During the past ~30-40 years, the field of Structural Biology has exploded. From a string of nucleotides, 3-dimensional structures (of e.g., DNA, RNA) can be proposed. From a string of amino acids, cytoplasmic proteins as well as membrane-bound proteins can be predicted. In the latter case, the obvious next step would be to study substrate-enzyme active-site interactions and ligand-receptor interactions. But — how accurate are these computer-generated models in Real Life?
For example, the active binding-site of AHR (aryl hydrocarbon receptor) is characterized by an internal cavity with a volume typically ranging from 200 to 600 ų — depending on the activation state and the specific ligand that is bound. Experimental and computational studies [says Google AI overview] provide the following approximate dimensions for the ligand-binding pocket: pocket radius ~3.64 Å; total cavity surface area ~178.3 Ų. In molecular docking studies, a grid box of 60 × 50 × 50 ų is often used to encompass the entire ligand-binding cavity and its surrounding flexible loops. Internal volume of mouse AHR ~201.5 ų. Internal volume of human AHR can expand to 400–600 ų during molecular dynamics simulations — to accommodate larger ligands such as indirubin or 6-formylindolo[2-b]carbazole (FICZ). The pocket of AHR is described as being bipartite, consisting of a geometrically constrained primary binding site that determines specificity, and a secondary site that allows for AHR’s characteristic ligand promiscuity; this flexibility enables it to bind a wide variety of molecules, from smaller planar compounds such as 2,3,7,8-tetrachlorodibenzo-p-dioxin, TCDD ( ~220 Å) to the larger, bulkier ligand agonists.
One extremely bizarre example comes to mind — when R D Palmiter & E R Mulvihill (Science 1978; 201: 356-358) reported that the insecticide and fungicide chlordecone (Kepone) is apparently a ligand that binds to the estrogen receptor (ESR1) — to induce estrogenic effects in chick oviduct. If ANYONE can see any resemblance between these two diverse chemical structures (yet they are both ligands for the same receptor) — please let me know. This is a great example of receptor promiscuity…!! Perhaps A.I. and/or Machine Learning can provide an answer. 😊
kepone chemical structure from en.wikipedia.org Estrogen Molecular Structure Oestrogen Skeletal Chemical
One more extremely bizarre example comes to mind (especially because it came from my lab): SLC39A8 (ZIP8) is a membrane-bound cation influx transporter — whose endogenous substrates include Mn2+, Zn2+, Fe2+, Co2+, and Se4+ in the form of selenite (HSeO3—). The transporter protein apparent has both its NH2-terminal (N-terminal) and COOH-terminal (C-terminal) ends located on the same side of the membrane (extracellularly for plasma membrane-localized ZIP8). Depending upon the tissue or cell type being studied, however, Mn2+ is preferentially taken up by some organs, whereas Zn2+ is preferentially taken up by other organs. Somehow the membrane-bound transporter changes its shape — to preferentially influx one or other divalent cation. How this is accomplished might be discovered through A.I. and/or ML.
Artificial Intelligence (AI) is the broad concept of machines mimicking human intelligence to perform tasks (in very short periods of time, compared to the human brain), whereas Machine Learning (ML) is a subset of AI focused on training algorithms to learn patterns from data, and improve on these algorithms over time. Essentially, AI is the goal or system, and ML is a primary technique used to achieve it.
There are two articles and two topics in this email. The first one [see attached pdf article] describes how scientists are seeking to decipher the role of non-coding DNA in the human genome — helped by a suite of artificial-intelligence tools. The second article [pasted far below] describes AlphaFold4, and similar AI tools, that are being designed to predict how proteins interact with potential therapeutic molecules (e.g., drugs under development).
[1] Deciphering the role of non-coding DNA in the human genome
Evo is an AI model trained on some 300 billion nucleotide bases, including 80,000 microbial whole-genome sequences. When prompted with ‘#’ — Evo will create a new sequence of mobile DNA (it does this, on the basis of other such biological systems to which the model has been exposed) (see go.nature.com/3jvp922). Given a prompt such as ‘030’, an AI tool called regLM can (in one or two seconds) spit out 200-base sequences that are predicted to exhibit regulatory activity in any of three human cell culture lines (go.nature.com/4jpttm8).
Evo and regLM are part of a fast-growing suite of tools that aim to internalize, decode, interpret, and build on the grammar of the genome — especially the vast portion (~98.5%) that apparently does not code for proteins. Think AlphaFold, but for regulatory DNA, which are sequences that control gene expression.
When Google DeepMind released AlphaFold in 2020, the company claimed it had solved a decades-old ‘grand challenge’
in biology — predicting a protein’s 3D shape from its sequence alone. However, the non-coding fraction of the genome could prove to be an even greater challenge.
