Diffusion for Molecules: From Plausible Structures to Useful Drug Design

Strong binding is only the beginning

Finding a drug resembles finding a key for a lock, Evan Feinberg explains. The lock is usually a protein; the key might be a small molecule, peptide, or antibody. Binding can inhibit an enzyme or activate a protein, changing a biological pathway. But strong binding is necessary, not sufficient: the molecule must also avoid unwanted targets, reach the right tissue, and meet safety requirements.

A longstanding hypothesis offered a computational foothold: predict the three-dimensional coordinates of a protein and its bound drug accurately enough, and that structure could improve predictions of binding affinity or potency. Feinberg says earlier predictions of these arrangements, or poses, were too poor to test the hypothesis meaningfully. Experimental structures could cost tens of thousands of dollars and take months or years. Faster, accurate predictions could change both the speed of discovery and which targets researchers could pursue.

The problem connects the guests’ careers. Feinberg’s Stanford PhD work in Vijay Pande’s lab treated molecules as graphs of atoms, bonds, and spatial interactions; large GPU simulations later became useful ingredients for molecular foundation models. Sergey describes moving from physics into software engineering and AI research at Facebook, then returning to physics through Genesis as CTO. The companion article identifies him as Sergey Edunov and specifies his earlier leadership of Llama 2 training and Llama 3 pretraining at Meta.

Feinberg presents Genesis Molecular AI as a Stanford research spinout pursuing compounding progress rather than a single decisive breakthrough. He reports that better protein–ligand structure predictions have recently improved potency prediction, though he gives no effect size or evaluation protocol for that relationship. The broader goal is a system whose structural predictions, property models, and discovery workflows improve together.

Co-folding without an internet of molecular data

[PEARL](https://www.genesis.ml/news/introducing-pearl) predicts a protein and ligand together. It takes a protein sequence and a representation of the ligand—the molecule intended to bind the protein—and predicts their joint three-dimensional structure. This task is called co-folding. The companion article expands PEARL as Place Every Atom at the Right Location. Genesis focuses on small-molecule interactions with proteins, with a broader discovery program that also includes medium-sized molecules such as macrocycles and peptides.

Small molecules still create an enormous search problem. The discussion estimates roughly 10^60 possible drug-like molecules, with rotations and different configurations adding further possibilities. Searching this space requires useful predictions far from familiar examples. A host asks whether scaling merely produces better pattern matching; the response identifies generalization beyond training data as a central challenge for molecular modeling.

Training data is a constraint. The guests describe the Protein Data Bank, or PDB, as containing a couple hundred thousand structures: information-rich examples, but nothing resembling the internet-scale supply available for language modeling. Experimental growth is expensive and slow. Physics simulations provide additional training examples, particularly for small molecules whose behavior is more tractable to model than that of much larger proteins.

Genesis describes three roles for physical knowledge: - Expand training inputs: simulations generate additional pretraining examples. - Constrain what the architecture must learn: physical priors reduce the need to rediscover physics and aim to limit overfitting. - Produce usable coordinates: output constraints help make predictions suitable for medicinal chemists and downstream physical models.

That last requirement makes interoperability part of model quality. Predicted coordinates must make physical sense when passed into existing computational chemistry tools. A diffusion head, the model’s structure-generating component, refines the prediction over multiple steps. Those steps create opportunities for physics-based guidance to steer generation toward usable structures.

The team compares this additional computation to inference-time scaling in language models. Here, however, computation revisits internal structural representations rather than language tokens or a succession of fully materialized structures. Genesis reports substantial improvements from physics-guided inference, without a numerical gain or compute budget. The internal mechanism remains an open interpretability question: a host’s suggestion that the system alternates between learned and physical force fields is not confirmed as an architectural description.

Knowing the target does not provide a treatment

Drug development begins with identifying a disease mechanism and target, then finding a potent, selective molecule that reaches the relevant tissue. Toxicology and preparation for an investigational new drug application precede clinical trials. Phase one typically emphasizes safety; later trials usually grow in size, with regulatory approval as the eventual goal. Target identification, molecular discovery, regulatory preparation, and patient segmentation are distinct modeling problems, even when they contribute to the same medicine.

