The thermostability-activity trade-off is one of the oldest observations in enzyme engineering. Increase the thermal stability of most enzymes and you reduce their specific activity at physiological temperature. The two properties pull in opposite directions with surprising consistency across protein families. Understanding why this happens at the structural level, and knowing the conditions under which it can be bypassed, is prerequisite knowledge for any computationally guided enzyme optimization campaign.
The Structural Basis of the Trade-Off
Catalysis requires conformational flexibility. Enzymes that function near physiological temperature have evolved active sites with a specific degree of conformational sampling: the substrate must be able to enter and exit, catalytic residues must adopt the correct geometry for chemistry, and the product must be able to leave. This requires that the active site region and the loops and helices that govern access to it be flexible enough to breathe on timescales relevant to the catalytic cycle, typically microseconds to milliseconds.
Thermostabilizing substitutions work by restricting this flexibility. They add buried hydrogen bonds, remove buried waters, introduce disulfide bridges in hinge regions, pack the hydrophobic core more tightly, or rigidify surface loops that were previously dynamic. All of these strategies are effective at raising Tm. And most of them, if applied near or at the active site, also reduce the conformational sampling that activity depends on.
The exceptions, the cases where you actually get both, tend to fall into a specific pattern: the stabilizing substitutions are distant from the active site, in regions where rigidification does not propagate conformationally to the catalytic machinery. This is sometimes called remote stabilization, and it is the category of substitution where engineering is most reliably possible without activity loss.
The challenge is identifying which positions are genuinely remote in this sense. A position can be 20 angstroms from the catalytic residues and still be conformationally coupled to them through a network of correlated motions. Crystal structures show you static averages, not dynamic coupling networks. This is one reason that purely structure-based stability engineering has a poor track record when the goal is stability with preserved activity: the structure does not tell you which positions are dynamically coupled to catalysis.
When Thermophile Homologs Do and Do Not Help
A common approach to rapid thermostabilization is to identify a thermophilic homolog of your mesophilic target and graft stabilizing features from it. This works reasonably well for raising Tm, but the activity transfer is often incomplete. Thermophilic enzymes have evolved to function at 60 to 90 degrees Celsius. Their activity at 37 degrees or 25 degrees is substantially lower than that of their mesophilic counterparts, precisely because their active sites are too rigid at lower temperatures to support efficient catalysis.
A consensus-sequence approach (averaging across thermophilic homologs to identify the modal residue at each position) has similar properties: it raises stability reliably but activity at the operating temperature typically drops. The consensus sequence reflects what works on average across the thermophilic clade, not what optimizes the specific enzyme's catalytic mechanism.
We are not saying thermophile grafting is wrong. If your process temperature is 60 degrees or above, thermophile-guided engineering is appropriate and often efficient. But if your operating temperature is ambient or physiological, thermophile-guided stabilization will trade activity away to get the stability you need. The degree of that trade-off depends on the protein family and the specific positions you are targeting.
What Generative Models Can and Cannot Do Here
A generative sequence model trained on evolutionary data has implicitly learned the correlation structure between stability-related positions and activity-related positions, to the extent that this correlation is captured in the training distribution. This gives it an advantage over purely rational design: when you ask it to propose sequences optimized jointly for high stability and high activity, it can draw on co-evolutionary patterns that suggest which combinations of substitutions have appeared together in natural variants.
But this advantage has a hard limit. If the evolutionary record contains very few examples of mesophilic-activity thermostable variants (because evolution did not encounter this selection pressure in any lineage), the model will have weak signal on this joint objective. The model knows what mutations tend to co-occur; it does not know about physical mechanisms that might allow exceptions to the usual trade-off.
In practice, we see generative models perform best on the thermostability-activity problem when there is a sufficiently diverse MSA covering organisms with different native operating temperatures, and when the model is conditioned on a fitness objective that explicitly penalizes activity loss alongside thermostability gain. Running the model with thermostability as the sole objective will produce sequences that look like natural thermophiles. That may not be what you want.
A Scenario Where Threading the Needle Worked
Consider a case we worked through with a cellulase optimization campaign. The starting enzyme was a mesophilic GH12 family endoglucanase with good activity at 50 degrees Celsius and poor stability above 55 degrees (Tm approximately 57 degrees). The process target was 65 degrees operating temperature with acceptable activity loss, targeting no more than 30 percent reduction in kcat/Km relative to wild-type measured at the optimal temperature.
We ran an MSA of 340 sequences spanning from psychrophilic marine bacteria to moderate thermophiles. The sequence landscape showed clear co-variation between two positions in a beta-loop region flanking the substrate cleft and four positions in a peripheral helix on the opposite face of the protein. The co-variation was specific: certain combinations at the loop positions strongly co-occurred with certain combinations at the peripheral helix, across organisms with different temperature optima.
The generative proposals targeting Tm improvement while holding MSA-based fitness scores high concentrated on combinations of the peripheral helix positions. The active site loop positions were left near wild-type because the model's fitness score dropped sharply when those were altered. We proposed 48 variants in the first round. 11 showed Tm increases above 5 degrees. Of those, 9 retained more than 75 percent of wild-type activity at 50 degrees. The best variant showed a 9-degree Tm increase with approximately 88 percent retained activity.
The key was that the stabilizing substitutions were at positions structurally remote from the catalytic center and not co-varying with active site residues in the MSA. The generative model found this region efficiently. A rational approach focused only on the structural proximity might have missed it entirely, because the peripheral helix looks disconnected from function in the crystal structure.
Practical Guidance for Campaign Setup
If you are setting up an enzyme optimization campaign with joint stability-activity objectives, the following framing helps.
First, separate the positions into three classes: active site adjacent (within 8 to 10 angstroms of catalytic residues in the static structure), dynamic coupling candidates (identified by MD simulation or by co-variation patterns in the MSA linking them to active site positions), and remote candidates (no structural proximity, no MSA coupling to catalytic residues). Restrict thermostabilizing mutations to the remote class in your first round. This is a conservative prior that often gives you 5 to 8 degrees of Tm increase without touching activity.
Second, use a two-objective fitness criterion when generating candidates, not a single stability score. Score candidates simultaneously on stability indicators (positions contributing to known stabilization modes: core packing, buried polar contacts, surface charge distribution) and on activity proxies (conservation scores at active site positions, compatibility with substrate-binding residues). Candidates that score well on both deserve assay priority over candidates that excel on one.
Third, accept that some campaigns will hit a ceiling. If your protein's active site is tightly coupled to the flexibility that confers thermostability, you will reach a point where no further Tm improvement is available without sacrificing activity. That ceiling is a scientific reality, not an engineering failure. At that point, the question becomes whether the achieved Tm is sufficient for your process or whether a scaffold switch is necessary.
The trade-off is real and structural. But it is not universal. Remote stabilization and co-evolution-guided design give you real room to maneuver, if you set up the problem correctly from the start.