As a general rule, a given sequence of amino acids will most likely fold into the same shape, whatever the cellular context. That predictability is not true of the genome, however, in which short, functional sequence motifs (gene promoter regions, and enhancers, transcription start and stop sites, etc.) can be scattered across long stretches of seemingly useless DNA. These motifs might overlap, interact with other DNA regions over long distances, bind to competing protein factors, or respond to signals that are only present in specific cells, or at certain times in development (i.e., an epigenetic effect). They are also tightly wrapped within chromatin, a complex of DNA and protein, which might be more or less accessible to external proteins depending on what the cell is doing… … …
Read the rest of the [attached pdf article]… …
In closing, researchers agree that sequence-to-function models generally work as advertised. But what they and other AI models can and should be used for — remains up for debate. In two studies published in late 2023, computational biologists Nilah Ioannidis et.al. (University of California, Berkeley), and Sara Mostafavi et.al. (University of Washington, Seattle) independently demonstrated that genomic models struggle with a key task: explaining why variation in gene expression differs from person to person — why one person expresses a given gene more than another, given an individual’s unique constellation of gene variants.
To complicate matters, existing benchmarks of AI performance don’t necessarily address the questions that researchers want to ask — especially regarding interpersonal genetic variation. Kundaje’s team (Cornell University, Ithaca, NY) has created its own benchmark tool, called DART-Eval, to help make the evaluation of models more systematic and comprehensive. So has Ioannidis, with her tool called GUANinE.
Anshul Kundaje is even less impressed with unsupervised models. Although they perform well — with coding sequences and small genomes — when it comes to mammalian regulatory DNA, he considers them “catastrophic failures”.
Unsupervised models are largely unaware of the many layers of epigenetic regulation, such as transcription-factor binding and histone modification, which make genomes work; they have no idea what happens when sequences are perturbed. And, of course, not every base in the genome is meaningful. Still — genomics-AI advocates see a promising future. Casey Green (a bioinformatician at the University of Colorado Anschutz Medical Campus, Aurora) foresees a day when AI tools will be
able to design an exquisitely fine-tuned piece of DNA — just as natural-language instructions today can generate code — a process sometimes called vibe coding. “I just want to vibe-code the genome,” he says.
Whether he’ll be able to do that — with a single-character prompt — remains to be seen.
Nature 21 Aug 2025; 644: 829-832
[2] Predicting how proteins interact with drugs and other potential therapeutic molecules
19 February 2026
‘An AlphaFold 4’ – scientists marvel at DeepMind drug spin-off’s exclusive new AI
Isomorphic Lab’s proprietary drug-discovery model is a major advance, but scientists developing open-source tools are left guessing how to achieve similar results.
By Ewen Callaway
A computer generated visualisation of the structure of a protein-protein interaction prediction.
This AI tool includes predictions of how proteins interact with potential therapeutic molecules.
Nearly two years after Google DeepMind released an updated AlphaFold3 geared at drug discovery, its biopharmaceuticals spin-off, Isomorphic Labs, announced an even more powerful artificial-intelligence model — and they’re keeping it all to themselves.
Isomorphic Labs, based in London, touted the capacities of its ‘drug-discovery engine’ — which it calls IsoDDE — in a 27-page technical report, released on 10 February. Achievements, including precise predictions of how proteins interact with potential drugs and antibody structures, have impressed scientists working in the field.
Yet, unlike the AlphaFold AI systems for predicting protein structure — which were made accessible to other researchers and described in depth in journal articles1 2 — IsoDDE is proprietary, and the technical paper offers scant insight into how to achieve similar results.
“It’s a major advance, on the scale of an AlphaFold4,” referring to an unreleased future generation of Google DeepMind’s technology, says Mohammed Al Quraishi, a computational biologist at Columbia University in New York City who is working to develop fully open-source versions of AlphaFold. “The problem, of course, is that we know nothing of the details.”
Drug–protein interactions
AlphaFold 3 was developed with drug discovery in mind. Unlike its Nobel-prizewinning predecessor AlphaFold2, the model could predict the structures of proteins interacting with other molecules — including potential drugs.
Similar AIs modeled after AlphaFold 3 have come close to fully matching its performance and have new capabilities. An open-source model called Boltz-2, developed by scientists at the Massachusetts Institute of Technology in Cambridge and released last year3, could predict the strength to which potential drugs bind proteins, or their binding affinity. This is a key property for developing therapeutics and is usually predicted with computationally intensive physics-based methods.
According to Isomorphic’s report, its new AI outperforms both Boltz-2 and physics-based methods at determining binding affinity. Predictions of how antibodies — which form the basis for therapies that rack up tens of billions of pounds in sales annually — interact with their targets is also state of the art, the report claims.
