IntFold: Deep Learning for Biomolecular Structures

• IntFold is a deep learning model featuring a four-stage pipeline and a custom FlashAttentionPairBias kernel to achieve high atomic accuracy and efficient prediction.
• It employs modular adapter mechanisms, such as LoRA and structural constraint adapters, to allow precise guidance in folding, binding affinity estimation, and structure ranking.
• Benchmark results show state-of-the-art performance in docking, affinity prediction, and structure ranking, enhancing applications in structural biology and drug design.

IntFold is a foundation-scale, controllable deep learning model for general and specialized biomolecular structure prediction, designed to combine end-to-end atomic accuracy with modular, adapter-driven mechanisms for downstream specialization. IntFold employs a custom high-performance attention kernel and supports guided folding, binding affinity estimation, and structure ranking, enabling applications across structural biology and drug discovery through precise user intervention (Team et al., 2 Jul 2025).

1. Architecture and Core Computational Modules

IntFold’s architecture follows a four-stage pipeline. Input biomolecular sequences (amino-acid or nucleic-acid), optionally augmented with multiple sequence alignments (MSAs) and coordinate templates, are projected into token and pairwise feature representations through an embedding trunk. These are processed by the “Pairformer” stack—alternating transformer-like layers over both single residues and residue–residue pair features for long-range effective modeling.

A key innovation is the FlashAttentionPairBias kernel, which integrates dot-product attention with a learnable pairwise bias. Specifically, attention weights for each head are computed as:

$$\alpha_{ij} = \mathrm{softmax}_j \left( \frac{Q_i\cdot K_j^T}{\sqrt{d_k}} + B_{ij} \right), \quad \mathrm{Attention}(Q, V) = \sum_{j=1}^L \alpha_{ij} V_j.$$

This operator fuses the dot product, bias addition, softmax, and aggregation into a single custom GPU kernel, achieving up to 1.8× speedup and ~30% memory reduction relative to DeepSpeed or standard NVIDIA implementations. Following the Pairformer trunk, a coordinate diffusion module iteratively refines all-atom positions, with a confidence head predicting pLDDT and pTM metrics for model reliability (Team et al., 2 Jul 2025).

2. Training Objectives and Dataset Composition

Training utilizes a composite loss function:

$$L = \lambda_\text{FAPE} L_\text{FAPE} + \lambda_\text{diff} L_\text{diff} + \lambda_\text{dist} L_\text{dist} + \lambda_\text{ang} L_\text{ang} + \lambda_\text{conf} L_\text{conf},$$

where

• $L_\text{FAPE}$ is the Frame-Aligned Point Error for atomic accuracy post rigid-body frame alignment,
• $L_\text{diff}$ is an expected squared error for diffusion-based atomic reconstruction,
• $L_\text{dist}$ and $L_\text{ang}$ are cross-entropy losses for inter-residue distances and backbone torsion bins,
• $L_\text{conf}$ is cross-entropy over pLDDT and pTM bins.

The core training dataset comprises ~1 million entries from the PDB (proteins, nucleic acids, and complexes, mmCIF format, with ions and low-confidence ligands filtered), augmented with three distilled sets:

• Monomer predictions from the AlphaFold database (filtered for pLDDT >85, ≤30% sequence identity, ≥0.5L long-range contacts),
• Disordered-region assemblies from AlphaFold-Multimer v2.3,
• Antibody–antigen interfaces from PLAbDab (60% identity clustering, interface diversity selection) and IntFold-predicted self-distillation.

For affinity tasks, IntFold leverages an integrated dataset of ~3 million log-scaled IC₅₀, $K_i$, and $K_d$ measurements from ChEMBL, BindingDB, GalaxyDB, BioLip, and PubChem, harmonized using ExCAPE-DB protocols. Data augmentation strategies include cropping, sequence masking, MSA/template masking, Gaussian noise for diffusion robustness, and balanced fine-tuning for IntFold+ (Team et al., 2 Jul 2025).

3. Controllability and Modular Adaptation

A principal advancement is IntFold’s adapter-based controllability. The main model (>1B parameters) remains frozen during specialization and task steering is introduced via lightweight, pluggable modules:

• LoRA Adapters: In each Pairformer block, low-rank updates $\Delta W = AB^\top$ (with $$L = \lambda_\text{FAPE} L_\text{FAPE} + \lambda_\text{diff} L_\text{diff} + \lambda_\text{dist} L_\text{dist} + \lambda_\text{ang} L_\text{ang} + \lambda_\text{conf} L_\text{conf},$$0) are inserted, adding only ~1% extra parameters per task. Fine-tuning on ~200 CDK2-inhibitor complexes (10 allosteric, ~190 orthosteric) enables allosteric state recovery: the base model predicts 0/5 closed allosteric states on a 40-target holdout; the LoRA-adapted model recovers 4/5 while maintaining 35/35 orthosteric predictions.
• Structural Constraint Adapters: An interface embedder encodes binary residue masks into pairwise features, supporting explicit residue-level constraints. Application in guided docking increases PoseBusters pose success from 79.5% to 89.7%, and antibody–antigen docking accuracy from 37.6% to 69.0%.
• Post-hoc Affinity Module: For binding affinity prediction, the model exports final representation tensors to a 4-block Pairformer “affinity head,” yielding continuous fitness predictions. Fine-tuning on affinity datasets achieves AUPR of 0.74 on Davis and 0.69 on BindingDB, surpassing Boltz-2 and sequence-only baselines (Team et al., 2 Jul 2025).

