AlphaFold3, SUTRA, Plot2Code, DrEureka & Is Sora a World Sim?

Week 2, May 2024

Greetings,

Welcome to the latest edition of the State of AI! In this issue, we explore the innovative strides being made in the field of artificial intelligence. You’ll discover how AlphaFold 3 is revolutionizing the accurate structure prediction of biomolecular interactions, take a deep dive into the scalable architecture of the multilingual language model SUTRA, and uncover the groundbreaking potential of Plot2Code for generating code from scientific plots using multi-modal large language models.

We also turn our focus to DrEureka and how language models can guide sim-to-real transfer, transforming the realm of simulations. Lastly, we ponder the capabilities of Sora, examining if it truly stands as a world simulator within the context of general world models and beyond.

Each of these exciting advancements illuminates the diverse and dynamic nature of AI, promising a captivating and thought-provoking read.

Enjoy!

Best regards,

State of AI

---

Contents

1. Accurate structure prediction of biomolecular interactions with AlphaFold 3

2. SUTRA: Scalable Multilingual Language Model Architecture

3. Plot2Code: A Comprehensive Benchmark for Evaluating Multi-modal Large Language Models in Code Generation from Scientific Plots

4. DrEureka: Language Model Guided Sim-To-Real Transfer

5. Is Sora a World Simulator? A Comprehensive Survey on General World Models and Beyond

---

Accurate Structure Prediction of Biomolecular Interactions with AlphaFold 3

Authors: Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, Lindsay Willmore, Andrew J. Ballard, Joshua Bambrick, Sebastian W. Bodenstein, David A. Evans, Chia-Chun Hung, Michael O’Neill, David Reiman, Kathryn Tunyasuvunakool, Zachary Wu, Akvilė Žemgulytė, Eirini Arvaniti, Charles Beattie, Ottavia Bertolli, Alex Bridgland, Alexey Cherepanov, Miles Congreve, Alexander I. Cowen-Rivers, Andrew Cowie, Michael Figurnov, Fabian B. Fuchs, Hannah Gladman, Rishub Jain, Yousuf A. Khan, Caroline M. R. Low, Kuba Perlin, Anna Potapenko, Pascal Savy, Sukhdeep Singh, Adrian Stecula, Ashok Thillaisundaram, Catherine Tong, Sergei Yakneen, Ellen D. Zhong, Michal Zielinski, Augustin Žídek, Victor Bapst, Pushmeet Kohli, Max Jaderberg, Demis Hassabis & John M. Jumper

Source and references: https://www.nature.com/articles/s41586-024-07487-w

---

Introduction

AlphaFold's previous iterations have already made significant ripples in the world of protein structure prediction. Now, with AlphaFold 3, the game has changed yet again. This new version of their model showcases an impressive ability to predict complex interactions between proteins and various other biomolecules, creating a unified tool for the myriad interactions underpinning biological systems.

Welcome to the Next Generation of Molecular Modelling

The past few years have been revolutionary for protein structure prediction. AlphaFold's first and second versions amazed the scientific community with their unprecedented accuracy. They leveraged deep learning to solve one of biology's most significant challenges: predicting a protein’s 3D structure from its amino acid sequence.

Now comes AlphaFold 3, a step further into the future. Google DeepMind's latest iteration isn't just about protein structures anymore. The team aimed at a broader spectrum, targeting nearly every molecule you might find in the Protein Data Bank. The model now tackles not just proteins but also nucleic acids, small molecules, ions, and even modified residues.

The Technical Marvel Behind the Scenes

So, what makes AlphaFold 3 tick? It’s a substantial upgrade over its predecessors in terms of architecture and methodology.

The most notable innovation in AlphaFold 3 is its diffusion-based architecture. Unlike previous models, which relied heavily on multiple sequence alignment (MSA) processing through the Evoformer module, AlphaFold 3 introduces the Pairformer Module. This module is both simpler and more effective, reducing data input requirements while improving performance.