Al Quraishi says he is especially impressed by IsoDDE’s ability to predict drug–protein interactions of molecules that are vastly different from the data that the model was trained on. “That’s the really hard problem, and suggests that they must’ve done something pretty novel,” he says.
Secret sauce
The models behind IsoDDE are “profoundly different” from other efforts, says Max Jaderberg, Isomorphic’s president. But the company has no plans to reveal the ‘secret sauce’ behind it. “Like with most big machine-learning and AI advancements, it’s a combination of compute, data [and] algorithms,” Jaderberg adds. He hopes his team’s report will “galvanize” the efforts of other teams building drug-discovery AIs.
“This report comes after extensive efforts to partner with industry and potentially access their private structural data, so we don’t know how impactful that extra data are” to IsoDDE’s performance, wrote Diego del Alamo (a computational structural biologist at Takeda Pharmaceuticals, who is based in Cambridge) — on the social-media site X.
Isomorphic has struck drug-development deals, potentially worth billions of pounds, with pharmaceutical companies Johnson & Johnson, Eli Lilly and Novartis. It also has its own internal pipeline, with clinical trials on the horizon. Jaderberg says that the company has developed different versions of IsoDDE from the one used for the technical report, including for work with its partners, that incorporate different data sources.
His colleague Michael Schaarschmidt, Isomorphic’s director of machine learning, says the company’s data strategy is “quite comprehensive”, incorporating publicly available data, synthetic training data, and data sources that they will “try to license”.
Gabriele Corso, a machine-learning scientist who co-developed Boltz-2 and now leads the company Boltz in London, does not think that proprietary data played an essential part in the reported performance of Isomorphic’s tool, on the basis of gains his team is seeing. “There are a lot of improvements we can make with the data that are out there,” he says. “I think this is a new baseline to match — but also to pass.”
doi: https://doi.org/10.1038/d41586-026-00365-7
References
Jumper, J. et al. Nature 596, 583–589 (2021).
Abramson, J. et al. Nature 630, 493–500 (2024).
Passaro, S. et al. Preprint at bioRxiv https://doi.org/10.1101/2025.06.14.659707 (2025).
COMMENTS:
Dear Anonymous,
Your email is humble yet brave. Probably more than half of the 200+ on this mailing list are in the same boat — but don’t have the time, or the interest, or they’re too embarrassed to ask such a question. For the “most crisp” all-inconclusive answer to your question, I simply asked AI Google:
Artificial Intelligence (AI) accelerates drug discovery by replacing slow, manual, and expensive laboratory experiments with rapid, computer-simulated, and data-driven predictions. Rather than testing millions of compounds one by one, AI systems simulate these interactions within virtual environments to pinpoint the most likely candidates for success.
Here is a precise breakdown of how AI accomplishes this:
1. Accurate Protein Structure Prediction (AlphaFold)
Before knowing how a drug binds, researchers need to know the exact 3D shape of the receptor or enzyme.
Method: Tools like AlphaFold2/3 and ESMFold analyze the primary amino acid sequence of a protein and predict its 3D-folded structure with atomic accuracy.
Outcome: These tools can predict protein structures — even when no similar experimental structure exists, allowing researchers to study “undruggable” targets and identify hidden, cryptic binding pockets.
2. Rapid Ligand-Binding Prediction (Molecular Docking & Diffusion)
Once the structure is known, AI simulates how potential drugs (ligands) fit into the receptor or enzyme active-site.
DiffDock & Diffusion Models: These AI systems model the “diffusion” of a ligand from a random position in space into the binding-site, predicting its final docking position with high accuracy.
Speed: AI-driven platforms can scan millions of compounds in days rather than months.
Protein-Ligand Co-generation: Advanced systems, including AlphaFold3, can generate the structures of both the protein and the ligand simultaneously, adapting the pocket structure to fit the ligand (induced fit).
3. Binding Affinity Scoring Functions
After placing a drug in the active-site, the AI must determine how strongly it binds (potency).
Method: Geometric Deep Learning and Graph Neural Networks (GNNs) are used to learn from huge datasets (e.g., PDBbind) of known protein-ligand complexes.
Mechanism: Instead of using traditional physics formulas, AI models learn complex non-linear patterns — such as hydrogen bonds, hydrophobic contacts, and electrostatic interactions — to predict binding affinity more accurately than traditional methods.
4. De Novo Drug Design (Generative AI)
Rather than searching existing libraries, AI can create entirely new molecular structures tailored for a specific enzyme pocket.
GANs and VAEs: Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) learn from large datasets of active compounds. They generate novel molecular structures that fit specific binding-sites and satisfy desired chemical properties.
Reinforcement Learning (RL): RL agents receive rewards for creating molecules that show high predicted binding affinity, stability, and ease of synthesis.