4. Prediction Ranking and Consensus Selection

Each structure prediction target is sampled stochastically ($$L = \lambda_\text{FAPE} L_\text{FAPE} + \lambda_\text{diff} L_\text{diff} + \lambda_\text{dist} L_\text{dist} + \lambda_\text{ang} L_\text{ang} + \lambda_\text{conf} L_\text{conf},$$1, from 5 diffusion seeds × 5 random initializations). Pairwise DockQ scores $$L = \lambda_\text{FAPE} L_\text{FAPE} + \lambda_\text{diff} L_\text{diff} + \lambda_\text{dist} L_\text{dist} + \lambda_\text{ang} L_\text{ang} + \lambda_\text{conf} L_\text{conf},$$2 are computed, and the predicted structure with the highest mean similarity is selected by:

$$L = \lambda_\text{FAPE} L_\text{FAPE} + \lambda_\text{diff} L_\text{diff} + \lambda_\text{dist} L_\text{dist} + \lambda_\text{ang} L_\text{ang} + \lambda_\text{conf} L_\text{conf},$$3

This consensus, training-free ranking scheme is model-agnostic. On antibody–antigen benchmarks, it increases selection of high-quality decoys by ~3% over random, without retraining model confidence heads (Team et al., 2 Jul 2025).

5. Benchmark Performance and Quantitative Evaluation

IntFold exhibits state-of-the-art all-atom accuracy across comprehensive FoldBench tasks, matching or exceeding major prior models, particularly in challenging interface and ligand-binding scenarios. Summary statistics include:

Task	IntFold	IntFold+	AlphaFold 3	Boltz-1	Chai-1
Protein monomer LDDT	0.88	—	0.88	—	—
Protein–protein (%)	72.9	—	72.9	—	68.5
Antibody–antigen (%)	37.6	43.2	47.9	—	—
Protein–ligand (%)	58.5	61.8	64.9	55.0	—
Protein–DNA (%)	74.1	—	79.2	71.0	—
Protein–RNA (%)	58.9	—	62.3	56.9	—
RNA monomer LDDT	0.63	—	0.61	—	—
DNA monomer LDDT	0.50	—	0.53	—	—

Against Boltz-2 on post-2024 targets, IntFold achieves higher rates: protein–ligand 58.17% vs 53.90%, protein–protein 72.13% vs 70.54%, and antibody–antigen 40.27% vs 25.00%. Binding affinity correlation on CASP16 L1000 reaches Pearson $$L = \lambda_\text{FAPE} L_\text{FAPE} + \lambda_\text{diff} L_\text{diff} + \lambda_\text{dist} L_\text{dist} + \lambda_\text{ang} L_\text{ang} + \lambda_\text{conf} L_\text{conf},$$4 vs Boltz-2’s $$L = \lambda_\text{FAPE} L_\text{FAPE} + \lambda_\text{diff} L_\text{diff} + \lambda_\text{dist} L_\text{dist} + \lambda_\text{ang} L_\text{ang} + \lambda_\text{conf} L_\text{conf},$$5, remaining comparable on L3000 (Team et al., 2 Jul 2025).

6. Applications in Drug Design and Structural Biology

IntFold supports specialized applications via controllable adapters:

• Allosteric inhibitor modeling (e.g., CDK2): LoRA-adapted model recovers rare closed-pocket conformers involved in selective inhibition.
• Pocket/epitope-guided docking: Interface residue constraints can steer antibody–antigen docking to correct DockQ > 0.5 poses.
• Virtual screening: Integration of IntFold+ structure generation with the affinity head increases hit enrichment by ~30% over Boltz-2 in retrospective settings.

These features enable rapid hypothesis testing, lead optimization, and antibody engineering with explicit structural priors central to modern drug discovery (Team et al., 2 Jul 2025).

7. Implementation Details and User Deployment

The full IntFold model contains ~1.2 billion parameters. Training utilizes mixed precision (bfloat16 for backbone, float32 for diffusion) on the Jinghe Cloud platform. Inference requires approximately 4 GB GPU memory per prediction and scales quadratically with sequence length.

Access is provided via REST API and Python SDK (https://server.intfold.com/), supporting submission of sequences, MSAs/templates, and JSON-encoded constraints. Adapter fine-tuning for new tasks (e.g., novel conformers or binding pockets) is streamlined with dedicated automation scripts (Team et al., 2 Jul 2025).

In summary, IntFold integrates scalable accuracy with precise, user-driven control, extending foundation model methodology to versatile downstream structure prediction and drug design scenarios.