Technical Summary: LineageFlow

Problem Statement

Protein engineering requires generating sequences that are biophysically plausible and remain recognizable members of a target family or domain while exploring within-family diversity. Current discrete generative models, including diffusion and flow-based approaches, typically initialize generation from uniform simplex noise or masked-token corruption. The authors argue that these generic priors are "family-agnostic" and discard strong position-specific constraints induced by evolution. Consequently, the model is forced to reconstruct nearly every residue, including highly conserved ones, from heavily corrupted states ("from-scratch synthesis"). This burden leads to weak family control, low plausibility, and difficulty in maintaining recognizable family-domain structures.

Methodology: LineageFlow

The authors propose LineageFlow, a Dirichlet flow-matching model that reframes protein sequence generation as structured mutation from an evolved scaffold rather than synthesis from generic noise.

1. Lineage-Prior Initialization

Instead of a uniform prior, LineageFlow initializes the generative process using ancestral priors derived from Ancestral Sequence Reconstruction (ASR).

• Construction: For a given protein family, the authors construct a Multiple Sequence Alignment (MSA), infer a maximum-likelihood phylogenetic tree, and perform marginal ASR at the root to obtain a site-wise posterior distribution over residues.
• Representation: This root posterior is encoded as Dirichlet concentration parameters ($\alpha^{(h,l)}$), defining a family-specific ancestral prior $q_0^{(h)}(X) = \prod_l \text{Dir}(x^{(l)}; \alpha^{(h,l)})$.
• Mechanism: This approach preserves conserved scaffolds (where the family is invariant) while retaining uncertainty only at genuinely variable sites. The generation process is thus modeled as a continuous transport on the probability simplex from this ancestral distribution toward extant sequences.

2. Flow Matching on the Simplex

LineageFlow employs Dirichlet Flow Matching (DFM).

• Trajectory: A conditional probability path $p_t(x | x_1)$ is defined where the Dirichlet mass concentrates around the target vertex $x_1$ over time $t \in [0, 1]$.
• Vector Field: The model learns a neural time-dependent vector field $\hat{v}(X, t; \theta)$ that transports the ancestral prior toward the data distribution. The drift is reconstructed via a classifier parameterization $\hat{p}_\theta(x_1 | x_t, t)$, minimizing the sequence-averaged cross-entropy loss.
• Theoretical Guarantee: The authors prove that the lineage-specific analytic field satisfies the continuity equation on the simplex, ensuring probability mass conservation and keeping trajectories within the valid state space.

3. Rerouting for Objective-Guided Sampling

To incorporate external fitness objectives without per-step predictor guidance, the authors introduce Rerouting.

• Process: At an intermediate time $t_{int}$, the generation trajectory is paused. A "mutate–select–amplify" intervention is applied, inspired by directed evolution.
  • Mutate: A proposal kernel injects diversity into the population of particles.
  • Select: Particles are reweighted based on a fitness score $J(X)$ (e.g., plausibility or specific property predictors).
  • Amplify: Resampling is performed according to the weights.
• Theoretical Basis: The authors formalize rerouting as a KL-regularized selection problem, showing that the reweighted distribution maximizes expected fitness while minimizing KL divergence from the proposal law.
• Refinement: The selected particle is then integrated through the base flow from $t_{int}$ to $t=1$ to produce the final sample.

Key Contributions

1. Ancestral Priors vs. Generic Noise: The paper demonstrates that initializing from ASR-derived lineage priors significantly outperforms uniform or mask-initialized baselines. It transforms generation from "from-scratch" denoising to structured mutation, preserving conserved scaffolds.
2. Manifold-Preserving Guidance: Rerouting provides a single-step, intermediate-time intervention that steers samples toward objectives without requiring gradients at every Euler step, empirically preserving family validity.
3. High-Fidelity Generation: The model achieves family validity close to held-out natural sequences and improves predicted structural confidence (plausibility) while maintaining substantial novelty and diversity.

Experimental Results

The model was evaluated on 8,886 Pfam-A families (RP35 dataset) with 1024 generated sequences per method.

• Family Validity: LineageFlow achieved a top-1 family accuracy (Accfam) of 95.3%, compared to 0% for uniform-prior baselines (DFM, EvoDiff) and 92.8% for an ASR-only baseline (ASR-PSSM i.i.d.). This indicates that generic noise priors fail to produce recognizable family members even when conditioned on family labels.
• Plausibility (Foldability): LineageFlow achieved a mean pLDDT of 76.6, outperforming the ASR-only baseline (70.8) and significantly surpassing DFM (46.2) and EvoDiff (45.4). It also outperformed the pretrained ProtBFN (71.9) despite ProtBFN being trained on a corpus ~8x larger.
• Novelty and Diversity: Among foldable samples (pLDDT $\ge$ 70), 86.2% satisfied a nearest-neighbor identity (NNId) $< 0.8$, and 48.9% satisfied NNId $< 0.6$, demonstrating deep novelty while retaining family identity.
• Zero-Shot Enzyme Generation: In a zero-shot setting on three held-out enzyme families, LineageFlow with rerouting preserved conserved motifs and improved solubility and thermostability proxy scores without fine-tuning on the target families.
• Theoretical Validation: Analysis of the "hard regime" (early time steps) showed that ASR priors provide a higher recoverable signal for the denoiser compared to uniform priors, leading to higher Bayes-optimal accuracy and lower training loss.

Significance and Claims

The paper claims that LineageFlow establishes a new paradigm for protein sequence generation by leveraging evolutionary history directly in the generative prior.

• Mechanism: By replacing generic noise with phylogeny-informed ancestral scaffolds, the model reduces the ambiguity of the early generation steps, allowing the learned vector field to focus on modeling evolutionary variation rather than reconstructing conserved structure.
• Practicality: The introduction of Rerouting enables objective-guided sampling (e.g., optimizing for stability or solubility) without the computational cost of per-step guidance, making it suitable for directed evolution-style tasks.
• Limitations: The authors note that the method relies on high-quality MSAs and phylogenetic inference. Generation is tied to family-specific alignment coordinates (handling gaps as fixed masks) rather than explicitly modeling indels. Furthermore, evaluations rely on computational proxies (e.g., OmegaFold pLDDT, ESM-IF) and predictor-based scores, meaning real-world functional validation remains to be tested.

The work suggests that incorporating explicit evolutionary context (lineage priors) is critical for achieving high-fidelity, family-aware protein design, offering a robust alternative to standard noise-based generative processes.