ProGenMech: Interpreting Protein Generation

• ProGenMech is a mechanistic interpretability framework that exposes internal computations in generative protein language models during amino-acid generation.
• It employs a replacement model strategy with sparse cross-layer transcoders that reconstruct each layer from preceding latent variables to trace functional circuits.
• Evaluation on generative and zero-shot fitness tasks shows competitive performance while revealing biologically meaningful circuits underlying protein sequence generation.

ProGenMech is a mechanistic interpretability framework for generative protein LLMs, introduced to expose the internal computations of ProGen3 during amino-acid generation, span infilling, and zero-shot fitness estimation. Its core method is a sparse replacement model based on cross-layer transcoders (CLTs), which reconstruct each layer from sparse latent variables drawn from all preceding layers rather than from a single local layer. In the formulation reported for ProGen3-112M, ProGenMech targets the model’s Mixture-of-Experts (MoE) feedforward pathway, then uses the resulting sparse latent space to identify circuits responsible for sequence generation and functional scoring behavior (Tsui et al., 14 Jun 2026).

1. Historical position within the ProGen and protein-interpretability literature

ProGenMech belongs to the broader ProGen lineage but addresses a distinct problem. The original ProGen formulated protein engineering as conditional autoregressive sequence modeling over a unified stream of prepended control tags and amino-acid tokens, with controllability arising from ordinary next-token conditioning rather than side-channel classifiers (Madani et al., 2020). ProGen2 then scaled decoder-only protein LLMs to billions of parameters and argued that data distribution matters as much as raw scale, particularly because improved perplexity on the observed sequence distribution does not necessarily imply better zero-shot biological fitness prediction (Nijkamp et al., 2022).

The immediate methodological precursor is ProtoMech, which introduced cross-layer transcoders for ESM2 and showed that layerwise sparse decomposition is insufficient for recovering cross-layer computation in protein representation models (Tsui et al., 12 Feb 2026). ProGenMech extends that CLT idea from masked or bidirectional representation learning to a generative protein model, and the extension is nontrivial because autoregressive generation introduces sequential dependence, error accumulation, and a distributional evaluation target rather than a purely token-reconstruction target (Tsui et al., 14 Jun 2026).

This positioning implies a useful distinction. ProGen and ProGen2 are primarily generative sequence models; ProtoMech is primarily a circuit-tracing framework for a representation model; ProGenMech is a circuit-tracing framework specialized to a generative protein LLM. A plausible implication is that it addresses the interpretability gap that remained after large autoregressive protein generators became useful for design but still opaque in their internal computation.

2. Base model, data regime, and task formulation

The framework is developed on ProGen3-112M, described as a generative protein LLM that supports both causal language modeling (CLM) and generalized language modeling (GLM) for span infilling, with feedforward blocks implemented as a sparse MoE architecture. The experiments use the 112M-parameter version with $L = 10$ layers and model dimension $d_{\text{model}} = 384$ (Tsui et al., 14 Jun 2026).

For transcoder training, the paper uses 5 million protein sequences randomly sampled from UniRef50, with sequence length up to 1022 amino acids. Training examples mirror ProGen3’s dual-objective regime through a 2:1 split between CLM and GLM sequences. In the GLM branch, span lengths are sampled from a mixture of five Gaussians,

$$\mathcal{N}(10,5),\quad \mathcal{N}(30,10),\quad \mathcal{N}(70,20),\quad \mathcal{N}(200,50),\quad \mathcal{N}(400,100),$$

and the maximum masked fraction is sampled from

$\{0.15,0.25,0.5,0.8\}$

with probabilities

$\{0.28,0.30,0.28,0.14\}.$

The evaluation protocol covers three tasks. In CLM generation, the model receives the first 80% of a real protein sequence and must generate the remaining 20% autoregressively. In GLM span infilling, a span within a real sequence is masked and then infilled; for evaluation, masked span lengths are sampled from the two smallest Gaussian components, $\mathcal{N}(10,5)$ and $\mathcal{N}(30,10)$, with equal probability. In zero-shot fitness estimation, mutants are scored by sequence likelihood, with ProGen3’s score defined as the average log-likelihood between forward and reverse scoring directions (Tsui et al., 14 Jun 2026).

For the replacement-model notation, the residual stream at the start of layer $\ell$ is $\mathbf{x}_{\mathrm{pre}}^{\ell}$. After attention,

$$\mathbf{x}^{\ell} = \mathbf{x}_{\mathrm{pre}}^{\ell} + \sum_{h \in H_{\ell}} h(\mathbf{x}_{\mathrm{pre}}^{\ell}),$$

and the MoE sublayer produces

$d_{\text{model}} = 384$0

so the next residual stream is

$d_{\text{model}} = 384$1

ProGenMech reconstructs $d_{\text{model}} = 384$2, not the individual experts or routing decisions.