But the real magic happens in the Diffusion Module. This part of AlphaFold 3 directly predicts atomic coordinates from a noisy input structure, refining it repeatedly until it converges on a high-accuracy prediction. The training process for this module is generative, meaning it can handle a diverse array of molecules without needing special adjustments.

Effortless Accuracy Across the Board

AlphaFold 3's prowess comes through most strikingly in its performance benchmarks. The model was put through its paces against specialized tools and consistently outperformed them.

When it comes to protein-ligand interactions, the model leaves traditional docking methods in the dust. It doesn’t even need structural input to achieve this, as it triumphs purely based on the amino acid sequence and ligand information.

It’s not just protein-ligand interactions where AlphaFold 3 shines. The model also demonstrates exceptional accuracy in predicting protein-nucleic acid interactions, outperforming leading nucleic-acid-specific tools. Even complex tasks like antibody-antigen interactions, a challenging domain even for state-of-the-art models, are tackled with higher fidelity by AlphaFold 3.

Generalist Yet Focused

Specialized tools have traditionally been the go-to for specific interaction types. However, AlphaFold 3 defies this norm by being a generalist model that doesn’t compromise on accuracy. It handles a vast array of biomolecular complexes, and in all but one interaction category, it surpasses the performance of domain-specific tools.

This versatility stems from AlphaFold 3’s generative diffusion approach, which produces a distribution of possible structures. This allows the model to fine-tune local structures and maintain high fidelity in steric and bonding details, something previous methods struggled with.

Tackling Technical Challenges

Generative models like the one used in AlphaFold 3 come with their own set of challenges, the most notorious being hallucination—where the model invents plausible-looking structures in regions that might be unstructured. To combat this, the team introduced cross-distillation methods, training AlphaFold 3 with structures predicted by AlphaFold-Multimer v2, which typically represent unstructured areas as extended loops.

Another critical innovation is the confidence measure, a tool for estimating the level of accuracy in its predictions. This feature was essential in the previous AlphaFold models and continued to play a significant role here by providing a fail-safe against errant predictions.

Real-World Applications and Implications

The real-world implications of AlphaFold 3 are immense. Understanding the structure of biomolecular complexes is crucial for advances in drug design, enzymatic research, and the broader field of synthetic biology. AlphaFold 3 doesn’t just map out these structures; it also helps predict how these molecules interact with each other in precise detail.

Drug discovery, for example, stands to benefit enormously. Traditional docking methods are often slow and less accurate, but AlphaFold 3’s ability to predict ligand binding with high fidelity allows for faster and more precise identification of potential drug candidates.

Looking Towards the Future

AlphaFold 3 is more than just an incremental update; it’s a leap forward. Its successes suggest that high-accuracy modeling across various biomolecular interactions is not just feasible but can be achieved within a single, unified deep learning framework.

The future promises further refinements and updates, potentially pushing the boundaries of what’s possible even further. Imagine a world where predicting complex biochemical interactions is as commonplace as sequencing DNA—AlphaFold 3 is a significant step towards that reality.

In conclusion, AlphaFold 3 stands as a testament to how far we’ve come in biomolecular modeling and offers a tantalizing glimpse of the extraordinary possibilities that lie ahead.

---

SUTRA: Scalable Multilingual Language Model Architecture

Authors: Abhijit Bendale, Michael Sapienza, Steven Ripplinger, Simon Gibbs, Jaewon Lee, Pranav Mistry

Source and references: arXiv:2405.06694

---

Introduction

In their groundbreaking paper, Bendale et al. introduce SUTRA, a new multilingual Large Language Model (LLM) architecture designed to understand, reason, and generate text in over 50 languages. Leveraging a unique design that decouples core conceptual understanding from language-specific processing, SUTRA aims to deliver scalable, efficient, and highly responsive multilingual capabilities. This novel approach exhibits substantial improvements over existing models like GPT-3.5 and Llama2, particularly in multilingual settings, marking a significant leap in the field of AI and natural language processing.