top of page
Search

AlphaFold Solved Structure, but Can AI Solve Interaction? Moving from Static Folding to Dynamic Interaction

Digital illustration of protein structures in blue and multicolors connected by light wave patterns on a grid backdrop, suggesting data flow.

1. Introduction: The Post-Folding Landscape

The early 21st century of computational biology will likely be remembered for the resolution of the "protein folding problem"—a grand challenge that stood for fifty years as the primary obstacle to understanding biological structure. With the advent of deep learning architectures, most notably AlphaFold2, the scientific community gained the ability to predict the static, three-dimensional structure of monomeric proteins from their amino acid sequences with near-experimental accuracy. This achievement was monumental, earning its developers the Nobel Prize and fundamentally altering the trajectory of structural biology. Yet, as the dust settles on the folding revolution, a more complex and biologically significant reality has come into focus: biology is not a museum of static sculptures; it is a dynamic, chaotic dance of molecular interaction.

Proteins rarely function in isolation. They are the gears of cellular machinery, and their function is predicated on their ability to bind—to other proteins, to nucleic acids (DNA and RNA), and to small molecule ligands (metabolites and drugs). The "Folding Problem" has thus given way to the "Binding Problem." This new frontier is exponentially more complex. While folding is largely driven by the hydrophobic collapse of a single chain seeking its thermodynamic minimum, binding involves the recognition of two or more distinct entities in a crowded, solvent-filled environment. It is governed by nuanced balances of enthalpy and entropy, transient conformational shifts, and the probabilistic nature of molecular encounters.

This report provides an exhaustive analysis of the state-of-the-art computational systems designed to solve this interaction challenge. We are currently witnessing a divergence in algorithmic philosophy: one path, exemplified by AlphaFold 3 and its open-source counterpart OpenFold3, pursues a "holistic" generative approach using diffusion models to predict the joint structure of entire complexes. Another path, represented by the Boltz ecosystem (Boltz-1, Boltz-2, and BoltzGen), seeks to integrate deep learning with physical priors, aiming to predict not just the shape of a complex, but the thermodynamic strength of the interaction (binding affinity). Simultaneously, models like PopEVE are mapping the evolutionary fitness landscapes that constrain these interactions, utilizing population genetics to predict the functional consequences of variation.

We will explore how these technologies are moving the field from descriptive biology—predicting what exists—to prescriptive engineering—designing novel binders and therapeutics with atomic precision. We write this for the student of the life sciences who recognizes that the future of biology is computational, generative, and inextricably linked to the physics of binding.

2. The Biophysical Imperative: Why Binding is the "Hard Problem"

To appreciate the engineering marvels of AlphaFold 3 or Boltz-2, one must first confront the biophysical magnitude of the problem they attempt to solve. The transition from predicting a single protein's fold to predicting a protein-ligand complex is not merely a matter of adding more atoms; it is a fundamental shift in the thermodynamic landscape.

2.1 The Thermodynamics of Recognition

The binding of a ligand (L) to a protein receptor (R) to form a complex (RL) is governed by the change in Gibbs free energy (ΔG). This relationship is defined by the classic thermodynamic equation:


ΔG = ΔH - TΔ S

Here, ΔH (enthalpy) represents the energetic "bonus" gained from specific interactions at the binding interface: hydrogen bonds, salt bridges, and van der Waals forces.1 These are the interactions that traditional "lock-and-key" models focus on—the geometric fit of the ligand into the pocket.