Genesis concentrates on the gap between understanding disease and making a treatment. The motivating example is a patient whose genome or tumor has been sequenced: a physician can explain the disease but still has no selective therapy to offer. Biological validation and ease of drugging are separate dimensions. A well-understood target can remain difficult to address, even if many easier opportunities have already been pursued.

The company describes two complementary opportunities: - Find the first binder: discover molecules for a target with no known binders. - Improve existing compounds: optimize preclinical molecules or improve on clinical and approved agents. Later-generation ALK inhibitors illustrate the second opportunity. A guest describes substantial survival improvements over earlier generations, while explicitly saying Genesis is not working on that target. No survival figures or study conditions accompany the example.

An expanded pharmaceutical collaboration spans both ends of discovery: - Existing compounds: Genesis reports fine-tuning foundation models on partner data to move closer to selecting a development candidate—a molecule that could ultimately become the agent used in a phase-one trial. - No known chemical starting point: Genesis reports finding initial binding hits for a target without known binders or a ligand-bound structure, then progressing them toward inhibitors active in enzyme-based and living-cell assays. These are discovery-stage reports. They provide no quantitative assay results and do not establish clinical-trial entry or approval.

A plausible pose can hide the wrong interaction

Structural accuracy matters at the level of individual interactions. The guests argue that a prediction can pass a two-angstrom pose criterion while orienting an aromatic ring incorrectly. It may still look reasonable to a chemist, much as an enhanced image can contain a convincing but incorrect face. Plausibility can conceal an error that downstream models amplify.

A host challenges the emphasis on one pose: molecules occupy distributions of configurations, so why privilege a single structure? The response defends the pose as an inspectable intermediate result. Even when binding affinity or potency is the final prediction, a structural hypothesis offers ways to judge whether the answer makes sense. A lone number provides fewer handles. The host also points out that the right pose alone does not establish whether a molecule can enter the binding pocket or persist there.

The response distinguishes two parts of a bound ligand: - The protein-interacting core can have a sharply defined position. The guests point to experimentally resolved electron density as evidence that some local geometry is more than a convenient abstraction. - Solvent-exposed regions may remain dynamic and less sharply defined while still influencing solubility and other properties. Genesis argues that placing the interacting core accurately is especially valuable for downstream binding predictions.

Hydrogen bonds make the scale concrete. An angstrom, written Å, is one-tenth of a nanometer. A guest gives an illustrative donor-to-acceptor heavy-atom distance of 2.7–3.3 Å—a window only 0.6 Å wide—with shorter distances described as clashes and longer distances as rapidly weakening interactions. This motivates sub-angstrom accuracy in the interacting core: a small positional error can change the predicted interaction and the decision about which molecule to make next. It is a practical rationale, not a demonstrated universal one-angstrom cutoff.

The same requirement motivates Sapphire, Genesis’s developing platform for continuous agentic drug discovery. Orchestration amplifies the capabilities and errors of the models it uses. Structural, potency, and other property predictions must therefore be reliable enough that repeated use proposes molecules medicinal chemists would want to make. The image of hundreds of scientists working around the clock expresses the ambition; it is not a measured productivity result.

The evaluation objective changes the training stack

Asked how Genesis pursued higher accuracy, the answer centers on data, infrastructure, and evaluations. A stricter objective changes the whole training stack: which noisy examples to filter, whether to omit them later in training, how to design the architecture, and which loss to optimize. The team says internal and partner drug programs exposed failures that a two-angstrom criterion did not adequately capture.

A guest traces the positional-error convention called RMSD-below-two-angstroms to academic comparisons of docking software, arguing that AI inherited a passing threshold insufficient for drug development. PoseBusters adds physical-validity checks beyond positional error alone. The discussion identifies these checks and the additional structural metric lDDT as part of OpenBind’s broader evaluation. The companion directs readers to the PEARL technical report for its treatment of benchmark weaknesses.

