Protein Structure Prediction in the Post-AlphaFold Era

AlphaFold changed structural biology irreversibly. By demonstrating that a deep learning system could predict protein structures from amino acid sequences at experimental accuracy, DeepMind's work closed a problem that had defined the field for fifty years. The AlphaFold Protein Structure Database now contains predicted structures for over 200 million proteins, providing the structural biology community with an extraordinary resource. This achievement deserves unqualified recognition. But it is essential to understand precisely what was solved and what remains open, because the unsolved problems are where the next decade of biological discovery — and therapeutic development — will be won. AlphaFold predicts a single static structure for a given sequence, typically the most thermodynamically stable conformation under crystallization-like conditions. Proteins, however, are not static objects. They are dynamic molecular machines whose function depends on motion, conformational change, and the ability to adopt multiple distinct states.

The first major gap is protein dynamics. A kinase that toggles between active and inactive conformations, a transporter that alternates between inward-facing and outward-facing states, a signaling protein that allosterically couples distant binding events — these systems cannot be understood from a single structure. Their biology is encoded in the energy landscape that governs transitions between states, the timescales of those transitions, and how ligands, post-translational modifications, and binding partners reshape that landscape. AlphaFold's confidence metric (pLDDT) implicitly flags flexible regions by assigning them low scores, but this is a byproduct, not a prediction of dynamics. At Webbeon, we have developed extensions to ArcOne that directly predict conformational ensembles from sequence. Our approach trains on molecular dynamics trajectory data and experimentally determined ensemble structures (from NMR, cryo-EM multi-state refinements, and hydrogen-deuterium exchange mass spectrometry) to produce not a single structure but a distribution over conformations weighted by their Boltzmann probabilities. For a benchmark set of 147 proteins with experimentally characterized conformational changes, ArcOne's ensemble predictions capture the major alternative states for 78% of cases, compared to 12% for AlphaFold's single-structure output.

The second gap is protein-protein interactions. Most biological functions are carried out by protein complexes, not individual chains. AlphaFold-Multimer made progress on predicting complex structures, but its accuracy drops substantially for transient interactions, low-affinity complexes, and interactions mediated by intrinsically disordered regions (IDRs). IDRs — which account for roughly 30% of the human proteome — are particularly challenging because they do not adopt fixed structures in isolation; they fold upon binding, or remain disordered even in complex, forming "fuzzy" interactions that defy conventional structural description. Our approach to IDR interactions abandons the attempt to predict a single bound structure and instead predicts the ensemble of bound-state conformations along with their interaction energetics. ArcOne integrates sequence-level features (evolutionary couplings, predicted disorder propensity, short linear motif identification) with a physics-informed energy function that captures the entropic contributions unique to IDR binding. On a curated benchmark of 85 IDR-mediated interactions with experimentally measured binding affinities, our predictions achieve a Pearson correlation of 0.71 with experiment — a substantial advance over physics-based free energy calculations (0.45) and previous ML approaches (0.52).

The third gap is arguably the most consequential: connecting structure to biological function. Knowing that a protein folds into a TIM barrel tells you its topology but not whether it catalyzes an aldol condensation or an isomerization. Structure prediction gives us the "what" of molecular shape; the field now needs the "why" of molecular behavior. We are addressing this through multi-task models that jointly predict structure, dynamics, and functional annotations, enforcing consistency between these predictions rather than treating them as independent outputs. When ArcOne predicts that a protein has a flexible loop near a conserved catalytic triad, it simultaneously infers the likely enzymatic mechanism and substrate specificity, because these predictions are coupled through shared latent representations trained on the integrated experimental record of structure-function relationships.

Our work is not a replacement for AlphaFold; it is built on the foundation that AlphaFold established. We use AlphaFold-predicted structures as initialization for our dynamics models, as templates for our interaction predictions, and as structural context for our functional annotations. The post-AlphaFold era is not about surpassing a single benchmark — it is about expanding the scope of computational biology from static structure to the full complexity of protein behavior. The problems that remain are harder, messier, and more biologically relevant than fold prediction. They require models that reason about energy, motion, and molecular logic, not just spatial coordinates. This is the frontier we are pursuing.