However, the second term, -TΔS (entropy), presents the true computational challenge. When a drug molecule binds to a protein, it loses degrees of freedom. It stops tumbling and rotating freely in solution (loss of translational and rotational entropy) and adopts a restricted conformation (loss of conformational entropy). Simultaneously, water molecules that were previously ordered around the hydrophobic binding pocket are displaced into the bulk solvent, gaining entropy (the hydrophobic effect). Accurately predicting binding affinity requires accounting for these invisible, entropic costs and benefits.1

Traditional physics-based methods, such as Free Energy Perturbation (FEP) or Thermodynamic Integration (TI), attempt to calculate these values by simulating the physical transformation of one molecule into another (alchemical simulation) over millions of time steps. While accurate, FEP is computationally exorbitant, often requiring days of supercomputer time for a single binding pair.1 The goal of the new wave of AI models is to approximate this physical accuracy through geometric reasoning, compressing weeks of simulation into seconds of inference.2

2.2 The "Breathing" Protein: Induced Fit and Allostery

A critical failure mode of early computational docking was the "rigid body" assumption. Docking algorithms often treated the protein as a frozen rock and the ligand as a flexible key. In reality, proteins "breathe." They undergo continuous thermal fluctuations. Upon binding, a protein often undergoes induced fit—a conformational change where the binding pocket reshapes itself to accommodate the ligand.2

Furthermore, binding events often involve allostery, where a molecule binds to a distal site and induces a shape change at the active site. AlphaFold 3 and Boltz-2 represent a leap forward because they move beyond static snapshots. By utilizing diffusion processes (AF3) or training on Molecular Dynamics (MD) ensembles (Boltz-2), these models attempt to capture the plasticity of the protein, predicting the complex as a joint, flexible system rather than a collision of rigid bodies.4

3. AlphaFold 3: The Diffusion Paradigm

If AlphaFold2 was the triumph of the "Evoformer"—an attention-based architecture that extracted spatial constraints from evolutionary history—AlphaFold 3 (AF3) is the triumph of Diffusion. Released by Google DeepMind and Isomorphic Labs, AF3 represents a radical architectural overhaul designed to handle the universality of biological molecules.6

3.1 From Geometry to Generative Denoising

The most significant architectural shift in AF3 is the replacement of the structural generation module. AlphaFold2 used a deterministic "Structure Module" that rotated and translated rigid protein triangles to build the backbone. AF3 replaces this with a Diffusion Network, a technology adapted from generative AI models used in image creation (like DALL-E or Stable Diffusion).6

The diffusion process begins with a "cloud of atoms"—a noisy, disordered distribution of atomic positions with no discernable structure. The model then operates as a "denoiser." Conditioned on the sequence information and evolutionary data processed by the Evoformer, the network iteratively refines this cloud. Over many steps, the fog of atoms coalesces into a sharp, chemically valid molecular structure.6

This diffusion approach offers distinct advantages for binding:

  1. Joint Probability Distribution: Diffusion models predict the positions of all atoms (protein, ligand, DNA, water) simultaneously. This allows the model to capture the cooperative nature of binding, where the ligand stabilizes the protein loop and the loop stabilizes the ligand.6

  2. Generative Flexibility: Unlike rigid geometric modules, diffusion is inherently probabilistic. It can theoretically sample multiple valid conformations of a flexible loop or a disordered region, although AF3 is currently tuned to produce the single most likely structure.7

  3. Chemical Universality: The diffusion network does not need to be hard-coded with the geometry of peptide bonds. It learns chemical constraints from data. This flexibility allows AF3 to model non-protein entities—DNA double helices, RNA loops, and diverse small molecule drugs—within the same unified framework.5

3.2 The Scope of the "All-Atom" Model

AlphaFold 3 is described as a "holistic" model. It accepts inputs not just of protein sequences, but of nucleic acids and SMILES strings (simplified molecular-input line-entry system codes for small molecules).

3.2.1 Protein-Ligand Interaction

In the realm of drug discovery, AF3 has set a new benchmark. On the PoseBusters benchmark, which evaluates the accuracy of ligand positioning and chemical validity, AF3 achieved approximately 50% higher accuracy than the best traditional docking methods.6 This is particularly disruptive because AF3 performs "blind docking"—it predicts the binding site and the ligand pose without being told where the pocket is. This contrasts with traditional methods that usually require a pre-defined search box.5

