Physics-based affinity prediction enables better therapeutics
PrimaryDeep Origin's free-energy perturbation and relative binding free energy workflows rest on the causal theory that more accurate physics-based estimates of ligand binding affinity can improve medicinal chemistry decisions. If compound affinity changes can be predicted reproducibly before synthesis or extensive wet-lab testing, drug programs should prioritize molecules with stronger target engagement and eliminate weaker candidates earlier. A testable prediction is that programs using these workflows should show improved hit-to-lead or lead-optimization efficiency, with prospective affinity predictions correlating with measured biochemical or cellular potency. For age-related disease, the mechanism is indirect: better molecular optimization should increase the chance of producing effective drugs against disease-relevant targets, but the provided material does not establish a specific aging-pathway mechanism.
Popperian evaluation
The core premise is credible: ligand binding affinity is a real driver of medicinal chemistry, and physics-based RBFE/FEP methods can estimate relative affinity before synthesis. The weak point is practical, not conceptual. The benchmark shows accuracy depends heavily on the protein system, input structure quality, and ligand transformation, so the theory should read as conditional: these workflows help when the model is accurate enough for the specific series being optimized.
Supporting evidence: OpenFE benchmarked RBFE calculations across more than 1,700 ligands from public and blinded private datasets.; Public datasets had weighted RMSE around 1.73 kcal/mol, and 10 of 58 systems reached sub-kcal/mol accuracy.; The protocol produced reproducible results with enough throughput and convergence for industrial use.
Counter evidence: Private industrial datasets had weaker performance, with weighted RMSE around 2.44 kcal/mol and only 2 of 37 systems reaching sub-kcal/mol accuracy.; The benchmark found no single dominant error source, which means failures may be hard to predict in advance for a new program.; Better affinity does not automatically mean better cellular potency, pharmacokinetics, safety, or clinical efficacy.
The theory explains why a drug team might synthesize fewer weak compounds during lead optimization: if prospective affinity estimates rank analogs correctly, chemistry resources can move toward stronger binders earlier. It does not yet explain broad therapeutic success on its own. Alternative explanations, such as better target selection, assay quality, chemistry judgment, ADME filtering, and project management, could also produce faster hit-to-lead progress. For aging biology, the explanatory link is thinner because the material does not name a disease pathway that affinity prediction directly modifies.
Supporting evidence: The reasoning chain connects prospective affinity prediction to compound prioritization and earlier elimination of weaker candidates.; The OpenFE evidence supports reproducible production-scale RBFE performance, which is the technical condition needed for the theory to matter in real campaigns.; Resistance profiling evidence supports the broader principle that prospective prediction can improve targeted therapy decisions when predictions match later clinical or experimental outcomes.
Counter evidence: The evidence supplied is mostly benchmark performance, not direct proof that Deep Origin workflows improve hit-to-lead or lead-optimization outcomes.; Affinity is only one optimization axis; potency, selectivity, solubility, permeability, metabolism, toxicity, and developability can dominate project decisions.; The age-related disease claim is indirect and has low support because no specific aging mechanism is established.
This theory is testable and can fail cleanly. A prospective study could ask whether predicted affinity changes correlate with later biochemical or cellular potency, then compare synthesis efficiency, active-compound yield, and lead-optimization cycle time against matched programs that did not use reliable RBFE/FEP ranking. If predictions fail to rank compounds within the relevant chemical series, or if correct rankings do not improve decisions, the practical theory takes a direct hit.
Supporting evidence: The theory predicts that prospective affinity predictions should correlate with later measured biochemical or cellular potency.; It also predicts improved hit-to-lead or lead-optimization efficiency compared with programs lacking reliable prospective affinity prediction.; The OpenFE benchmark already uses measurable error statistics such as RMSE and sub-kcal/mol accuracy, which can be reused prospectively.
Counter evidence: The supplied material does not define a pass threshold for correlation, RMSE, decision lift, or cycle-time improvement.; Program-level efficiency has many confounders, so a weak study design could make the theory look right or wrong for the wrong reason.; Cellular potency may diverge from binding affinity for reasons unrelated to the affinity model.
Reasoning tree
Public endorsements
The public material here places Andrew Mochalskyy at Deep Origin as CTO and shows the company publicly talks about physics-informed drug discovery and in-silico models. It does not provide a direct public statement from Mochalskyy endorsing, explaining, or disputing the specific theory that physics-based affinity prediction improves medicinal chemistry decisions. On this record, he is publicly silent on the theory itself.
Evidence publication IDs: 60df4213-a51b-41e9-a0e2-9a72030e69c1, 312f46ed-5e59-4caf-ad84-e0bfbc5514ea
Papoian is a public face of Deep Origin as co-founder and CSO, and the supplied materials place him in public discussions of the company’s physics and molecular-modeling approach. He was scheduled to give a technical deep dive on multiscale molecular models, and company materials tie him to prospective computational drug discovery work. But the dossier does not give a direct Papoian quote endorsing the specific causal claim that physics-based affinity prediction improves medicinal chemistry decisions, so this is mention rather than a clean explicit endorsement.
Evidence publication IDs: 9cad53c0-f4f7-4632-9fba-07cf2c0ca823, 312f46ed-5e59-4caf-ad84-e0bfbc5514ea, 41bffb22-cce6-4588-a8ad-94127ddf7108
The provided public evidence does not show Garik Petrosyan endorsing, mentioning, or contradicting Deep Origin's specific theory that physics-based binding-affinity prediction improves medicinal chemistry decisions. The cited 2025 publication discusses active learning, distributed quantum chemistry, and molecular energy prediction, but not free-energy perturbation, relative binding free energy, or prospective ligand-affinity prediction for therapeutic optimization.
Evidence publication IDs: a96ffa15-65a1-4484-9bb0-f8fb213f3dd1
The provided evidence does not show Garik Petrosyan making a public statement about physics-based affinity prediction or Deep Origin's free-energy workflows. The only record is a Deep Origin team page identifying him as Head of AI, which establishes affiliation, not an explicit endorsement, mention, or contradiction of the theory.