Structure is one part of a larger prediction portfolio. ADMET, expanded in the conversation as absorption, distribution, metabolism, elimination, and toxicity, covers properties that determine whether a molecule can function safely as a drug. Genesis describes a multitask prediction problem spanning dozens of assay properties, each needing to fall within an acceptable range. Examples include solubility, oral bioavailability, inhibition of cytochrome P450 enzymes, and excessive inhibition of the hERG channel, which can cause cardiotoxicity.

These endpoints span different levels of biological complexity. Inhibiting a specific protein such as CYP3A4 is a focused molecular interaction; other endpoints combine several processes and pathways. Public datasets are sparse. A guest recalls earlier success with multitask graph neural networks on large pharmaceutical datasets, without comparative scores. Genesis sees structural modeling as useful beyond binding to the intended target, but does not claim PEARL alone handles every ADMET task.

Earlier tools made diffusion useful

The research history was less linear than a retrospective breakthrough story suggests. Around 2017–2018, generative adversarial networks looked promising for producing protein configurations and protein–ligand poses, but training difficulties limited their usefulness. Diffusion became a more productive primitive for molecular modeling. That experience underlies Feinberg’s assessment that three-dimensional structure prediction now hosts some of the most innovative diffusion research.

Genesis already had physics-based potency tools and molecular-generation methods when co-folding emerged. Those components became ingredients for synthetic-data pipelines and inference-time guidance. Sustained work on adjacent problems prepared the team for a new model family, even though it had not fully anticipated the route.

Laboratory measurements extend that system through a design–make–test–analyze cycle: 1. Design: models predict or propose compounds. 2. Make: an experimental partner synthesizes them. 3. Test: the partner measures downstream properties. 4. Analyze and update: results return to Genesis for fine-tuning or, where appropriate, retraining. Potency measurements can directly train predictive models. Genesis describes rapid partner feedback, but gives no throughput or turnaround measurements.

Reinforcement learning—improving models through feedback—enters as an early result and a development direction. A guest reports initial signs of success, describing a progression from physics-based feedback toward laboratory feedback. The aspiration is to perform more improvement with GPUs and reduce dependence on rapid experimental cycles. Some discovery problems still require experiments, and the conversation supplies no quantitative reinforcement-learning results or training details.

The laboratory is not a deterministic API

Synthesis has variable turnaround. A well-characterized reaction may work under standard conditions; another can fail despite published precedent, forcing chemists to try new conditions. Making a molecule requires coordinating reagents, catalysts, temperature, solvents, and protocols, then purifying and characterizing the product. Tests such as mass spectrometry help establish whether the intended substance was actually made.

A measurement can mislead because the material is wrong. Without purification, either a positive or negative assay result may be an artifact. The guests describe poor translation from some high-throughput screens to resynthesis and higher-fidelity experiments, without giving datasets or numerical correlations. Cleaner candidate selection through AI is a proposed benefit, not a demonstrated comparison in this discussion.

Molecular optimization also contains opposing pressures: - More hydrophobicity can improve binding while worsening solubility. - More polarity can improve solubility while making passage through a cell’s lipid membrane harder. Finding molecules that reconcile these demands may require novel chemistry. Restricting the search to reactions that are easy to automate can increase synthesis speed while limiting the quality and novelty of available compounds. Genesis would welcome more capable robotic synthesis, but says the capability needed for its cutting-edge work was not yet available at the time of the conversation.

This explains the emphasis on strong experimental partners and an internal wet lab. In-house capability trades cost and operational risk for control over the process. Genesis says its lab generates data, tests models in practical workflows, gathers candid feedback from medicinal chemists, and supports an internal drug pipeline.

The company describes its organization as a double helix of AI researchers and experienced drug developers. Its original Genesis Therapeutics name helped signal seriousness to biotechnology practitioners; the change to Genesis Molecular AI reflects its continuing identity as both a model builder and a molecule maker. Pharmaceutical partnerships remain its primary route to scale, combining Genesis’s AI with partners’ biological and clinical-development expertise.

