AlphaFlex

AlphaFlex is a two-stage framework for predicting the multiple conformational states a protein adopts while carrying out its function. Structure predictors such as AlphaFold 2 and ESMFold resolved the static single-structure problem, but they collapse onto one low-energy conformation because the co-evolutionary signal in a multiple sequence alignment (MSA) mixes constraints from every state a protein visits. Enzyme catalysis, allosteric regulation, and substrate transport all depend on the transitions between those states.

The prevailing workaround is to perturb the MSA — clustering it, subsampling it, or masking columns at random — and let AlphaFold 2 produce a spread of structures. AlphaFlex replaces that brute-force perturbation with a biologically directed one. A dedicated deep learning model first reads an input structure, experimental or predicted, and assigns every residue a flexibility probability between 0 and 1. Those probabilities drive column masking of the MSA at exactly the positions expected to move, relaxing co-evolutionary constraints in dynamic regions while leaving those on the structural core intact. The masked sub-MSAs are subsampled to a range of depths and passed through an unmodified AlphaFold 2, which returns a conformational ensemble.

AlphaFlex was developed by Lingyu Ge, Xinyue Cui, Kailong Zhao, Xiaogen Zhou, Yang Zhang, and Guijun Zhang at Zhejiang University of Technology and the National University of Singapore, and posted to bioRxiv in July 2025. It is a preprint awaiting peer review; the authors state that release of the inference code and model weights is in preparation, and that the training data is available from the corresponding author on request.

#Key Features

• Flexibility-guided MSA masking: Rather than masking MSA columns randomly, AlphaFlex masks contiguous fragments predicted to be highly flexible, using a sliding window of three residues that steps inward from both ends of each region.
• AlphaFold 2 left untouched: The sampling stage feeds ordinary sub-MSAs into a stock AlphaFold 2 with no retraining or architectural change.
• Multimodal flexibility predictor: Per-residue predictions combine geometric descriptors, BLOSUM62 conservation, Rosetta energy terms, MSA Transformer (ESM-MSA-1b) embeddings, and structural profiles built from remote homologs in the AlphaFold Database.
• Membrane transporter states: On a non-redundant set of five transporters, AlphaFlex resolves inward-facing and outward-facing conformations with comparable accuracy.
• Transition intermediates: Ensembles populate continuous paths between apo and holo states, sampling intermediates that are difficult to capture experimentally.

#Technical Details

The flexibility predictor fuses structural and physicochemical features in an integration module while MSA embeddings pass through a sequence-reweighting module that scores each aligned sequence. The concatenated features enter three layers of triangle updates and axial attention, then a 64-layer residual network whose blocks each use two 3x3 convolutions with instance normalization and ELU activation, ending in a fully connected layer and sigmoid. Training followed a transfer-learning schedule: pretraining on a PDB-derived set of 608,984 monomers reduced by MMseqs2 clustering to 20,932 non-redundant clusters, then fine-tuning on molecular dynamics trajectories for 1,390 ATLAS proteins sampled at 100 frames each. Optimization used Adam at a 1e-4 learning rate with focal loss on a single NVIDIA A100. At inference a 20% masking ratio was optimal, with sub-MSAs subsampled across depths from 16 to 5,120 sequences.

On 69 curated apo/holo pairs from CoDNaS — averaging 3.6 A RMSD between states — AlphaFlex reached a 42.0% success rate, defined as TM-score above 0.95 against both states, a 16% relative gain over the best baseline. Mean best-model RMSD was 2.008 A (apo) and 1.471 A (holo), against 2.975/2.487 A for AF-Cluster, 2.117/1.554 A for AF2_conformations, and 2.079/2.126 A for AlphaFlow, each generating 1,000 conformations per target. Flexible-residue prediction scored 0.810 ROC-AUC, 0.742 accuracy, and 0.521 F1.

#Applications

The method targets questions where a single static model is insufficient: mapping cryptic and transient pockets for structure-based drug design, characterizing allosteric mechanisms, and modeling the alternating-access cycles of membrane transporters. Case studies span large domain motions in adenylate kinase, mitochondrial HSPD1, DAHP synthase, and the D-ribose-binding protein, plus lid-loop closures in 3-dehydroquinate dehydratase and aspartate transcarbamoylase. Because the pipeline wraps a standard AlphaFold 2 installation, groups already running AF2 can adopt it without changing their structure-prediction stack.

#Impact

AlphaFlex belongs to a growing family of conformational-ensemble methods built on AlphaFold, and its contribution is the argument that where you perturb the alignment matters more than how much you perturb it. Ablations support this directly: swapping guided masking for random masking at the same rate costs roughly 4% accuracy on both states. The limitations are acknowledged. A 42% success rate means most targets still fail the dual-state threshold; the method needs an input structure and 1,000 sampled conformations per target; validation rests on 69 apo/holo pairs and 5 membrane proteins; and the sequence databases used for MSA construction are a known bottleneck. With code, weights, and training data not yet released, the results cannot currently be reproduced independently.