3.2.2 The Nucleic Acid Frontier

AF3 has also cracked the code for protein-nucleic acid interactions. It can predict the structure of transcription factors bound to DNA or RNA-modifying enzymes wrapped around RNA strands.7 This includes complex topologies like G-quadruplexes (four-stranded DNA knots) and pseudoknots in RNA aptamers.8 However, researchers have noted that while AF3 excels at canonical interactions found in the PDB, it can struggle with rare, non-genomic nucleic acid structures due to the scarcity of training data—a phenomenon known as "hallucination" where the model reverts to the most common fold when uncertain.8

3.3 SiteAF3: Controlling the Diffusion

Despite its power, the "black box" nature of AF3's diffusion can be a liability. In drug design, a chemist might want to target a specific "cryptic" pocket to achieve a distinct therapeutic effect (e.g., allosteric inhibition). The standard AF3 model decides for itself where the ligand goes.

To address this, the research community has developed SiteAF3, a method that modifies the diffusion process to introduce steerability.5 SiteAF3 utilizes "conditional diffusion," where the receptor structure is fixed or heavily biased, and the diffusion network is constrained to generate the ligand within a user-defined region (a binding pocket or near hotspot residues). This restores the control needed for rational drug design, allowing researchers to ask "what if" questions about specific sites rather than accepting the model's global preference.5

4. OpenFold3: The Infrastructure of Open Science

The release of AlphaFold 3 was met with both acclaim and criticism. Unlike AlphaFold2, DeepMind did not initially release the code or weights for AF3, offering only a restricted web server. This "closed" approach catalyzed a vigorous response from the open-science community, leading to the formation of the OpenFold Consortium and the development of OpenFold3.9

4.1 The Challenge of Bitwise Reproduction

OpenFold3 is not merely an imitation; it aims to be a "bitwise reproduction" of the AlphaFold 3 architecture, re-engineered for public and commercial use.9 This project is a massive undertaking in systems engineering, involving academic labs (like the AlQuraishi Lab at Columbia) and high-performance computing centers (like Lawrence Livermore National Laboratory).10

The training of such a model requires massive computational resources. The OpenFold3 team has developed ElMerFold, a highly optimized workflow designed to run on the world's fastest supercomputers. This workflow utilizes:

  • FlashAttention and Kernel Optimization: To handle the immense memory requirements of calculating pairwise interactions between thousands of residues.12

  • NVIDIA cuEquivariance: A library that accelerates the geometric operations required to ensure the model respects the physical laws of rotation and translation.13

  • DeepSpeed and Flux: Advanced workload managers that allow the training to scale across thousands of GPUs simultaneously, a necessity for processing the petabytes of training data.11

4.2 The Data Bottleneck: Distillation

A critical insight from the OpenFold3 technical report is the reliance on data distillation.12 The Protein Data Bank (PDB) contains only about 200,000 structures—too few to train a model with hundreds of millions of parameters without overfitting. AlphaFold 3 solved this by "self-distillation": using an early version of the model to predict structures for millions of uncharacterized sequences, creating a massive synthetic dataset.

OpenFold3 faces the challenge of recreating this synthetic dataset. The consortium aims to simulate a billion different "multimers" (protein complexes), a task that requires exascale computing.11 This reveals that the barrier to entry in modern AI is no longer just the architecture, but the sheer computational cost of generating the training data.

4.3 Federated Learning: Unlocking the "Dark Matter"

Perhaps the most innovative aspect of the OpenFold ecosystem is its embrace of Federated Learning (FL). The pharmaceutical industry possesses vast archives of high-quality protein-ligand crystal structures that are proprietary and never released to the PDB. This "dark matter" of structural biology is invaluable.

