Investigating errors in alchemical free energy predictions using random forest models and GaMD

State of the art in silico ∆∆G predictions for antibody-antigen complexes achieve an accuracy of ±1 kcal/mol. While this is sufficient for high throughput screening or affinity maturation, it is insufficient for assessing the criticality of post-translational modifications (PTMs) during clinical development. PTMs that impair binding by >50% pose a major risk to achieving the desired therapeutic bioactivity and must be controlled within defined limits to ensure product quality. A 50% loss in the dissociation constant (Kd) corresponds to a ∆∆G of +0.5 kcal/mol, thus requiring a ±0.5 kcal/mol accuracy threshold for in silico predictions to be practically actionable in clinical phases. In this work, we use conventional molecular dynamics thermodynamic integration (cMD-TI) to generate ∆∆G predictions and develop an error analysis approach using random forest models and end state Gaussian accelerated molecular dynamics (GaMD). This approach provides insight into inadequate sampling of key degrees of freedom (DOF) using only cMD-TI and end state GaMD. We identify bulky side chain undersampling and violation of energetically relevant interatomic interactions as major sources of error, and our GaMD-based error corrections lead to > 1 kcal/mol improvements in accuracy in our most erroneous cases. When applied to a set of 13 predictions, the GaMD-based error correction reduced the root mean square error (RMSE) from 1.01 kcal/mol to 0.69 kcal/mol. This work introduces the application of alchemical free energy predictions to estimating PTM impacts on bioactivity and addresses the current errors that limit their practical use in clinical development.