TLDR: The IMPRESS framework integrates AI and high-performance computing (HPC) to create an adaptive system for protein design. It uses tools like ProteinMPNN and AlphaFold in an iterative, “evaluate as you go” approach, dynamically allocating resources and executing tasks asynchronously. This leads to higher quality protein designs, better consistency, and significantly improved utilization of computational resources compared to traditional methods.
The field of computational protein design is undergoing a significant transformation, largely driven by advancements in Artificial Intelligence (AI) and Machine Learning (ML). However, the sheer number of possible protein sequences and structures is incredibly vast, making the process of designing new proteins and ensuring their predicted structures match the generated ones a computationally intensive challenge. To address this, researchers have introduced a new framework called the Integrated Machine-learning for Protein Structures at Scale (IMPRESS).
IMPRESS offers a novel approach by combining AI with high-performance computing (HPC) systems. This integration allows for the real-time evaluation of protein designs as they are developed, as well as the models and simulations used to generate data and train these models. The core idea is to enable an “evaluate as you go” methodology, where AI systems can guide HPC tasks and vice versa, creating a bidirectional influence that accelerates scientific discovery.
Traditional computational methods often suffer from inefficiencies such as idle resources and prolonged workflow times. IMPRESS tackles these issues by implementing adaptive decisions that can change the set of tasks being executed. It also supports asynchronous execution and dynamic resource allocation, meaning computational tasks can be adjusted in real-time based on available resources and specific requirements. This leads to better resource utilization, reduced workflow duration, and the ability to execute diverse tasks across different computing platforms.
At its heart, IMPRESS utilizes powerful tools like ProteinMPNN for generating protein sequences based on existing protein backbones, and AlphaFold for predicting the 3D structures of these candidate proteins. As a demonstration, the framework was used to optimize protein binder designs, specifically PDZ domains, for particular peptide targets. Designing these domains for high affinity and selectivity is crucial for drug development.
The IMPRESS Framework in Detail
The IMPRESS framework is designed to enhance protein design by integrating AI-driven generative models with HPC simulations, creating a real-time feedback loop. This not only improves the design and production of proteins but also helps in validating foundational models with experimental data. The framework consists of two main components: a pipelines coordinator and an execution runtime system, RADICAL-Pilot (RP).
The coordinator manages the iterative process of constructing and generating IMPRESS pipelines, submitting independent protein pipeline tasks concurrently, and making adaptive decisions based on the results. It maintains a global view of each pipeline’s outcomes and the quality of the resulting sequences, allowing it to re-process low-quality sequences with new pipelines if needed.
The IMPRESS pipeline itself is a series of stages: (i) generating multiple customizable sequences using ProteinMPNN, (ii) selecting and ranking these sequences, (iii) compiling the top sequences for structural prediction, (iv) predicting structures with AlphaFold and ranking them, (v) gathering quality metrics (like pLDDT, pTM, pAE), (vi) comparing current structure quality to previous iterations and adapting the next steps (e.g., re-running with a different sequence or feeding the improved structure back into ProteinMPNN), and (vii) iteratively cycling through these stages until a final design is returned.
Also Read:
- EvoIF: A Data-Efficient Framework for Protein Fitness Prediction Through Evolutionary Profiles
- DeepEvolve: An AI Agent That Learns and Builds Better Algorithms
Performance and Impact
The researchers evaluated IMPRESS by comparing its adaptive pipelines (IM-RP) against a non-adaptive control version (CONT-V) on an HPC system. The results were compelling. While the control version showed gradual improvement, the IM-RP pipeline achieved superior results in terms of protein quality metrics (higher pLDDT and pTM, lower inter-chain pAE medians) at every iteration. Furthermore, the adaptive protocol demonstrated greater consistency in design quality.
Crucially, IM-RP significantly outperformed CONT-V in resource utilization. IM-RP utilized approximately 88% of available CPU cores and 61% of GPUs, whereas CONT-V only managed about 18.3% CPU and a mere 1% GPU utilization. This under-utilization in CONT-V was attributed to sequential processing and idle GPUs during CPU-intensive phases. IMPRESS’s adaptive design smartly offloads newly created pipelines to idle resources, maximizing efficiency.
This research highlights that while the adaptive approach (IM-RP) might take slightly longer overall due to evaluating more design trajectories, it yields higher quality protein designs and makes much more efficient use of valuable computational resources. The paper’s main contributions include the development and implementation of this adaptive protein design protocol and its supporting computing infrastructure, leading to more consistent protein design quality and enhanced throughput.
Looking ahead, the team plans to generalize this pipeline to other protein design problems, such as improving catalytic activity in proteases, and to further enhance IMPRESS to support the real-time evaluation and optimization of foundational AI models using real-world experimental data. For more details, you can read the full research paper here.