OpenFold3, integrated with NVIDIA FLARE, enables a federated approach. The model can be sent to a pharmaceutical company's secure server, trained on their private data, and then the weight updates (gradients)—not the data itself—are sent back to the central server.13 This allows the model to learn from the collective wisdom of the entire industry without compromising trade secrets. It represents a shift from "competitive silos" to "shared infrastructure," potentially allowing OpenFold3 to surpass the accuracy of models trained solely on public data.16

5. Boltz-2: The Physical Foundation Model

While AlphaFold 3 focuses on broad structural coverage, Boltz-2 (developed by MIT's CSAIL and Recursion) targets the metric that matters most to drug developers: Binding Affinity (K_d, K_i, or IC_{50}).10

5.1 Beyond Shape: The Quest for Affinity

Knowing the shape of a complex is necessary but insufficient. A drug may fit perfectly into a pocket geometrically but bind with weak affinity due to poor electrostatic complementarity or high desolvation costs. Boltz-2 is the first biomolecular foundation model designed to jointly predict structure and binding affinity, effectively bridging the gap between geometric deep learning and physical chemistry.3

The architecture of Boltz-2 extends the AlphaFold-like trunk with a specialized Affinity Head.18 This module analyzes the predicted interface and outputs two key metrics:

  1. Binder Probability: A binary classification score indicating the likelihood that the molecule binds at all (distinguishing active drugs from inactive "decoys").

  2. pIC50/Kd Prediction: A continuous regression output predicting the strength of the binding.18

5.2 Learning from Dynamics, Not Just Statics

A key differentiator for Boltz-2 is its training data. While AF3 relies heavily on static crystal structures, Boltz-2 incorporates structural ensembles derived from Molecular Dynamics (MD) simulations and NMR data.19

This training strategy teaches the model to recognize the "fuzziness" of biology. It learns that protein loops are not fixed wires but flexible chains that occupy a volume of space. By supervising the model to predict B-factors (a measure of atomic flexibility) and RMSF (Root Mean Square Fluctuation), Boltz-2 implicitly learns the entropic components of binding.19

5.3 Benchmarking against Physics: The FEP Challenge

The gold standard for affinity prediction has long been Free Energy Perturbation (FEP). FEP is rigorous but slow. Boltz-2 claims to approach the accuracy of FEP while being 1000x faster.3

  • FEP+ Benchmark: On a standard industry dataset comprising targets like TYK2 and CDK2, Boltz-2 achieved a Pearson correlation of 0.66 with experimental results.19 While not perfect (experimental error limits the theoretical maximum correlation to ~0.8), this is significantly better than traditional docking scores and fast enough to screen millions of compounds.

  • Virtual Screening: In the CASP16 competition, Boltz-2 outperformed all other entries in the affinity prediction track.19 This suggests that for the first time, AI can reliably filter vast chemical libraries for high-affinity binders without the massive compute cost of physics-based simulations.

5.4 Steerability and Interaction Control

Boltz-2 also emphasizes user control. It introduces "steerability features" that allow researchers to inject experimental knowledge into the prediction.19

  • Pocket Constraints: A user can define a specific volume of the protein as the target site.

  • Template Integration: A user can force the model to use a specific homolog as a template.This "human-in-the-loop" capability is essential for handling difficult cases where the AI might otherwise "hallucinate" a binding site in an irrelevant region.21

6. BoltzGen: The Generative Leap

If Boltz-2 is the "discriminator" (evaluating how well a molecule binds), BoltzGen is the "generator" (creating the binder from scratch). It represents the transition from analyzing biology to engineering it.22

6.1 The Inverse Folding Problem and Beyond

Traditional protein design often framed the problem as "Inverse Folding" (e.g., ProteinMPNN): given a fixed backbone structure, find a sequence that folds into it. BoltzGen goes a step further, solving the co-generation problem. It generates both the backbone geometry and the amino acid sequence simultaneously to maximize binding affinity to a target.23

BoltzGen acts as an "all-atom generative model." It can design diverse modalities:

  • Miniproteins: Small, stable scaffolds that bind tight targets.

  • Peptides: Including cyclic peptides, which are often more stable and cell-permeable than linear chains.

  • Nanobodies: Single-domain antibodies derived from camelids, which are easier to manufacture than full antibodies.24

6.2 Programmable Biology

A defining feature of BoltzGen is its Design Specification Language. Users do not just press a "generate" button; they write a specification.

