Rowan's strain-prediction workflow estimates intramolecular strain by comparing a user-provided 3D structure against a carefully generated low-energy conformer ensemble, using a consistent quantum-chemistry or ML potential for both. The goal is to answer a precise question: "How much higher in energy is the submitted conformation than the best conformer this molecule can adopt under the same model?", while making sure that any optimization respects the original geometry closely enough to remain physically and structurally meaningful.
Computing the strain of a given pose is a common and important problem across various fields of chemistry, including catalysis (through distortion–interaction models) and structure-based drug design. Unfortunately, naïve approaches to strain computation often give bizarre or unphysical energies. Since different levels of theory predict different bond lengths, comparing a conformation produced by e.g. a Vina forcefield to a structure optimized with higher-level methods often predicts vast strain energies (>100 kcal/mol) because each individual bond is predicted to be "strained."
While technically correct within a given level of theory's paradigm, this result doesn't qualitatively match what scientists expect to get from a strain prediction; since the exact forcefield used to generate docked poses is known to be inaccurate, it's not helpful to point out that a visually "correct" conformation has minutely different bond lengths than DFT would predict. Instead, a good strain-prediction method should be able to evaluate the strain of the overall conformation without being overly sensitive to individual bond lengths or angles.
From the user-provided input structure, a conformer search is run using the requested settings and level of theory. This search generates a diverse ensemble of low-energy conformers and refines them with the specified optimization protocol, recording single-point energies along the way. The resulting ensemble is treated as an approximate reference distribution of accessible conformations; in particular, its minimum energy defines the baseline "unstrained" state used for later comparisons.
In parallel, starting from the original structure, Rowan runs an optimization using the same level of theory but augmented with a harmonic restraint that penalizes deviations from the starting coordinates. Heavy atoms (and optionally hydrogens) are kept close to their initial positions via this added quadratic term, so the optimizer can eliminate obvious artifacts—bad angles, clashes, minor distortions—without allowing the structure to collapse into an unrelated conformer. If multiple optimization stages are requested, earlier stages use their own settings but retain the same harmonic-restraint logic, stepping the structure towards a locally relaxed, chemically sensible version of the submitted pose. This approach, also used by OpenEye's FreeForm, prevents large-scale reorganization while allowing local optimization.
Once this restrained optimization is complete, we evaluate the energy of the refined pose and compare it to the conformer ensemble. The optimized pose is re-evaluated without changing its geometry, and its electronic (or ML) energy is obtained. In parallel, the conformer ensemble is scanned to determine its lowest energy under the same scoring scheme; if the optimized pose appears to be lower in energy than this provisional minimum, an additional unconstrained refinement is triggered internally to update the reference and avoid artifacts from the restraint. The final strain energy is reported as the difference between the pose energy and the best ensemble energy, converted to kcal/mol.