The Diffusion Model Revolution: The Next Frontier in AI Drug Discovery

From Large Language Models to Molecular Science: A Telling Pivot

While the broader tech industry remains consumed by the LLM arms race, some of the best AI researchers are quietly choosing a different path. Sergey Edunov, former head of Meta's Llama project, has left Meta to join Genesis Molecular AI, diving headfirst into drug discovery. That choice alone sends a telling signal: the most cutting-edge, most exciting diffusion model research may not be happening in the LLM space — it may be happening in molecular science.

This conviction was repeatedly confirmed in conversations with Genesis Molecular AI's Evan Feinberg and Sergey Edunov. Having been deeply involved in developing the Llama model family, Edunov understands the engineering logic of large-scale generative models better than almost anyone. His pivot to drug discovery carries a clear message: the field of molecular modeling is at an inflection point — one that feels a lot like the moment just before the LLM explosion.

Why Diffusion Models Are a Natural Fit for Drug Discovery

Drug discovery is a high-value, high-difficulty scientific challenge. Traditional development cycles span over a decade, cost billions of dollars, and still fail most of the time. Generative AI — and diffusion models in particular — offer the potential for a genuine paradigm shift.

Diffusion models are a class of generative models that learn data distributions by progressively adding noise to data and then learning to reverse that process. The core idea is rooted in non-equilibrium thermodynamics, formalized around 2020 with work like DDPM, before achieving breakthrough success in image generation (e.g., Stable Diffusion). Unlike language, molecules and proteins are inherently three-dimensional geometric objects — specifically, arrangements of atoms in 3D Euclidean space or SE(3) symmetric group space, where each atom carries continuous and discrete attributes like coordinates, type, and bonding relationships. This is precisely the kind of continuous-space generation task that diffusion models excel at. By contrast, natural language is a discrete sequence, and applying diffusion models to discrete spaces requires additional techniques (such as MDLM or D3PM) — far less natural than working in continuous space. This is why molecular models like DiffDock and AlphaFold3 can leverage diffusion frameworks to their full potential. And it's the core reason why top researchers like Edunov have concluded that molecular science is where diffusion models truly belong.

PEARL: Zero-Shot Victory on the OpenBind Benchmark

The most technically significant highlight of these conversations is Genesis's PEARL model achieving state-of-the-art zero-shot performance on the OpenBind benchmark.

Co-Folding and Binding Prediction: A Core Challenge in Drug Discovery

One of the central problems in drug discovery is predicting how a small-molecule drug (a ligand) binds to a target protein. This is known as "co-folding" — simultaneously predicting the 3D structures of both the protein and the ligand, and the way they interact.

The concept of co-folding gained enormous attention after AlphaFold2 (DeepMind, 2021) revolutionized single-protein structure prediction, and AlphaFold3 (2024) extended those capabilities to joint prediction of protein-small molecule and protein-nucleic acid systems — a landmark in co-folding technology. Traditional molecular docking methods (such as AutoDock and Glide) typically assume a fixed protein conformation, only allowing the ligand to be flexible. Co-folding, by contrast, allows the protein and ligand to mutually adapt — capturing the "induced fit" effect and more closely mirroring real biological processes. Accurately predicting binding conformations and affinities is the foundation for screening drug candidates, with RMSD (root mean square deviation) and binding affinity prediction accuracy as the core evaluation metrics.

OpenBind is a standardized benchmark designed around exactly these real-world challenges. PEARL achieving top performance in a zero-shot setting — without targeted training for this task — demonstrates that the model has genuinely learned generalizable molecular interaction principles, rather than simply memorizing training data. This generalization capability is critical for real-world AI drug discovery, where new drug targets are often completely novel challenges the model has never encountered.

What Zero-Shot Capability Actually Means

In machine learning, zero-shot capability means a model can perform a task directly without task-specific training. In structural biology, this has a particularly important meaning: the human proteome contains roughly 20,000 protein-coding genes, and many disease-relevant targets lack experimentally resolved complex structures in structural databases — they exist in "data-scarce regions." Traditional deep learning docking methods tend to perform well on target families covered by training data, but accuracy drops sharply when facing entirely novel targets.

Zero-shot performance thus becomes a litmus test for whether a model has truly internalized physicochemical principles — such as hydrophobic effects, hydrogen bond geometric constraints, and π-π stacking — rather than relying on statistical pattern memorization from the training set.

The core value of zero-shot capability is this: the model can be directly applied to entirely new protein-ligand combinations without needing to collect new data or retrain for each new target. In therapeutic contexts with scarce data — orphan diseases, novel pathogens — this capability is enormously practical, and serves as strong evidence that the model genuinely "understands" the physics of molecular interactions.

Crossing the Accuracy Threshold: Unlocking New Possibilities

Another compelling argument from these conversations is the question of what new applications become possible when co-folding technology crosses an accuracy threshold.

From "Can See" to "Can Actually Use"

Throughout the history of scientific tools, accuracy often has a decisive tipping point. Below it, a tool can only serve as a reference or source of inspiration. Once crossed, the tool can be genuinely embedded in real workflows — becoming reliable, production-grade infrastructure. AlphaFold's revolution in protein structure prediction is the canonical example of crossing exactly such a threshold. Before AlphaFold2, X-ray crystallography and cryo-EM were the dominant methods for resolving protein structures, with each experimental structure requiring months or even years to obtain. AlphaFold2 elevated prediction accuracy to a level comparable to experimental methods (median TM-score above 0.9), allowing researchers to obtain trustworthy structural models in hours — fundamentally transforming structural biology.

Once co-folding technology reaches sufficient accuracy, researchers will be able to efficiently screen and design drug molecules computationally, dramatically reducing the number of expensive wet lab iterations required. This not only accelerates development timelines but could also make previously cost-prohibitive targets tractable again.

Generative Molecular Design: From Screening Tool to Design Engine

When predictions become reliable enough, diffusion models can do more than "evaluate" given molecules — they can "generate" entirely new molecular structures with desired binding properties. This means AI will evolve from a passive screening tool into an active molecular design engine, realizing the vision of "designing drugs on demand" — one of the most anticipated breakthroughs in AI-driven drug discovery.

A Quiet Migration of Top Talent

Edunov's move from Meta to Genesis may be just one data point in a much larger trend. As the LLM field gradually enters a phase of engineering optimization and commercialization, a cohort of top researchers is turning its attention toward vertical domains with higher scientific value and greater technical challenges.

Drug discovery, materials science, protein engineering — these fields share a common technical foundation. Geometric deep learning is particularly central: unlike traditional deep learning, which processes regular grid data like images and text, geometric deep learning is specifically designed for non-Euclidean data with symmetry constraints. Molecules naturally possess SE(3)/E(3) equivariance — their physical properties do not change under global rotation and translation. Representative work includes equivariant neural network architectures such as SchNet, EGNN, and SE(3)-Transformer, as well as more recent models like FrameDiff and FoldFlow that integrate these ideas into diffusion frameworks. By embedding physical symmetry priors directly into network architecture, these models achieve smaller parameter counts, stronger generalization, and greater robustness to limited data — all of which are crucial in the life sciences, where experimental data is expensive and scarce. It is precisely the combination of geometric deep learning and diffusion generative models that forms the primary technical engine of the current AI molecular science wave, carrying enormous potential to reshape human health and industry.

As the most talented researchers and the most advanced methods begin converging here, the golden age of AI for Science may truly be arriving.

For practitioners following the AI frontier, this signal deserves serious attention: the true cutting edge of innovation is quietly shifting — from general-purpose conversational systems toward scientific applications capable of solving humanity's most fundamental problems.

Key Takeaways