Rat liver microsome (RLM) assays are cheap, fast, and run early enough in lead optimization to actually influence scaffold selection. They are also a surprisingly unreliable predictor of human hepatic clearance for a meaningful fraction of drug-like chemical space. The species translation problem — inferring human PK parameters from rodent metabolic data — has been understood qualitatively for decades. What has changed is that we now have enough human data, and enough structural biology on the relevant CYP isoforms, to model the translation computationally rather than relying on empirical allometric scaling or default correction factors. This post describes where the gap comes from, what drives it at the structural level, and where in silico species translation models are currently useful versus where they still fall short.
The Mechanistic Source of the Discrepancy
The core problem is CYP isoform portfolio divergence. Rats and humans share the major CYP families — CYP3A, CYP2C, CYP2D — but the specific isoforms within those families have meaningfully different substrate preferences. CYP3A1/2 in rats is the rough functional analog of CYP3A4 in humans, but "rough functional analog" hides considerable differences in active site topology that translate directly to different regioselectivity on the same substrate. A molecule that is primarily a CYP3A1/2 substrate in rats may be a CYP2C8 substrate in humans, or it may recruit a minor CYP pathway in rats that becomes the dominant clearance route in human liver. The direction of error is not systematic. Some molecules look more stable in rats than humans (underprediction of human clearance, a safety risk). Others look less stable (overprediction, leading to false attrition of good candidates). Without knowing the CYP contribution fingerprint in each species for the specific molecule, you can't know which direction to adjust. The scaling factor approach assumes a stable ratio between RLM and HLM clearance that simply doesn't hold across chemical space.
Where Allometric Scaling Succeeds and Fails
Allometric scaling — using body weight or metabolic rate to project from animal to human — works acceptably for compounds where the same CYP isoform is the dominant metabolizer in all species, and where that CYP isoform's expression level scales predictably with body size. High-clearance compounds eliminated predominantly by CYP3A4/3A1/2 often translate reasonably with simple allometric corrections because CYP3A4 is the dominant hepatic CYP across most mammalian species and its expression level does scale with liver size. The failures cluster in three places. First, compounds with significant non-CYP clearance routes: aldehyde oxidase, FMO3, and UGT contributions are poorly predicted from RLM data because those enzymes have different expression levels and substrate preferences across species that don't follow metabolic rate scaling. Second, compounds metabolized by CYP2C19, which is essentially absent in rats — any rat stability data on a CYP2C19 substrate is uninformative about human clearance. Third, compounds with low intrinsic clearance in both species, where the absolute measurement uncertainty is large relative to the difference you're trying to translate.
Computational Species Translation: What We've Built
Our in silico species translation model starts from a CYP contribution prediction — for a given structure, what fraction of metabolism in human liver microsome is attributable to each major CYP isoform, and what is the analogous fingerprint in rat. This is not a single model but a panel: separate regressors for CYP3A4, CYP2C9, CYP2C19, CYP2D6, CYP1A2, CYP2C8 in human, and the rat CYP orthologs where available. The panel is trained on matched human and rat microsomal clearance data with CYP reaction phenotyping data where available — roughly 18,000 compounds with varying degrees of experimental CYP annotation. Where CYP contribution is predicted to be isoform-specific (high confidence in the regressor, narrow prediction interval), we apply isoform-specific scaling factors derived from in vitro–in vivo correlation data for that isoform. Where the CYP contribution is mixed or uncertain, we widen the prediction interval and flag the compound for early human microsome testing before synthesis prioritization. The practical effect is that the model doesn't give you a single point prediction for human clearance from RLM data — it gives you a probability distribution over human clearance, with the width of that distribution reflecting how well the rat-to-human translation is expected to perform for that specific molecule's metabolism fingerprint.
The Non-CYP Pathways Problem
Aldehyde oxidase has become increasingly relevant as CYP liabilities are successfully engineered out of scaffolds. AO is a cytosolic enzyme with high expression in human liver and low, variable expression in rats — a compound that is AO substrate will appear metabolically stable in RLM (which largely lacks AO) and in rat in vivo, and then clear rapidly in human. This is exactly the scenario that produces false confidence. AO metabolism is structurally predictable to a degree: sp2 carbons adjacent to electronegative atoms, particularly in N-heterocyclic systems, are the typical sites. We have a separate AO substrate prediction model that flags high-risk scaffolds for human S9 or hepatocyte testing before they enter lead optimization. The model is not highly accurate in absolute terms — AO site of metabolism prediction is a hard problem — but its precision on high-confidence flagged structures is sufficient to redirect synthesis effort. FMO3 contributions are harder still: the structural features that make a compound an FMO3 substrate are well-characterized for soft nucleophiles like thioethers and tertiary amines, but the enzyme's promiscuity outside those structural alerts is difficult to predict from first principles, and the training data is sparse.
Practical Guidance for Computational Programs
The clearest lesson from running this analysis across our pipeline is that early human microsome data, even if just a single-concentration stability measurement, is worth obtaining at the same time as rat data rather than waiting for in vivo confirmation. The cost differential is small, and it catches the species translation failures before they drive three rounds of synthesis in the wrong direction. For compounds where the computational species translation model predicts high uncertainty — wide prediction interval on human clearance given the rat data — we recommend not relying on RLM data as the primary clearance filter at all. The second lesson is scaffold-specific: N-containing heterocycles, particularly those with AO alert substructures, should be characterized in human S9 early regardless of RLM stability. The third lesson is that the allometric factors published in most PK prediction tools were derived from datasets with different chemical space coverage than modern lead series. Re-deriving correction factors from your own matched RLM/HLM data, even a small internal set of 50–100 compounds, substantially improves translation accuracy for the chemical space you're actually optimizing in.
What This Means for In Silico Campaign Design
For a fully computational campaign, where no wet-lab data exists at lead selection time, the RLM proxy is exactly what we have — predicted, not measured. Our computational ADMET stack predicts both rat and human microsomal clearance independently rather than using rat as a proxy for human. The two predictions are correlated but not constrained to be consistent with a fixed scaling factor, which means the model can express the RLM-to-HLM divergence that characterizes problem scaffolds. Compounds that show the most divergence between predicted RLM and predicted HLM clearance are treated as high-uncertainty on human PK and down-weighted in the Pareto ranking unless they have other compensating advantages. This is not a solution to the species translation problem — it's an acknowledgment of the uncertainty and a way to route that uncertainty into the decision framework explicitly rather than hiding it behind a single predicted number.