Agents lower the tool barrier and raise an oversight problem

Genesis explicitly defines its agents as large language models that orchestrate scientific tools built over years. Their intended users include scientists who have never written code. The appeal is practical: mastering many specialized tools and configuring their parameters correctly is difficult. Agents can coordinate those tools and interpret model outputs during iterative molecular design.

The discussion outlines three possible approaches to understanding a predicted structure: - Interpret an image: reason over a visual representation. - Use analysis tools: call software that examines the structure. - Learn a structural modality: train a model to understand structural representations directly. These are possible approaches; the exchange does not specify which Genesis has deployed.

The companion article describes the internal SAPPHIRE system as reasoning about poses, forming hypotheses, reading literature, invoking tools, and proposing candidates for subsequent iterations. It presents continuous molecule design and testing through laboratory partnerships as an approaching possibility, rather than an already demonstrated round-the-clock workflow.

A host identifies the oversight problem: an agent may choose the wrong parameter, while a scientist unfamiliar with that parameter may fail to notice. The response acknowledges the risk without explaining concrete safeguards. The envisioned division of work keeps humans responsible for strategic direction and continuing feedback, with agents becoming more independent in execution. The intended benefit is more time for scientific judgment and less repetitive work.

OpenBind tests whether the protein can move

Late in the conversation, OpenBind brings generalization back to a concrete case. Genesis reports running PEARL out of the box on a target it had not encountered during training or development, with substantially better results than other published, openly available models. The challenge concerns one previously unseen target, which the guests acknowledge as both a useful test and a limit on interpreting the comparison.

The protein has a flexible loop that must move to accommodate the ligand. This is induced fit: the bound protein adopts a different configuration from the protein in isolation. Jointly predicting protein and ligand lets PEARL represent that adjustment. Genesis attributes its behavior to the joint training objective, induced-fit examples throughout training, and help from physics-based simulations. It reports predicting the relevant loop movement correctly for essentially every evaluated pose.

The companion article adds that the evaluation contained 802 previously unseen co-complexes on a target identified there as EV-A71. It describes binding as closing the ligand’s entry path into the pocket. Holding the protein fixed makes ligand placement difficult; explicitly resolving its motion is computationally costly. The article reports that PEARL handled this induced fit without long molecular-dynamics simulations.

For PEARL’s OpenBind evaluation, the companion article reports no fine-tuning or use of data on the target or homologous targets, and says the template PDB was released after PEARL’s training cutoff. Those conditions make the result relevant to extrapolation. Neither the conversation nor its companion gives numerical scores or per-metric comparisons, however, and many complexes on one target do not establish performance across many targets. Genesis says the advantage resembles confidential partner-program results, which remain company reports.

Compute, architecture, and the next researchers

Asked which bottleneck they would remove, the guests converge on GPU access. They describe rising prices and competition from language-model companies, saying a large H100 or H200 cluster would be especially valuable. Genesis reports two NVIDIA investments, joint kernel optimization, and collaboration on PEARL. One guest predicts greater chipmaker investment in life sciences because demand for medicines will persist—a forecast about the market, not an announced commitment.

The closing recruiting pitch returns to architecture research. A guest who moved from language models into drug discovery describes finding the field intellectually engaging without prior drug-discovery experience. Genesis presents molecular modeling as an opportunity to develop architectures substantially different from familiar language-model components, and points to architecture work it says it had recently published and open-sourced. At the time of the conversation, it was hiring researchers to connect those methods to medicines.

The companion’s closing argument is that tight integration of machine learning, large-scale computation, and practical drug discovery enabled the progress—and that academic structures make this combination difficult. It sees related advances at Isomorphic and Deep Origin, while acknowledging that closed systems prevent an honest comparison. Its suggestion that protein–ligand binding might eventually be considered solved remains an outlook, with the practical work of making useful medicines still ahead.