3. Cross-layer transcoder architecture

The central architectural claim is that a local, per-layer sparse bottleneck cannot faithfully recover inter-layer generative computation. ProGenMech therefore uses CLTs, in which each layer is reconstructed from latent variables drawn from all previous layers. At layer $d_{\text{model}} = 384$3, the residual stream activation $d_{\text{model}} = 384$4 is encoded into a sparse latent vector $d_{\text{model}} = 384$5, with

$d_{\text{model}} = 384$6

a $d_{\text{model}} = 384$7 expansion over $d_{\text{model}} = 384$8. Sparsity is imposed by TopK with

$d_{\text{model}} = 384$9

The encoder is

$$\mathcal{N}(10,5),\quad \mathcal{N}(30,10),\quad \mathcal{N}(70,20),\quad \mathcal{N}(200,50),\quad \mathcal{N}(400,100),$$0

The defining CLT decoder equation is

$$\mathcal{N}(10,5),\quad \mathcal{N}(30,10),\quad \mathcal{N}(70,20),\quad \mathcal{N}(200,50),\quad \mathcal{N}(400,100),$$1

This means that the reconstructed MoE output at layer $$\mathcal{N}(10,5),\quad \mathcal{N}(30,10),\quad \mathcal{N}(70,20),\quad \mathcal{N}(200,50),\quad \mathcal{N}(400,100),$$2 is allowed to depend explicitly on sparse features discovered anywhere from layer 1 through layer $$\mathcal{N}(10,5),\quad \mathcal{N}(30,10),\quad \mathcal{N}(70,20),\quad \mathcal{N}(200,50),\quad \mathcal{N}(400,100),$$3. By contrast, the per-layer transcoder baseline reconstructs only locally,

$$\mathcal{N}(10,5),\quad \mathcal{N}(30,10),\quad \mathcal{N}(70,20),\quad \mathcal{N}(200,50),\quad \mathcal{N}(400,100),$$4

The main reconstruction loss is

$$\mathcal{N}(10,5),\quad \mathcal{N}(30,10),\quad \mathcal{N}(70,20),\quad \mathcal{N}(200,50),\quad \mathcal{N}(400,100),$$5

With residual

$$\mathcal{N}(10,5),\quad \mathcal{N}(30,10),\quad \mathcal{N}(70,20),\quad \mathcal{N}(200,50),\quad \mathcal{N}(400,100),$$6

the auxiliary loss is

$$\mathcal{N}(10,5),\quad \mathcal{N}(30,10),\quad \mathcal{N}(70,20),\quad \mathcal{N}(200,50),\quad \mathcal{N}(400,100),$$7

with

$$\mathcal{N}(10,5),\quad \mathcal{N}(30,10),\quad \mathcal{N}(70,20),\quad \mathcal{N}(200,50),\quad \mathcal{N}(400,100),$$8

and total objective

$$\mathcal{N}(10,5),\quad \mathcal{N}(30,10),\quad \mathcal{N}(70,20),\quad \mathcal{N}(200,50),\quad \mathcal{N}(400,100),$$9

The implementation details reported for ProGenMech are batch size 16, AdamW, learning rate $\{0.15,0.25,0.5,0.8\}$0, gradient clipping 1, and weight decay $\{0.15,0.25,0.5,0.8\}$1. Held-out validation normalized MSE after training is reported as $\{0.15,0.25,0.5,0.8\}$2 on average, and the CLT has about 115M parameters, reflecting the $\{0.15,0.25,0.5,0.8\}$3 growth in cross-layer decoder connections (Tsui et al., 14 Jun 2026).

4. Replacement models and circuit-discovery procedure

ProGenMech is used as a replacement model for ProGen3’s MoE pathway. The paper studies three replacement regimes. Direct replacement uses ground-truth residual-stream activations $\{0.15,0.25,0.5,0.8\}$4 at every layer and reconstructs only the final layer; it serves as an upper bound on representational fidelity. Sequential replacement replaces the true MoE output at each layer with its CLT reconstruction,

$\{0.15,0.25,0.5,0.8\}$5

while holding attention outputs fixed to ProGen3’s true attention outputs. Full replacement recursively replaces both MoE and attention from reconstructed internal states, and is reported to degrade substantially because of compounded errors (Tsui et al., 14 Jun 2026).

Circuit discovery is zero-shot and task-specific. A circuit is defined as a set of latent variables whose activity is responsible for a particular model behavior. For attribution, the latent importance score is

$\{0.15,0.25,0.5,0.8\}$6

For generation tasks this is instantiated as

$\{0.15,0.25,0.5,0.8\}$7

where $\{0.15,0.25,0.5,0.8\}$8 is computed only over the logits of generated tokens. For zero-shot fitness it becomes

$\{0.15,0.25,0.5,0.8\}$9

with $\{0.28,0.30,0.28,0.14\}.$0 defined over the entire final logit matrix.

The search is greedy. For generation circuits, the full-latent replacement model yields $\{0.28,0.30,0.28,0.14\}.$1, and the stopping threshold is

$\{0.28,0.30,0.28,0.14\}.$2

Latents are ranked by attribution, added in batches of 32, and the process stops once the sparse circuit’s mean KL divergence is at most $\{0.28,0.30,0.28,0.14\}.$3, or when the circuit reaches 1000 latents. For zero-shot fitness circuits, with $\{0.28,0.30,0.28,0.14\}.$4 the original ProGen3 Spearman correlation and $\{0.28,0.30,0.28,0.14\}.$5 the full-latent replacement Spearman, the target is

$\{0.28,0.30,0.28,0.14\}.$6

again with batch size 32 and a 1000-latent ceiling (Tsui et al., 14 Jun 2026).

This procedure makes circuit identity global at the feature level but sequence- and position-specific at runtime through the actual latent activations. A plausible implication is that ProGenMech treats “mechanism” as a sparse cross-layer computation graph rather than as an isolated motif detector or a purely local feature basis.

5. Evaluation setup and quantitative performance

The generation benchmarks use Swiss-Prot sequences clustered at 30% sequence identity, with 1000 randomly sampled sequences for circuit discovery and evaluation. Generated sequences are filtered to remove low-quality repetitive artifacts by requiring less than 25% low-complexity regions, where low-complexity is defined as tandem repeats spanning at least six consecutive residues. Generation uses top-$\{0.28,0.30,0.28,0.14\}.$7 sampling with $\{0.28,0.30,0.28,0.14\}.$8 and temperature $\{0.28,0.30,0.28,0.14\}.$9. To avoid temporal drift, the replacement model receives the ground-truth ProGen3 activations from the preceding step at each generative step (Tsui et al., 14 Jun 2026).

For zero-shot function prediction, the paper uses eight ProteinGym deep mutational scanning assays: A4_HUMAN_Seuma, CAPSD_AAV2S_Sinai, F7YBW8_MESOW_Ding, GFP_AEQVI_Sarkisyan, GRB2_HUMAN_Faure, RASK_HUMAN_Weng_abundance, SPG1_STRSG_Olson, and YAP1_HUMAN_Araya. For each assay, 256 mutants are sampled for circuit construction, balanced between functional and nonfunctional bins, and 1000 random mutants are used as test set across 5 folds. The main metric is Spearman correlation between model log-likelihood scores and experimental fitness (Tsui et al., 14 Jun 2026).

The core quantitative results are as follows.

Task	ProGenMech result	Comparison
CLM generation	Full-latent NLL $\mathcal{N}(10,5)$0; circuit NLL $\mathcal{N}(10,5)$1	Original ProGen3 $\mathcal{N}(10,5)$2; PLT full-latent $\mathcal{N}(10,5)$3; PLT circuit $\mathcal{N}(10,5)$4
GLM infilling	Full-latent NLL $\mathcal{N}(10,5)$5; circuit NLL $\mathcal{N}(10,5)$6	Original ProGen3 $\mathcal{N}(10,5)$7; PLT full-latent $\mathcal{N}(10,5)$8; PLT circuit $\mathcal{N}(10,5)$9
Zero-shot fitness	Full-latent Spearman $\mathcal{N}(30,10)$0; circuit $\mathcal{N}(30,10)$1	Original ProGen3 $\mathcal{N}(30,10)$2; PLT full-latent $\mathcal{N}(30,10)$3; PLT circuit $\mathcal{N}(30,10)$4

In CLM generation, ProGenMech full-latent replacement recovers about 60% of the original model’s likelihood, while the discovered circuits recover about 58%. The average CLM circuit size is $\mathcal{N}(30,10)$5 latents, described as less than 2% of the total latent space. In GLM generation, ProGenMech essentially matches the original model’s NLL and also matches the PLT baseline rather than clearly exceeding it. In zero-shot fitness estimation, full-latent ProGenMech recovers about 95% of ProGen3’s average Spearman correlation, and sparse circuits recover about 80%, with average circuit size $\mathcal{N}(30,10)$6 latents, about 0.6% of the latent space (Tsui et al., 14 Jun 2026).

The appendix further reports that direct replacement has the highest fidelity, that sequential replacement is the main fair comparison setting, and that full replacement is much worse because replacing attention induces strong error accumulation. It also reports a U-shaped layerwise latent density: many latents in early layers, fewer in middle layers, and many again in final layers. The authors interpret this as early biochemical or local-dependency detectors, more compressed middle-layer motif representations, and late task-specific features (Tsui et al., 14 Jun 2026).

6. Biological circuit interpretations

A central claim of ProGenMech is that the recovered circuits correspond to biologically meaningful motifs and functional regions rather than to uninterpretable sparse surrogates. The main examples are a kinase sequence, UniProt ID P83104, and the C-terminal SH3 domain of human GRB2 (Tsui et al., 14 Jun 2026).

For the kinase, the paper analyzes the HRD motif at positions 133–135 and the DFG motif at positions 152–154. In the HRD circuit, L1/3183 and L2/1754 activate on arginine residues, these feed into L5/1090, which identifies the broader catalytic loop, and then into L7/2070, which narrows to the HRD motif within that loop and also highlights interacting parts of the ATP-binding site. L8/897, the latent with the largest attribution in the circuit, is mainly responsible for detecting the aspartic acid in the HRD motif. In the DFG circuit, L1/934 and L3/1068 activate on phenylalanine, these feed into L4/2366, which detects glycine in the DFG motif and has the largest attribution, and then into L7/2070; L8/1710 detects the kinase N-terminal lobe involved in ATP anchoring.

These examples support the paper’s hierarchical interpretation: early layers detect residue identities or simple biochemical patterns, middle layers assemble them into broader motifs or loops, and later layers refine them into motif-specific or function-specific generative decisions. The claim is not that ProGenMech simulates enzymatic mechanism; rather, it traces the internal computation by which the LLM builds amino-acid probability mass around conserved motifs.

The GRB2 case study addresses zero-shot fitness rather than generation. For the wildtype sequence, L1/2693 detects aspartic acid, L1/1522 detects tryptophan, L9/2113 activates on the homodimer interface, L10/3297 activates on GRB2 sites interacting with GAB2, and L10/225 activates on those same binding sites and residues stabilizing the hydrophobic core. For the known high-fitness mutant H26D, L1/2693 activates on the introduced aspartic acid and the attribution score of L10/225 increases by more than 58%. For the known low-fitness mutant Y51D, L1/2693 again activates on the introduced aspartic acid, but the attribution score of L10/3297 decreases by about 57%. Since Y51 is a known GAB2 binding site, the paper interprets this as reduced binding capability, consistent with lower fitness (Tsui et al., 14 Jun 2026).

The biological validation strategy in these examples is qualitative but structured: the authors compare top-activating Swiss-Prot sequences for each latent, align activations with conserved motifs, project activations onto structure, and check consistency with known biochemical annotations and mutational effects.

7. Steering, limitations, and significance

The paper also explores whether fitness-related circuits can be used to steer generation, but the reported results are largely negative at the 112M scale. Interventions on zero-shot fitness circuits do not yield sequences whose scores are meaningfully improved relative to baseline ProGen3 generations. The authors note that both baseline and steered generations often fail the low-complexity filter and conclude that, in the 112M model, there is a mismatch between ProGen3’s scoring mode, where likelihood correlates with fitness, and its generation mode, where the model does not generate sequences within the same desirable fitness distribution (Tsui et al., 14 Jun 2026).

Several limitations are explicit. First, ProGenMech treats each sparse MoE block as a single input–output computation and therefore abstracts away expert routing logic; it does not explain which experts are selected or what each expert specializes in. Second, the CLT decoder count scales as $\mathcal{N}(30,10)$7, and the resulting ProGenMech model has about 115M parameters, creating a nontrivial scaling cost. Third, full replacement remains weak because fidelity depends on frozen true attention outputs. Fourth, the GLM conclusions are constrained by the apparent weakness of the 112M ProGen3 model in infilling. Fifth, biological interpretation still depends on manual examination of latents and annotations. Sixth, steering is not yet effective at this model scale (Tsui et al., 14 Jun 2026).

Despite these constraints, the broader significance claimed for ProGenMech is clear. It provides a path toward interpretable protein generation by tracing which latent features and cross-layer pathways drive amino-acid probabilities and fitness-related likelihoods. It also supports an auditing perspective: one can inspect whether a generative protein LLM relies on known conserved motifs, functional regions, or suspicious artifacts. The paper’s strongest conceptual conclusion is that generative circuits are inherently cross-layer objects. If amino-acid predictions are built by progressively composing early biochemical cues into later motif-level decisions, then layer-local interpretability methods will systematically miss part of the mechanism (Tsui et al., 14 Jun 2026).

In that sense, ProGenMech marks a shift in the ProGen literature from controllable generation and scale toward internal explanatory structure. ProGen framed control through prepended metadata (Madani et al., 2020); ProGen2 emphasized scale and data distribution (Nijkamp et al., 2022); ProGenMech turns to the question of how a generative protein LLM actually computes its outputs across depth. This suggests a broader research program in which protein design systems are not only powerful and conditional, but also circuit-level interpretable.