5. Prediction of Druggability (ADMET/Tox)
AI ensures the drug works in the body, not just in the computer.
ADMET Profiling: AI models predict Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) early. They identify potential side effects or safety issues before synthesis.
Filtering: AI helps weed out molecules that are toxic or incapable of being orally absorbed, focusing resources on promising, safe candidates.
6. Molecular Dynamics (MD) Simulation Enhancement
AI models the movement and stability of the drug once it has bound to the receptor.
AI-based Force Fields: Systems like TorchMD use AI to make MD simulations much faster and more accurate, predicting the stability of the bound drug over time.
In summary, AI turns drug discovery from a process of “experimental testing” into a “smart search” and “rational design,” effectively accelerating hit identification, lead optimization, and the design of novel therapeutics.
* * * * * * * * * * * * * * * * * * * * * * * *
Machine Learning (ML) accelerates drug discovery by replacing slow, physics-based simulations with fast, data-driven predictions of molecular interactions. Instead of calculating every force interaction from scratch, ML models learn complex patterns from thousands of previously known protein-ligand structures to predict new binding possibilities in seconds.
Here is how ML works precisely in this field:
1. Representing Data for Machines (Featurization)
Before an algorithm can predict binding, molecular structures (protein pockets and ligands) must be converted into numerical data.
SMILES Strings (1D): Small molecules are converted into “SMILES” strings—a text representation of atoms and bonds — which are processed by deep learning to identify functional groups that might interact with a target.
Molecular Graphs (2D/3D): Molecules are treated as graphs, where atoms are nodes and bonds are edges. Graph Neural Networks (GNNs) learn the connectivity and topological information of the drug.
Voxelization (3D Grid): The active-site of an enzyme is divided into a 3D grid (voxels), similar to 3D pixels, where each voxel contains info on local atomic charges and hydrophobic properties. 3D Convolutional Neural Networks (CNNs) “scan” these voxels to detect binding pockets.
2. Identifying Active Sites (Pocket Prediction)
ML algorithms, such as P2Rank or DeepSurf, can analyze protein surface topology to find potential binding pockets on a protein — even if no ligand is present (apo-structure).
Geometric Deep Learning: Algorithms learn the “fingerprint” of a pocket’s shape, using surface representations like Voronoi diagrams (alpha-spheres) to identify suitable cavities for docking.
3. Predicting Binding Affinity (Scoring)
ML models act as “scoring functions” to determine the strength of binding (Kd, Ki, IC50).
Random Forest & SVM: Traditional ML uses handcrafted features — like the number of hydrogen bonds, hydrophobic contacts, or the count of specific atom type pairs — to predict binding affinity.
Deep Learning/3D CNNs: Models like KDEEP or DeepAtom use 3D voxel representations to learn high-dimensional patterns, often capturing non-linear relationships that classical, rule-based scoring functions miss.
△ML Strategy: ML is used to predict the error in traditional scoring functions. By adding the predicted correction factor △ (Greek letter, delta) to a classical docking score (e.g., AutoDock Vina), the overall accuracy is significantly increased.
4. Predicting Binding Poses (Docking)
ML is transforming how a ligand is oriented (docked) within a pocket.
Diffusion Models (e.g., DiffDock): Instead of exploring thousands of orientations, generative AI uses diffusion models to directly predict the most favorable binding pose in “one shot,” treating it as a translation, rotation, and torsion angle regression problem.
Summary of Techniques Used
Graph Neural Networks (GNNs): Best for analyzing 2D chemical topological structures.
3D CNNs: Used for voxel-based 3D spatial scanning of protein active sites.
Reinforcement Learning (RL): Used to generate new, optimal molecules (de novo design) that fit into a predicted pocket.
Why ML is revolutionary: These methods are exponentially faster than traditional molecular dynamics, allowing for the virtual screening of billions of compounds in a short time.
* * * * * * * * * * * * * * * * * * * * * * * *
From: Anonymous
Sent: Monday, March 30, 2026 10:53 AM
To: Nebert, Daniel
Subject: COMPUTATIONAL STRUCTURAL BIOLOGY — Is this where A.I. and Machine Learning can really make a difference…??
Dear Dr. Nebert:
Your long GEITP email blog from March 26th [below] is all well and good — about describing the problems that need answers at the next level, concerning “Deciphering the role of non-coding DNA in the human genome” and “Predicting how proteins interact with drugs and other potential therapeutic molecules,” and it superficially concludes how Artificial Intelligence and Machine Learning are becoming so helpful in providing answers at the next level. As a completely naïve practicing physician, however, there seems to me to be a disconnect between “what the problems are” and “how A.I. and M..L. will be SO helpful in solving these problems.”
P.S. I <
Warmest wishes, —- ———