AlphaMut

Helix propensity scales have been measured on isolated peptides for decades: alanine and leucine favor a helix, glycine and proline break one. AlphaMut's own results show how poorly that ranking travels. The propensity scale it derives for helices inside folded proteins correlates with the scale for the same helices in isolation at a Spearman rho of just 0.43 — a substitution that wrecks a free-standing helix can be inconsequential once tertiary contacts hold it in place.

AlphaMut, from Prathith Bhargav and Arnab Mukherjee at IISER Pune, casts the choice of which residues to mutate as a sequential decision problem instead of a lookup. A protein sequence is the state, a point mutation is an action, and the reward comes from folding the mutated sequence and measuring how much of the target helix survived. A policy trained with proximal policy optimization (PPO) learns which combinations of substitutions destroy a helix; because a single mutation is rarely enough, an episode may apply several, and the agent is credited for the set.

Every network in the loop except that policy is frozen and pretrained. ESM-2 at 8M parameters embeds the sequence into the state vector, while ESMFold folds each candidate mutant and the P-SEA algorithm assigns its secondary structure to produce the reward. The framing inverts the usual mutation question: AlphaMissense and stability predictors such as RaSP, ThermoMPNN and DeepDDG score mutations a user supplies, whereas AlphaMut nominates them.

#Key Features

• Mutation sets rather than single substitutions: An episode applies point mutations up to a fixed budget — 13 for the proline-free helix-in-protein agent — so the policy can learn combinations whose effect emerges only jointly. Reaching a 70% loss of helical content scores +10, each non-disruptive step −0.01, and exhausting the budget −25.
• Two environments under one policy family: A Helix-only agent works on 30-residue helices with no protein context; a Helix-in-protein agent sees a 15-residue target helix and the full sequence surrounding it, concatenating ESM-2 embeddings of each.
• A structure predictor as the reward function: Helical content is recomputed after every mutation, so the agent optimizes against a structural readout rather than a sequence heuristic.
• A proline-free action space: Proline is so obvious a helix breaker that leaving it available collapses the policy onto it — chosen 53% of the time in the Helix-only setting and over 95% in Helix-in-protein. Removing it forces richer solutions, dominated by aspartate at 47% and by glycine.
• Free-energy validation: Predicted mutants are re-examined with well-tempered metadynamics in GROMACS patched with PLUMED rather than accepted on the folding model's word.

#Technical Details

Helices come from a therapeutic peptides database of PDB helical motifs. The Helix-only corpus is 2,066 non-redundant 30-residue helices split 90:10 into 1,874 training and 192 validation examples; the Helix-in-protein corpus pairs 15-residue helices with their parent sequences for 32,775 training and 3,643 validation entries, and all reported results use the subset of proteins of 250 residues or fewer (8,794 training, 1,009 validation). PPO was selected over A2C and DQN, which failed to converge; the Helix-only model trained for 168,000 steps and the Helix-in-protein model for 190,000. The proline-free Helix-only agent disrupts its target roughly 80% of the time against over 95% for the proline-enabled version, while the proline-free Helix-in-protein agent reports a learning efficiency of 0.54 — a disrupting set found in about one episode in two — with a mutation entropy of 1.57. Metadynamics checks the proposals: the frataxin tetramutant L186D/A188G/L190G/K192D shifts the most stable state of the 182–196 helix from a backbone psi angle near −40° to roughly +47°, a random coil, and the protein-L triple mutant E41F/S45G/A51D produces a comparable helix-to-coil shift.

#Applications

The intended user is a researcher asking whether a particular helix is load-bearing for a protein's structure or function. Given a sequence and a chosen helical segment, the trained Helix-in-protein agent returns candidate mutant sets to prioritize for expression and biophysical characterization, in place of intuition or exhaustive alanine scanning. The authors demonstrate it on frataxin, implicated in Friedreich's ataxia, and on protein-L, a folding model system, and report predicted mutants for insulin, hemoglobin, myoglobin, a zinc-finger protein and HSP-90 — all from one trained checkpoint, with no per-target retraining.

#Impact

AlphaMut demonstrates a reusable pattern: wrap frozen pretrained models — a protein language model for representation, a structure predictor for evaluation — in a reinforcement learning loop whose only trainable component is the policy network, keeping it cheap to reproduce and easy to retarget at other structural motifs. Its concrete contribution, published in the Journal of Chemical Theory and Computation, is a helix propensity scale computed from disruptive mutations rather than from residue frequencies in existing helices. The limits are equally concrete: helix lengths are fixed at 30 and 15 residues, in-protein results cover only chains of 250 residues or fewer, and every validation is computational, with no experimental characterization of any proposed mutant.