  • "Design a cyclic peptide of 12 residues."

  • "Design a binder that contacts residues H41 and K89 on the target."

  • "Design a nanobody but fix the framework region to be human-compatible."

This programmability allows BoltzGen to be integrated into complex drug discovery pipelines. In wet-lab validations, the model has demonstrated a stunning success rate. For a set of diverse targets, including those with no known binders, BoltzGen produced nanomolar-affinity binders for 66% of the targets.23 This success rate rivals or exceeds established methods like phage display, potentially compressing years of antibody discovery into weeks of computation.

7. PopEVE: The Evolutionary Landscape of Fitness

While AlphaFold and Boltz focus on the physics of atoms, PopEVE focuses on the genetics of fitness. Developed by the Marks Lab at Harvard, PopEVE (Population-calibrated Evolutionary Model of Variant Effect) provides the essential context of "allowable variation".25

7.1 The Limits of Physics-Based Scoring

Physics-based models can predict if a mutation destabilizes a protein (folding energy). However, they often fail to predict if that destabilization matters biologically. A mutation might destroy a protein's stability, but if the protein is expressed at high levels or has a chaperone, the organism might survive. Conversely, a subtle mutation might not affect stability but could disrupt a critical binding interface or regulatory site.

PopEVE solves this by looking at the "fossil record" of the protein. It uses Deep Generative Models (specifically Variational Autoencoders, or VAEs) trained on millions of years of evolutionary sequence data (the EVE component). If a residue has been conserved across yeast, mice, and humans, it is likely critical.27

7.2 Integrating Human Population Data

PopEVE adds a crucial second layer: Human Population Genetics. Using data from the UK Biobank and gnomAD, it analyzes the "shallow" evolution of humans. It identifies which genes are under constraint in the current human population.

  • The Latent Gaussian Process: PopEVE uses a Gaussian process to learn a mapping between the deep evolutionary score (from EVE) and the observed constraint in humans. This allows it to calibrate scores across the entire proteome.27

  • Proteome-Wide Comparison: Previous tools could rank mutations within a gene (Mutation A is worse than Mutation B). PopEVE can rank mutations across genes (Mutation A in Gene X is worse than Mutation B in Gene Y). This provides a universal "Richter scale" for genetic damage.28

7.3 Clinical and Engineering Applications

The primary application of PopEVE is Rare Disease Diagnosis. It has successfully identified over 100 novel candidate genes for Severe Developmental Disorders (SDD) by flagging "orphan" variants that other models classified as benign.25 In "singleton" cases (where parents are not available for sequencing), PopEVE can prioritize causal variants with 98% accuracy.29

For Protein Design, PopEVE defines the "safe zones" of the sequence space. If BoltzGen suggests a high-affinity binder, PopEVE can be used to check if the proposed mutations are "evolutionarily valid" or if they are likely to cause aggregation, immunogenicity, or off-target effects based on the fitness landscape.30

8. Comparative Analysis: The Ecosystem of Binding

We can now view these tools not as competitors, but as complementary components of a modern computational biology stack.

Table 1: Feature Comparison of Leading Interaction Models

Feature

AlphaFold 3

OpenFold3

Boltz-2

PopEVE

Core Architecture

Diffusion + Evoformer

Diffusion (Bitwise Reproduction)

Geometric DL + Ensemble

VAE + Gaussian Process

Primary Output

Static Structure (Complex)

Static Structure (Complex)

Structure + Affinity (K_d)

Variant Fitness Score

Training Data

PDB + Self-Distillation

PDB + Open Distillation

PDB + MD Ensembles

MSAs + Population Data

Key Innovation

All-Atom Diffusion

Federated Learning

Affinity Prediction

Proteome-Wide Calibration

Licensing

Proprietary / Server

Open Source (Apache 2.0)

Open Source (MIT)

Open Source

Steerability

Low (via SiteAF3)

High

Native (Constraints)

N/A

Speed

Moderate

Scalable (HPC)

1000x vs FEP

Fast (Inference)

8.1 The Convergence of Methods

The future lies in the integration of these approaches.

  1. Structure Generation: Use AlphaFold 3 or OpenFold3 to generate the base structure of a novel complex.

  2. Affinity Scoring: Use Boltz-2 to predict the binding affinity and filter out weak binders, leveraging its sensitivity to dynamics.

  3. Generative Optimization: Use BoltzGen to redesign the interface for higher affinity or specificity.

  4. Fitness Check: Use PopEVE to ensure the designed sequences are robust, stable, and non-pathogenic in a biological context.

9. Conclusion: The Generative Future

The transition from "Protein Folding" to "Biological Binding" marks the maturation of computational biology. We have moved beyond the passive observation of static structures to the active simulation and engineering of dynamic interactions.

AlphaFold 3 has proven that diffusion models can capture the joint probability of life's molecules, treating DNA, RNA, and proteins as a unified chemical language. OpenFold3 has demonstrated that this power can be democratized and scaled through federated infrastructure, turning the "dark matter" of private data into a public good. Boltz-2 has bridged the gap between deep learning and thermodynamics, offering a "computational microscope" that sees not just where atoms are, but how strongly they hold together. Finally, PopEVE reminds us that these physical interactions operate within the constraints of evolutionary history.

For the undergraduate researcher entering this field, the message is clear: the era of solving single structures is over. The era of designing molecular interactions—of rewriting the code of life to bind, inhibit, activate, and cure—has just begun.

Citations

1

Works cited

  1. Thermodynamic Integration in 3n Dimensions Without Biases or Alchemy for Protein Interactions - Frontiers, accessed December 19, 2025, https://www.frontiersin.org/journals/physics/articles/10.3389/fphy.2020.00202/full

  2. Computational Methods in Biophysics and Medicinal Chemistry: Applications and Challenges - AMS Tesi di Dottorato, accessed December 19, 2025, https://amsdottorato.unibo.it/id/eprint/6329/1/Dissertation_Di_Martino.pdf

  3. Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction | bioRxiv, accessed December 19, 2025, https://www.biorxiv.org/content/10.1101/2025.06.14.659707v1

  4. Thermodynamics and Kinetics of Drug-Target Binding by Molecular Simulation | Chemical Reviews - ACS Publications, accessed December 19, 2025, https://pubs.acs.org/doi/10.1021/acs.chemrev.0c00534

  5. Accurate Site-specific Folding via Conditional Diffusion Based on Alphafold3 | bioRxiv, accessed December 19, 2025, https://www.biorxiv.org/content/10.1101/2025.07.06.663385v1.full-text

  6. AlphaFold 3 predicts the structure and interactions of all of life's ..., accessed December 19, 2025, https://www.isomorphiclabs.com/articles/alphafold-3-predicts-the-structure-and-interactions-of-all-of-lifes-molecules

  7. Accurate site-specific folding via conditional diffusion based on AlphaFold3 - PNAS, accessed December 19, 2025, https://www.pnas.org/doi/10.1073/pnas.2521048122

  8. Direct Modeling of DNA and RNA Aptamers with AlphaFold 3: A Promising Tool for Predicting Aptamer Structures and Aptamer–Target Interactions | ACS Synthetic Biology - ACS Publications, accessed December 19, 2025, https://pubs.acs.org/doi/10.1021/acssynbio.5c00196

  9. aqlaboratory/openfold-3: A fully open source biomolecular structure prediction model based on AlphaFold3 - GitHub, accessed December 19, 2025, https://github.com/aqlaboratory/openfold-3

  10. The Great Divergence: Why OpenFold3 Represents a New, Foundational Platform for Bio-AI, accessed December 19, 2025, https://medium.com/@dapscience/the-great-divergence-why-openfold3-represents-a-new-foundational-platform-for-bio-ai-72fd13b9a774

  11. LLNL and partners launch record-breaking protein-folding workflow on world's fastest supercomputer, accessed December 19, 2025, https://www.llnl.gov/article/53581/llnl-partners-launch-record-breaking-protein-folding-workflow-worlds-fastest-supercomputer

