Command Palette
Search for a command to run...
IsoDDE를 이용한 새로운 생체 분자 상호작용의 정확한 예측
IsoDDE를 이용한 새로운 생체 분자 상호작용의 정확한 예측
Isomorphic Labs Team
초록
생체 분자 상호작용을 예측하는 것은 합리적 신약 설계(rational drug design)의 핵심이지만, 실험적 정확도를 확보하고 새로운 화학 공간(chemical space)에 대해 일반화(generalisability)를 달성하는 것은 여전히 중대한 병목 현상으로 남아 있습니다. AlphaFold 3와 같은 딥러닝 방식이 구조 예측 분야를 발전시켰음에도 불구하고, benchmark 결과에 따르면 미탐사 분자 공간 영역으로의 일반화, 결합 친화도(binding affinity) 추정, 그리고 이전에 특성이 밝혀지지 않은 단백질 표면에서의 분자 결합 부위 탐지 측면에서는 여전히 한계가 존재함이 드러났습니다.본 논문에서는 이러한 한계를 해결하기 위해 설계된 통합 계산 시스템인 Isomorphic Labs Drug Design Engine (IsoDDE)를 소개합니다. 당사는 까다로운 단백질-리간드 일반화 benchmark에서 IsoDDE가 AlphaFold 3보다 정확도를 두 배 이상 높였음을 입증하였으며, 유도 적합(induced fits)과 같은 복잡한 out-of-distribution 이벤트를 성공적으로 모델링하고 새로운 결합 포켓(binding pockets)을 정확하게 식별해냈습니다. 바이오 의약품(biologics) 분야에서 IsoDDE는 기존 모델들을 크게 능가하며, 항체-항원 인터페이스 예측 및 CDR-H3 루프 모델링 분야에서 새로운 state of the art를 제시합니다. 마지막으로, 저분자 결합제(small molecule binders)의 경우, IsoDDE의 친화도 예측은 골드 스탠다드인 물리 기반 방식(physics-based methods)을 능가하며, 기존 물리 기반 워크플로의 과도한 계산 비용(computational overhead) 없이도 실험실 수준의 정밀도에 도달하여 그 간극을 메웠습니다.본 연구 결과는 IsoDDE가 AI 신약 설계의 확장 가능한 기반을 제공하며, 전례 없는 정확도로 새로운 생물학적 시스템을 탐색하는 데 필요한 예측 충실도(predictive fidelity)를 제공함을 입증합니다.
One-sentence Summary
The Isomorphic Labs Drug Design Engine (IsoDDE) is a unified computational system that outperforms AlphaFold 3 on protein-ligand generalization benchmarks by modeling complex out-of-distribution events like induced fits, achieving state-of-the-art accuracy in antibody-antigen interface and CDR-H3 loop modeling, and providing small molecule binding affinity predictions that exceed gold-standard physics-based methods.
Key Contributions
- The paper introduces the Isomorphic Labs Drug Design Engine (IsoDDE), a unified computational system that achieves a step change in unconditional accuracy for biomolecular interaction prediction. This architecture demonstrates the ability to more than double the accuracy of AlphaFold 3 on protein-ligand generalization benchmarks and successfully models complex out-of-distribution events like induced fits.
- This work presents a high-performance model for the biologics domain that establishes a new state of the art in antibody-antigen interface prediction and CDR-H3 loop modeling. The system also enables blind pocket identification, allowing for the discovery of novel binding sites on previously uncharacterized protein surfaces without requiring a specified ligand.
- The research provides a quantitative binding affinity estimation capability that outperforms gold-standard physics-based methods for small molecule binders. This approach bridges the gap toward experimental-grade precision while avoiding the high computational overhead associated with traditional free energy perturbation workflows.
Introduction
Accurate prediction of biomolecular interactions is essential for rational drug design and the discovery of new biological mechanisms. While deep learning models like AlphaFold 3 have advanced structure prediction, they often struggle to generalize to unexplored chemical spaces, fail to identify novel binding pockets, and lack the ability to provide precise quantitative binding affinity. Current methods for estimating affinity are either computationally expensive physics-based simulations or deep learning models that correlate poorly with experimental data. The authors introduce the Isomorphic Labs Drug Design Engine (IsoDDE), a unified computational system that achieves a step change in accuracy and generalization. IsoDDE more than doubles the accuracy of previous state-of-the-art models on protein-ligand benchmarks, provides superior antibody-antigen interface modeling, and delivers binding affinity predictions that exceed gold-standard physics-based methods.
Experiment
IsoDDE was evaluated across multiple benchmarks, including Runs N' Poses, FoldBench, and antibody-antigen datasets, to assess its structural modeling and generalization capabilities. The results demonstrate that the model significantly outperforms previous state-of-the-art methods like AlphaFold 3 and Boltz-2, particularly when predicting highly novel protein-ligand interfaces and complex antibody-antigen structures. Furthermore, IsoDDE shows exceptional performance in predicting binding affinities and identifying cryptic pockets, proving to be a robust tool for navigating diverse chemical and biological spaces in drug discovery.
The authors compare the performance of IsoDDE against AF3 across different subsets of the Runs N' Poses benchmark. Results show that IsoDDE achieves a higher success rate than AF3 in the full dataset, the filtered dataset, and the filtered plus clustered dataset. IsoDDE demonstrates a consistent performance advantage over AF3 across all evaluated data subsets The mean gain in success rate remains substantial whether using the full or filtered datasets IsoDDE maintains superior performance even when the data is filtered and clustered to reduce redundancy
The authors evaluate the performance of IsoDDE against Boltz-2 across different data subsets. Results show that IsoDDE maintains a higher success rate than Boltz-2 in the full, filtered, and clustered datasets. IsoDDE demonstrates a substantial mean gain in performance compared to Boltz-2 across all tested sets The model maintains consistent improvements even when data is filtered or clustered Performance advantages are observed in both the full dataset and more specialized subsets
The authors evaluate IsoDDE's structural modeling performance on the FoldBench dataset across three distinct interface types. Results show that IsoDDE achieves superior success rates compared to several external models in antibody-antigen, protein-ligand, and protein-protein benchmarks. IsoDDE demonstrates higher success rates in antibody-antigen interface prediction than all compared models The model shows improved performance in protein-ligand modeling relative to the other evaluated methods IsoDDE maintains a competitive advantage across protein-protein interaction benchmarks
IsoDDE was evaluated against AF3, Boltz-2, and various external models using the Runs N' Poses and FoldBench benchmarks to assess its structural modeling capabilities. Across diverse datasets and interface types, including antibody-antigen, protein-ligand, and protein-protein interactions, IsoDDE consistently demonstrated superior success rates. The results indicate that the model maintains a robust performance advantage even when data is filtered or clustered to reduce redundancy.