Problem
Engineering proteins/enzymes to hit target fitness (activity, stability)
Given a parent protein or enzyme, the task is to propose sequence variants that improve a measured property - catalytic activity, thermostability, expression level, or substrate specificity - without degrading the others. This matters commercially because enzymes are the production route for pharmaceutical intermediates, detergents, food ingredients and plastic recycling, and today each improvement still costs many rounds of directed evolution and wet-lab screening. Two acceptance criteria are in public use. On prediction: average Spearman correlation against held-out deep mutational scanning assays on the ProteinGym DMS-substitution benchmark (217 assays, roughly 3M mutants), where sequence-only zero-shot baselines such as ESM-2 650M sit near 0.52 and the leaderboard is topped by multimodal models combining MSAs and structure. On design: the fraction of computationally designed variants that beat the wild-type parent in an independent wet-lab assay while still expressing and remaining stable. The Align Foundation Protein Engineering Tournament made the second criterion concrete - of roughly 1200 experimentally characterised designed variants in its first edition, only two teams exceeded a 50 percent success rate for meeting both expression and stability criteria. Solved would mean a model that, on a protein family it has never seen assayed, returns a high fraction of improved, expressible, stable variants in a single design round.
Funding
Raised by the 6 companies working on it.
Why it is still open
Wet-lab tournaments show computational design still fails multi-parameter optimisation: in the Protein Engineering Tournament only two participating teams achieved above 50 percent success rates for designed variants meeting both expression and stability criteria, most designed sequences carried 15 or more mutations, and the authors state that multiparameter optimisation was challenging for many teams. On the prediction side ProteinGym zero-shot correlations remain far below assay reproducibility, and a 2025 scaling study of protein language models found fitness performance improves non-monotonically with more pretraining data, with unsupervised models not yet consistently beating supervised baselines.
Source: pmc.ncbi.nlm.nih.gov
Who is working on it
Also active
Funding for these is not counted: the problem is a small part of what they do, or they contribute data and tools.
Prizes
| Prize | Status | Pays for | Amount |
|---|---|---|---|
| 2025 Protein Engineering Tournament (PETase) - Align Foundation | Closed, unawarded | Solving | $30k |
What counts as solved
import Mathlib /-! # Engineering proteins/enzymes to hit target fitness (activity, stability) Two ways a board can pay, both stated here: 1. `SolvedCampaign` — a design campaign: propose at most `maxVariants` sequences of a named parent enzyme; the lab expresses and assays them; success means at least one variant improves catalytic efficiency k_cat/K_M by a fold factor AND does not lose thermostability (ΔT_m ≥ bound), within the screening budget. This is the Align-to-Innovate / Cradle-style "one round, few variants" setting. 2. `SolvedPrediction` — a ProteinGym-style held-out assay: predict the fitness of unseen variants, scored by Spearman rank correlation against the measured DMS values, beating a pinned baseline by δ on every assay. Concrete in Lean: fold-improvement arithmetic, the Spearman definition over a finite held-out variant list, both acceptance predicates, and a proved monotonicity lemma. Opaque oracle: `Assay` (expression, measured k_cat/K_M in M⁻¹s⁻¹, T_m in °C, all under a pinned protocol, buffer and temperature) and `Dms` (the hidden deep-mutational-scanning measurements). Attesters: the contract lab for the campaign; the dataset owner (ProteinGym maintainers or the board's lab) for the prediction track, who must confirm the labels were withheld until the submission hash was committed. -/ namespace OpenSpur.Problems.EnzymeProteinFitness structure Variant where sequence : String /-- Measured properties under the pinned protocol (oracle). -/ structure Assay where expressed : Variant → Prop catEff : Variant → ℝ -- k_cat/K_M, M⁻¹ s⁻¹ tm : Variant → ℝ -- °C parent : Variant structure Thresholds where maxVariants : ℕ -- screening budget, e.g. 20 foldImprove : ℝ -- e.g. 10 maxTmLoss : ℝ -- °C, e.g. 2 /-- A variant hits the engineering target. -/ def Improved (A : Assay) (T : Thresholds) (v : Variant) : Prop := A.expressed v ∧ T.foldImprove * A.catEff A.parent ≤ A.catEff v ∧ A.tm A.parent - T.maxTmLoss ≤ A.tm v def SolvedCampaign (A : Assay) (T : Thresholds) (vs : List Variant) : Prop := vs.length ≤ T.maxVariants ∧ ∃ v ∈ vs, Improved A T v noncomputable section def avgRank {n : ℕ} (x : Fin n → ℝ) (i : Fin n) : ℝ := ((Finset.univ.filter fun j => x j < x i).card : ℝ) + (((Finset.univ.filter fun j => x j = x i).card : ℝ) + 1) / 2 def mean {n : ℕ} (x : Fin n → ℝ) : ℝ := (∑ i, x i) / n def pearson {n : ℕ} (x y : Fin n → ℝ) : ℝ := (∑ i, (x i - mean x) * (y i - mean y)) / Real.sqrt ((∑ i, (x i - mean x) ^ 2) * (∑ i, (y i - mean y) ^ 2)) def spearman {n : ℕ} (x y : Fin n → ℝ) : ℝ := pearson (avgRank x) (avgRank y) /-- One hidden DMS assay (oracle): variants, measured fitness, pinned baseline ρ. -/ structure Dms where n : ℕ variant : Fin n → Variant fitness : Fin n → ℝ baselineRho : ℝ def rho (d : Dms) (f : Variant → ℝ) : ℝ := spearman (fun i => f (d.variant i)) d.fitness def SolvedPrediction {m : ℕ} (assays : Fin m → Dms) (f : Variant → ℝ) (minN : ℕ) (margin : ℝ) : Prop := 0 < m ∧ ∀ a, minN ≤ (assays a).n ∧ (assays a).baselineRho + margin ≤ rho (assays a) f end /-- Fewer folds of improvement is a weaker target, given a nonnegative parent activity. -/ theorem improved_mono (A : Assay) (T : Thresholds) (v : Variant) (c : ℝ) (hp : 0 ≤ A.catEff A.parent) (hc : c ≤ T.foldImprove) (h : Improved A T v) : Improved A { T with foldImprove := c } v := by obtain ⟨h1, h2, h3⟩ := h refine ⟨h1, ?_, h3⟩ exact le_trans (mul_le_mul_of_nonneg_right hc hp) h2 end OpenSpur.Problems.EnzymeProteinFitness
Public resources
- ProteinGym - DMS substitution and indel protein fitness benchmarks with 90+ baselines
- ProteinGym NeurIPS 2023 Datasets and Benchmarks paper
- Align Foundation Protein Engineering Tournament benchmarks (predictive plus generative, wet-lab scored)
- Protein Engineering Tournament results paper (Proteins, 2025)
- Align Data Portal - protein tournament datasets
- TAPE fluorescence and meltome splits (Cradle Bio, Hugging Face)
- Profluent E1 protein representation models (open weights)
- Scaling and Data Saturation in Protein Language Models (2025)