  12. Scalable prediction of symmetric protein complex structures - bioRxiv, accessed December 19, 2025, https://www.biorxiv.org/content/biorxiv/early/2025/11/17/2025.11.14.688531.full.pdf

  13. How to Predict Biomolecular Structures Using the OpenFold3 NIM | NVIDIA Technical Blog, accessed December 19, 2025, https://developer.nvidia.com/blog/how-to-predict-biomolecular-structures-using-the-openfold3-nim/

  14. How Novo Nordisk, Columbia University and AWS collaborated to create OpenFold3 with the OpenFold AI Consortium, accessed December 19, 2025, https://aws.amazon.com/blogs/hpc/how-novo-nordisk-columbia-university-and-aws-collaborated-to-create-openfold3/

  15. Advancing protein prediction with Federated Learning on NVIDIA DGX Cloud - Apheris, accessed December 19, 2025, https://www.apheris.com/resources/blog/advancing-protein-prediction-with-federated-learning-on-nvidia-dgx-cloud

  16. The AI model OpenFold3 takes a crucial step in making protein predictions - Science News, accessed December 19, 2025, https://www.sciencenews.org/article/ai-model-openfold3-protein-predictions

  17. MIT and Recursion Release Boltz-2: Next Generation AI Model to Predict Binding Affinity at Unprecedented Speed, Scale, and Accuracy, accessed December 19, 2025, https://ir.recursion.com/news-releases/news-release-details/mit-and-recursion-release-boltz-2-next-generation-ai-model

  18. The Boltz-2 FAQ - Rowan, accessed December 19, 2025, https://www.rowansci.com/blog/boltz2-faq

  19. Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction | bioRxiv, accessed December 19, 2025, https://www.biorxiv.org/content/10.1101/2025.06.14.659707v1.full-text

  20. How have AlphaFold 3's predictions been validated? - EMBL-EBI, accessed December 19, 2025, https://www.ebi.ac.uk/training/online/courses/alphafold/alphafold-3-and-alphafold-server/introducing-alphafold-3/how-have-alphafold-3s-predictions-been-validated/

  21. BoltzGen: - Hannes Stärk, accessed December 19, 2025, https://hannes-stark.com/assets/boltzgen.pdf

  22. MIT scientists debut a generative AI model that could create molecules addressing hard-to-treat diseases, accessed December 19, 2025, https://news.mit.edu/2025/mit-scientists-debut-generative-ai-model-that-could-create-molecules-addressing-hard-to-treat-diseases-1125

  23. BoltzGen: Toward Universal Binder Design | bioRxiv, accessed December 19, 2025, https://www.biorxiv.org/content/10.1101/2025.11.20.689494v1

  24. Designing Protein and Peptide Binders with BoltzGen - Rowan, accessed December 19, 2025, https://www.rowansci.com/tools/boltzgen

  25. New AI model could speed rare disease diagnosis | Broad Institute, accessed December 19, 2025, https://www.broadinstitute.org/news/new-ai-model-could-speed-rare-disease-diagnosis

  26. Debbie Marks Lab, accessed December 19, 2025, https://www.deboramarkslab.com/

  27. Proteome-wide model for human disease genetics - medRxiv, accessed December 19, 2025, https://www.medrxiv.org/content/10.1101/2023.11.27.23299062v2.full-text

  28. Deep generative modeling of the human proteome reveals over a hundred novel genes involved in rare genetic disorders - PMC, accessed December 19, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC10802723/

  29. New AI model popEVE predicts harm from ultra-rare DNA mutations - AADC News, accessed December 19, 2025, https://aadcnews.com/news/new-ai-model-popeve-predicts-harm-ultra-rare-dna-mutations/

  30. Protein Engineering with AI: OpenFold3 vs Boltz 2 vs AlphaFold 3 - Blackthorn AI, accessed December 19, 2025, https://blackthorn.ai/blog/protein-engineering-with-ai/

  31. (PDF) Protein design and variant prediction using autoregressive generative models, accessed December 19, 2025, https://www.researchgate.net/publication/351080967_Protein_design_and_variant_prediction_using_autoregressive_generative_models

  32. AlphaFold3: An Overview of Applications and Performance Insights - PMC - PubMed Central, accessed December 19, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC12027460/

  33. 6 ways AI reshaped scientific software in 2025 - R&D World, accessed December 19, 2025, https://www.rdworldonline.com/6-ways-ai-reshaped-scientific-software-in-2025/

  34. OpenFold Consortium, accessed December 19, 2025, https://openfold.io/

  35. Boltz-ABFE: Free Energy Perturbation without Crystal Structures - arXiv, accessed December 19, 2025, https://arxiv.org/html/2508.19385v1

  36. Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction - PMC - PubMed Central, accessed December 19, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC12262699/

  37. Learn to Use Boltz-2 for Docking and Affinity Predictions. - YouTube, accessed December 19, 2025, https://www.youtube.com/watch?v=w3pQcr_KpMI

  38. AI model ranks genetic variants from severe to mild disease mutations - News-Medical.Net, accessed December 19, 2025, https://www.news-medical.net/news/20251203/AI-model-ranks-genetic-variants-from-severe-to-mild-disease-mutations.aspx

  39. Thermodynamics of Biological Processes - PMC - PubMed Central, accessed December 19, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC3264492/

  40. Achievements and challenges in structural bioinformatics and computational biophysics - Oxford Academic, accessed December 19, 2025, https://academic.oup.com/bioinformatics/article/31/1/146/2366253

  41. openfold3 Model by Openfold - NVIDIA NIM APIs, accessed December 19, 2025, https://build.nvidia.com/openfold/openfold3/modelcard

  42. (PDF) Deep generative modeling of the human proteome reveals over a hundred novel genes involved in rare genetic disorders - ResearchGate, accessed December 19, 2025, https://www.researchgate.net/publication/376011842_Deep_generative_modeling_of_the_human_proteome_reveals_over_a_hundred_novel_genes_involved_in_rare_genetic_disorders

  43. OpenFold3 Released Redefining The Limits Of Protein Prediction Models, accessed December 19, 2025, https://www.drugdiscoveryonline.com/doc/openfold-released-redefining-the-limits-of-protein-prediction-models-0001

  44. Training Federated AI Models to Predict Protein Properties | NVIDIA Technical Blog, accessed December 19, 2025, https://developer.nvidia.com/blog/training-federated-ai-models-to-predict-protein-properties/

  45. Why do some predicted protein structures fold poorly? Benchmarking AlphaFold, ESMFold, and Boltz in maize | bioRxiv, accessed December 19, 2025, https://www.biorxiv.org/content/10.1101/2025.07.05.663230v1.full-text

  46. AlphaFold 3, Demystified: I Wrote a Technical Breakdown of Its Complete Architecture., accessed December 19, 2025, https://www.reddit.com/r/bioinformatics/comments/1l7xcp3/alphafold_3_demystified_i_wrote_a_technical/

  47. Machine Learning for RNA-Targeting Drug Design - arXiv, accessed December 19, 2025, https://arxiv.org/html/2512.15645v1

  48. Boltzina: Efficient and Accurate Virtual Screening via Docking-Guided Binding Prediction with Boltz-2 - arXiv, accessed December 19, 2025, https://arxiv.org/html/2508.17555v1

  49. Molecular Biology Institute of Barcelona - ResearchGate, accessed December 19, 2025, https://www.researchgate.net/institution/Molecular_Biology_Institute_of_Barcelona

  50. Activation mechanism of a human SK-calmodulin channel complex elucidated by cryo-EM structures - ResearchGate, accessed December 19, 2025, https://www.researchgate.net/publication/324933856_Activation_mechanism_of_a_human_SK-calmodulin_channel_complex_elucidated_by_cryo-EM_structures

Comments

bottom of page