flow maps." This approach teaches the model the average direction of travel between the starting noise and the final molecular structure. By understanding the broader trajectory, the flow map allows the AI to jump directly from one point to another, effectively traversing the entire generation process in just a handful of leaps rather than hundreds of microscopic steps.[1][2]

The computational savings are massive. Where a full simulation using a standard diffusion model requires 200 or more Neural Function Evaluations to accurately cofold a protein and a ligand, DeCAF-Pearl accomplishes the same task using only 40 NFEs. This 80 percent reduction in compute load translates directly to a 5x inference speedup. For laboratories and pharmaceutical companies renting time on expensive cloud supercomputers, this efficiency fundamentally changes the economics of early-stage drug discovery.[2]

To prove the model's real-world viability, the research team benchmarked DeCAF-Pearl against a massive virtual screening task. Utilizing a cluster of 64 graphics processing units (GPUs), the model successfully screened one million distinct molecules against a protein target in approximately 18 hours. Previously, executing a screen of this magnitude with full diffusion-based accuracy would have taken nearly a week on the same hardware, or required a prohibitively expensive expansion of computing infrastructure.[1]

Crucially, this speed does not come at the expense of quality. When tested against a rigorous benchmark of 196 complex protein-ligand structures that the AI had never seen before, DeCAF-Pearl matched the success rate of its slower "teacher" model, Pearl. Furthermore, despite using a fraction of the computational steps, the flow-map model actually outperformed several other leading full-simulation tools, including AlphaFold 3 and Boltz-2, in generating physically valid and accurate poses.[2]

Beyond simply screening existing libraries of chemicals, the 5x speedup unlocks a second, equally vital capability: scalable synthetic data generation. Modern AI models are notoriously data-hungry, and in the realm of structural biology, high-quality examples of proteins binding to drugs are in short supply. Researchers rely on AI to generate synthetic examples to train downstream models, such as affinity predictors that estimate how tightly a drug will bind to its target.

Because DeCAF-Pearl operates so efficiently, research teams can now generate an order of magnitude more synthetic training data per unit of compute. And because the model maintains strict structural accuracy, this synthetic data preserves the delicate physical signals that downstream models depend on to learn. This creates a compounding effect, where faster generation leads to better training data, which in turn produces smarter, more specialized AI tools for the pharmaceutical pipeline.

Dr. Joey Bose, an Assistant Professor at Imperial College London's Department of Computing and a senior author of the study, emphasized that the field is undergoing a fundamental transition. According to Bose, the era of simply training massive foundation models to prove that AI can understand biology is ending. The new frontier is scaling inference—making these models fast and cheap enough to generate the millions of optimized samples required to actually invent new medicines.[1]

This push for efficiency mirrors broader trends across the biotechnology sector, where institutions like MIT are deploying specialized language models to optimize the genetic sequences of protein drugs, aiming to slash manufacturing costs. As AI moves from theoretical research into industrial application, the focus has shifted entirely toward practical deployment, ensuring that breakthroughs in the digital realm can survive the economic realities of physical drug development.[5]

The development of gene set foundation models and flow-map architectures signals a maturing landscape where AI is no longer a blunt instrument, but a highly tuned scientific apparatus. By integrating diverse biological contexts and streamlining the underlying mathematics, researchers are building a unified, accessible representation of human biology that can be queried at unprecedented speeds.[4]

Ultimately, the true impact of DeCAF-Pearl lies in its democratizing potential. By slashing the computational barrier to entry, the flow-map framework allows smaller biotech startups, academic institutions, and independent researchers to perform the kind of massive virtual screening that was once the exclusive domain of multinational pharmaceutical giants. As these tools become faster and more accessible, the timeline between a biological discovery and a life-saving treatment stands to shrink dramatically.[3]