Inside AlphaProteo, Google DeepMind’s New Model for Next Generation Protein Design

Author(s): Jesus Rodriguez

Originally published on Towards AI.

Inside AlphaProteo, Google DeepMind’s New Model for Next Generation Protein Design

I recently started an AI-focused educational newsletter, that already has over 170,000 subscribers. TheSequence is a no-BS (meaning no hype, no news, etc) ML-oriented newsletter that takes 5 minutes to read. The goal is to keep you up to date with machine learning projects, research papers, and concepts. Please give it a try by subscribing below:

Google DeepMind has been at the forefront of using AI for protein design for the last decade. DeepMind’s AlphaFold model set new milestones in protein design and its been widely adopted by the research community. Beyond the structure of proteins, one key area of research is their interactions patterns which are essential to understanding functions required in drug development. To address this challenge, scientists have created proteins that “bind” to specific molecules streamlining that manipulation process. DeepMind doubled down in this area with the recent publication of AlphaProteo, an AI model poised to transform protein binder design.

Proteins, the workhorses of biology, interact in a delicate dance, dictating countless cellular processes. The ability to design proteins that bind specific targets, termed “protein binders,” opens doors in research, diagnostics, and therapeutics. Imagine blocking a disease-causing protein interaction or directing an enzyme to a specific cellular location — the possibilities are vast.

Traditional protein binder design, however, is laborious and time-consuming. Imagine finding a needle in a haystack, except you’re also designing the needle! Techniques like immunization or directed evolution rely on trial-and-error, often requiring extensive screening and optimization. AlphaProteo offers a radical departure, harnessing the power of machine learning to predict and design these intricate interactions with unprecedented speed and efficiency.

AlphaProteo

AlphaProteo isn’t a single entity but rather a sophisticated system with two core components:

1. Generative Model: This model, trained on a massive trove of protein structures and sequences from the PDB (Protein Data Bank) and augmented with AlphaFold predictions, acts as the creative engine of the system. It takes the target protein’s structure, and optionally, “hotspot” residues representing the desired binding site, and generates a multitude of candidate binder structures and sequences.
2. Filter: With a plethora of potential binders in hand, the filter steps in to separate the wheat from the chaff. It evaluates each candidate, predicting its likelihood of successfully binding to the target. This crucial step ensures that only the most promising designs advance to experimental validation.

A Triumphant Debut: Experimental Validation

The report meticulously details the experimental validation of AlphaProteo, showcasing its ability to design high-affinity binders for a diverse set of eight target proteins. These targets weren’t chosen randomly; they represent a range of design difficulties and biological significance.

Success Rates

The first litmus test for AlphaProteo was its experimental success rate — the proportion of designed binders that actually bound their targets in laboratory assays. The results were nothing short of impressive. For seven out of the eight targets, AlphaProteo achieved success rates ranging from 9% to a remarkable 88%. This means that for some targets, nearly 9 out of every 10 designs successfully bound!

To put this in perspective, these success rates surpass the best existing design methods by a significant margin. In the case of IL-17A, a protein involved in inflammatory responses, AlphaProteo achieved a success rate 700 times higher than any previously reported method. This dramatic leap in efficiency has the potential to drastically reduce the time and resources needed for protein binder development.

High Affinity: The Hallmark of a Strong Binder

A high success rate is just one part of the equation. For a binder to be truly useful, it needs to bind its target tightly, a property quantified by its binding affinity, often represented as the dissociation constant (KD). The lower the KD, the stronger the binding.

AlphaProteo again shines in this regard. The report highlights that for four of the targets, the best binders designed by AlphaProteo had KD values in the picomolar range (pM) — a testament to their incredibly tight binding. This achievement is particularly remarkable when compared to previously designed binders, where AlphaProteo consistently produced binders with significantly higher affinity.

Precision Targeting and Specificity

The researchers didn’t stop at demonstrating binding; they went on to confirm the precision and specificity of the designed binders. Through clever experiments, they verified that the binders interacted with the intended “hotspot” regions on the target proteins, just as they were designed to do.

Furthermore, specificity tests revealed that the binders were highly selective, exhibiting negligible binding to unintended targets. This high specificity is crucial for many applications, especially therapeutic development, where off-target binding can lead to undesirable side effects.

AlphaProteo in Action: Inhibiting Viral Infection and Cell Signaling

DeepMind’s AlphaProteo offers tantalizing glimpses into AlphaProteo’s potential applications. One striking example is the design of binders against the SARS-CoV-2 receptor binding domain (SC2RBD), the very protein responsible for the virus’s entry into human cells.

AlphaProteo successfully generated binders that potently neutralized several SARS-CoV-2 variants, effectively blocking viral infection in laboratory assays. This achievement highlights the system’s potential in rapidly developing antiviral therapies against emerging viral threats.

In another compelling demonstration, researchers used AlphaProteo to design binders that effectively inhibited the activity of VEGF-A, a protein implicated in the formation of new blood vessels (angiogenesis). By blocking VEGF-A, these binders have the potential to disrupt tumor growth, offering a promising avenue for developing novel cancer therapies.

Peering into the Future of Protein Design

While the initial successes of AlphaProteo are undeniable, the researchers acknowledge that this is just the tip of the iceberg. The system, though powerful, is still under development, and several exciting avenues for future research are outlined.

Expanding the Repertoire of Targetable Proteins

One area of focus is expanding AlphaProteo’s capabilities to handle even more challenging protein targets. Currently, the system relies on the availability of a target protein’s structure. However, many proteins, particularly those embedded within cell membranes or those that rapidly change their shape, lack readily available structures. Developing methods to overcome these limitations will unlock a vast new world of potential targets for AlphaProteo.

Beyond Binding: Engineering Functionality

The current focus of AlphaProteo is on designing high-affinity binders. However, the report hints at a future where the system’s capabilities extend beyond mere binding, towards engineering proteins with specific functionalities. Imagine designing proteins that act as biosensors, detecting specific molecules in their environment, or enzymes with tailor-made catalytic activities — the possibilities are endless.

A New Era of Protein Design

AlphaProteo represents a paradigm shift in protein design. Its innovative approach, combining the power of machine learning with structural biology, offers a glimpse into a future where designing protein binders is no longer a daunting task but rather a streamlined and efficient process. The implications are far-reaching, with the potential to revolutionize research, diagnostics, and therapeutic development.

It’s important to remember that the information presented here originates solely from the provided technical report. Further independent research may be necessary to verify these findings and explore the full breadth of AlphaProteo’s capabilities and limitations.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI