Plants store enormous amounts of carbon in cell walls built like reinforced concrete: cellulose fibers as the rebar, hemicellulose as wrapping, and lignin — a rigid, tangled aromatic polymer — as the concrete encasing it all. This architecture, called biomass recalcitrance, is the main economic barrier to turning crop residues and poplar into fuel, because the enzymes that digest cellulose are physically blocked by lignin and, worse, get stuck to it: residual lignin droplets left after pretreatment adsorb cellulases non-productively so the enzyme sits on the surface doing nothing (Effect of lignin content on changes occurring in poplar cellulose ultrastructure during dilute acid pretreatment). At the same time, a separate research community has learned to open up protein language models — transformers trained on millions of amino-acid sequences, in the same way text models are trained on text — and find individual, human-readable "concepts" hidden in their activations using sparse autoencoders, which reconstruct dense internal vectors through a wide but mostly-off bottleneck so each latent dimension tends to mean one thing.
This proposal claims the two problems need each other. The physical phenotypes that matter for recalcitrance — hydrophobic surface patches that grab lignin, and the flexibility of the loops that grip cellulose — have no curated database labels, which is exactly the regime where existing interpretability tools are weakest. Conversely, recalcitrance research is stuck doing one residue at a time with expensive simulations. We will use sparse-autoencoder interpretability to discover, causally verify, and then steer away the internal features responsible for lignin adsorption in fungal cellulases, producing engineered variants whose behavior is checked against the small number of physics-grade kinetic measurements that exist — and whose design is conditioned on the measured physical state of the substrate, something no protein language model currently sees.
Core Claim
The central hypothesis is that a protein language model (ESM-2) already internally separates two coupled properties of cellobiohydrolases — the processive enzymes that slide along cellulose chains — namely the surface hydrophobicity that drives non-productive lignin binding and the loop flexibility that governs the tradeoff between catalytic speed and getting stuck on the crystal. These features are invisible to ontology-guided methods because Swiss-Prot has no "lignin-binding patch" tag. We propose to recover them by training sparse autoencoders on a corpus enriched for carbohydrate-active enzymes and validating the discovered features not against annotations but against experimental kinetics: the D262V mutation in Trichoderma reesei CBHI, which lowered both the catalytic and the dissociation free-energy barriers (by 2.2 and 2.4 kcal/mol) and raised $k_{cat}$ by 53.1% and cellobiose yield on microcrystalline cellulose by 49.8% (Boosting Enzyme Activity in Biomass Conversion by Modulating the Hydrolysis Process of Cellobiohydrolases), and the two Phanerochaete chrysosporium AA9 oxidative enzymes that boosted sugar release from corn straw by 1.34- and 1.33-fold (Identification and characterization of two AA9 lytic polysaccharide monooxygenases isolated from the enzymatic arsenal in Phanerochaete chrysosporium). The distinctive ingredient is substrate-gated steering: the strength with which we suppress the lignin-binding feature during variant design is modulated by measured substrate descriptors, calibrated on the poplar pretreatment dataset, rather than being a fixed knob.
Research Program
The program has three stages. First, discovery: we curate all GH7 cellobiohydrolases and AA9 lytic polysaccharide monooxygenases from UniRef50 and UniProtKB, train sparse autoencoders (TopK and gated variants) on ESM-2 650M residue embeddings of this corpus, and ask which latent dimensions activate on known hydrophobic surface regions and on the substrate-enclosing loop B3 that Liu et al. identified as the bottleneck. Validation labels come from a new ~200-mutant library we will express in Pichia pastoris (the same host Liu et al. used), scanning loop B3 and surface-exposed aromatic residues, measuring $k_{cat}$, $K_m$, Avicel hydrolysis, and lignin-adsorption isotherms. Second, causal verification: because correlational probing is the known weak point of this literature — the toxin-detection work itself flags that its layer attributions were correlational rather than causally ablated (Toxin Feature Hierarchy in ESM-2: Mechanistic Interpretability reveals Why Frozen Probes Resist ProteinMPNN Redesign) — we will patch and ablate candidate features and require that the model's predicted phenotype moves in the measured direction. A secondary analysis asks where these features live: the "hunchback" intrinsic-dimension trajectory (Geometry of Decision Making in Language Models) and the early "cornerstone" layers found by Shapley attribution (Investigating Layer Importance in Large Language Models) predict specific layerwise signatures that we will test for CAZymes, following the toxin paper's finding that function concentrates late (Layer 32 of 33). Third, steering: following ProtSAE's demonstration that intervening on a concept neuron generates sequences with a TM-score of 0.829 to natural proteins (ProtSAE: Disentangling and Interpreting Protein Language Models via Semantically-Guided Sparse Autoencoders), we steer a generative model with the lignin-binding feature suppressed, at a strength set by the substrate context.
The substrate gating is what makes the biomass direction a first-class ingredient rather than an application afterthought. Sun et al. showed a genuine tension: removing lignin before dilute acid pretreatment makes cellulose more crystalline (crystallite size growing from 2.9 to 4.2 nm) yet dramatically more digestible (glucose yield jumping from 10% to 80%, Simons'-stain accessibility from 21.8 to 72.5 mg/g) (Effect of lignin content on changes occurring in poplar cellulose ultrastructure during dilute acid pretreatment). A protein model that scores enzymes in isolation is blind to this: the same variant should be engineered differently for high-lignin residue (suppress lignin binding hard) versus delignified, coalesced-crystallite substrate (favor processivity and oxidative assistance instead). We calibrate a simple mapping from measured substrate descriptors — lignin content across 23.2%, 19.2%, 14.3%, crystallite size, accessibility — to the steering strength, and evaluate every design under both substrate regimes.
Proposed Experiments
The prediction experiment evaluates whether sparse features predict mutational effects on the held-out mutant library, measured by Spearman correlation between feature-silencing scores and measured activity changes, mirroring how Kim and Ha's concept-silencing metric tracked deep mutational scanning fitness (Decoding Loss-of-Function Variants with Sparse Concept Features of ESM-2). The design experiment expresses the top steered variants and tests them on Avicel, on lignin-rich pretreated poplar, and in combination with the AA9 enzymes under their optimal conditions (60°C, pH 6.0, ascorbic acid enhancement), with the 1.33–1.34-fold sugar-release bar from Xu et al. as the synergy reference. Decisive causal checks are counterfactual feature ablations: suppressing the lignin-binding feature must reduce predicted lignin adsorption without collapsing predicted cellulose affinity, and the ablation must transfer across enzyme families (GH7 hydrolases and AA9 oxidases).
Baselines
The do-nothing reference point is wild-type CBHI in an industrial cocktail on pretreated poplar, where the published anchor is 10% glucose yield for native-lignin (23.2%) biomass after 15 minutes of dilute acid pretreatment, rising to 80% for the delignified 14.3% variant (Effect of lignin content on changes occurring in poplar cellulose ultrastructure during dilute acid pretreatment); no reported result found on Lacuna for any engineered cellulase on lignin-containing pretreated poplar, which is precisely the gap we target. The strongest existing engineered method is the D262V mutant of Liu et al., which raised cellobiose yield on Avicel by 49.8% — but it was designed by molecular dynamics and QM/MM simulation and never tested against lignin; we will run D262V ourselves on our substrate panel under identical assay conditions, and beating it means exceeding +49.8% cellobiose yield on Avicel while outperforming both it and wild type on lignin-rich poplar. The strongest interpretability competitor is the concept-silencing pipeline of Kim and Ha, whose sparse concept features reached a median Spearman correlation of 0.719 across 201 ProteinGym assays, matching the 0.715 of the dense ESM-2 backbone (Decoding Loss-of-Function Variants with Sparse Concept Features of ESM-2); we will run it unchanged on our corpus, and ProtSAE with Gene Ontology guidance as well, whose defined-concept probing reached an average AUC of 0.797 against 0.565 for an unguided TopK sparse autoencoder (ProtSAE: Disentangling and Interpreting Protein Language Models via Semantically-Guided Sparse Autoencoders). Fairness holds by construction: every method scores the identical mutant library, reads the identical ESM-2 650M embeddings, and receives the same SAE training compute and the same random-search tuning budget over width and sparsity. Beating the interpretability baselines means matching or exceeding their Spearman on held-out mutants while producing features that survive causal ablation — which they cannot, since their concept alignment depends entirely on Swiss-Prot tags that do not exist for lignin adsorption or loop dynamics. The key ablation is our own pipeline stripped of its two new ingredients: an unguided TopK sparse autoencoder on the same corpus with no kinetic validation and fixed (non-substrate-gated) steering. If that stripped version matches the full pipeline, the experimental labels and substrate gating are doing no work and the claim collapses to "sparse autoencoders find hydrophobicity," which the toxin paper already established in another domain (Toxin Feature Hierarchy in ESM-2: Mechanistic Interpretability reveals Why Frozen Probes Resist ProteinMPNN Redesign).
Why not just…?
Why not just run molecular dynamics and QM/MM simulations? This is the method that actually produced the best existing variant, with quantified barrier reductions of 2.2 and 2.4 kcal/mol for D262V (Boosting Enzyme Activity in Biomass Conversion by Modulating the Hydrolysis Process of Cellobiohydrolases), and we do not dismiss it — we use it as the final mechanistic check on our top designs. Its limit is throughput: one study, one residue, and the authors themselves left lignin untested because the simulation burden was already enormous. Scanning hundreds of surface positions for lignin-adsorption effects across two enzyme families, then re-designing per substrate state, is out of reach for trajectory-based simulation; our pipeline uses the cheap model to propose and the expensive simulation to confirm.
Why not just use epistatic sequence models? Unsupervised inference of protein fitness landscapes from deep mutational scanning achieved Pearson correlations above 0.8 with experimental binding energies across five assays (Unsupervised inference of protein fitness landscape from deep mutational scan), and alignment-based methods like GEMME remain competitive on ProteinGym, though no reported result found on Lacuna for GEMME's zero-shot substitution scores. Two things rule it out as the whole answer here: accuracy degrades sharply for mutants several substitutions away from the wild type, exactly where design lives, and these models are scorers with no internal structure to steer — they cannot tell you which property of the enzyme to change, let alone separate lignin binding from catalysis. They are a baseline we will run, not the method.
Why not just fine-tune ESM-2 on the assay data? If a few hundred labeled mutants exist, a supervised head plausibly wins on raw prediction: supervised ProteinNPT led ProteinGym with a Spearman correlation of 0.613 (ProteinGym (NeurIPS 2023)), and Kim and Ha showed supervised-level accuracy is achievable at all. This is a real threat, and we will settle it with a label-efficiency curve: performance versus number of labeled mutants, from 10 to the full library. Our claim survives if our features match a supervised head trained on few labels (where the sparse structure supplies the inductive bias the labels cannot) and if the supervised head, interrogated afterward, turns out to rely on the features we found — in which case interpretability is what makes the predictor transferable to lignin-rich substrates the labels never covered. If a fully supervised black box dominates at every label budget and transfers across substrates without our features, we will report that our contribution is the mechanism and the steering, not the prediction.
Decisive Checks
The proposal succeeds if three things hold: sparse features validated only against ~200 kinetic measurements predict held-out mutational effects at backbone-level accuracy; counterfactual ablation of the lignin-binding feature causally changes predicted adsorption without destroying predicted hydrolysis; and a substrate-gated steered variant beats D262V's +49.8% Avicel cellobiose yield while outperforming every baseline on lignin-rich pretreated poplar, where no engineered cellulase has a published number. It fails cleanly — and informatively — if the features are correlational only, or if the unguided ablation matches the full pipeline, in which case the biomass domain gains a negative result about what protein language models do and do not encode about recalcitrance.
References
- Effect of lignin content on changes occurring in poplar cellulose ultrastructure during dilute acid pretreatment
- Boosting Enzyme Activity in Biomass Conversion by Modulating the Hydrolysis Process of Cellobiohydrolases
- Identification and characterization of two AA9 lytic polysaccharide monooxygenases isolated from the enzymatic arsenal in Phanerochaete chrysosporium
- Toxin Feature Hierarchy in ESM-2: Mechanistic Interpretability reveals Why Frozen Probes Resist ProteinMPNN Redesign
- Geometry of Decision Making in Language Models
- Investigating Layer Importance in Large Language Models
- ProtSAE: Disentangling and Interpreting Protein Language Models via Semantically-Guided Sparse Autoencoders
- Decoding Loss-of-Function Variants with Sparse Concept Features of ESM-2
- Unsupervised inference of protein fitness landscape from deep mutational scan
- ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design