In-Silico Screening

Computational methods for evaluating potential drug molecules before physical synthesis — replacing or filtering the experimental high-throughput screening (HTS) that processes millions of compounds in the wet lab. In-silico screening reduces the cost and time of the "make-test" cycle by eliminating candidates that are predicted to fail on absorption, distribution, metabolism, excretion, toxicity (ADMET) or binding affinity criteria.

The field has matured significantly since 2020. Three layers now exist:

1. Structural docking — Predicts how a small molecule fits into a protein binding site (Glide, AutoDock, Schrodinger Suite). Accuracy limited by protein conformational flexibility; AlphaFold3 improved this substantially.
2. Free energy perturbation (FEP) — Rigorous physics-based calculation of binding free energy; most accurate but computationally expensive. Schrodinger's FEP+ is the commercial standard; increasingly combined with ML for speedup.
3. ML-based ADMET prediction — Neural networks trained on experimental ADMET data predict solubility, permeability, CYP inhibition, hERG activity, etc. Certara's Simulations Plus and ADMET Predictor are commercial tools; deep learning models now rival or exceed expert rules.

The key shift in 2024–2026: Rather than screening a fixed library of known compounds, AI generative models propose novel compounds whose predicted properties already satisfy screening criteria — the virtual screen comes before synthesis, not after. This is the integration point between in-silico screening and generative biology.

Key Claims

• Schrodinger's FEP+ is the commercial gold standard for binding affinity — Used by top-10 pharma; Schrodinger's own pipeline (SDGR) is transitioning from platform licensing to internal asset development. Evidence: strong (Schrodinger SEC filings, multiple pharma citations)
• AlphaFold3 dramatically improved the quality of structure-based in-silico screening — All-atom prediction of protein-ligand complexes removes the need for experimental crystal structures in many cases. Evidence: strong (Nature 2024)
• Certara/Simulations Plus PBPK/ADMET models are used by regulators and pharma — Model-informed drug development (MIDD) is now FDA-accepted for dose optimization, reducing Phase 1/2 trial burden. Evidence: strong (FDA guidance on MIDD, Certara SEC filings)
• Virtual screening is faster than HTS by 3–4 orders of magnitude — HTS screens ~1M compounds in weeks; generative AI + virtual screen evaluates billions in hours. The caveat is that quality metrics diverge: HTS tests compounds in real biology. Evidence: moderate
• IsoDDE roughly doubles AlphaFold-3's accuracy on the hardest protein-ligand cases — Isomorphic's Drug Design Engine reports 50% accuracy vs AlphaFold 3's 23.3% on the hardest protein-ligand structure-prediction cases (technical report 2026-02-10) — a direct improvement in the docking/structure layer that underpins structure-based virtual screening. Evidence: moderate (Isomorphic IsoDDE)
• Recursion's Boltz-2 approaches FEP accuracy ~1,000× faster — Per the IsoDDE-context report, Recursion's Boltz-2 (with MIT) reaches free-energy-perturbation-grade binding accuracy at roughly 1,000× the speed, illustrating the trajectory from physics-based FEP toward ML surrogates. Evidence: weak (Isomorphic IsoDDE)
• ESMFold runs much faster than AlphaFold2 with only minor accuracy loss — The 2026 ESM survey confirms ESMFold's speed advantage, but flags that ESM still loses to CNNs on some structure tasks and generalizes poorly out-of-domain — bounding how far a protein-LM alone can carry structure-based screening. Evidence: strong (ESM survey 2026)
• Unified generative models are folding "screen" into "generate" for the structure step — UniGenX co-generates sequence + coordinates under functional objectives (23× protein induced-fit improvement, RMSD < 2 Å), pushing the structural-confidence step upstream into generation rather than running it as a separate post-hoc docking screen. Evidence: strong (UniGenX, arXiv:2503.06687)

Open Questions

• How well do ADMET prediction models generalize to novel chemical scaffolds outside their training distribution?
• At what point does in-silico confidence justify skipping HTS entirely vs using it as a first-pass filter?
• Can multi-objective in-silico optimization reliably predict oral bioavailability for complex molecules (PPI inhibitors, macrocycles)?

Related Concepts

• Protein & Molecule Foundation Models — foundation models improve the quality of virtual screening
• Lab-in-the-Loop — in-silico screening feeds the design step of the closed loop
• The Data Moat Question — ADMET models are only as good as the experimental data they train on

Changelog

• 2026-06-15 — Initial compilation; structural docking, FEP, ML-ADMET, AlphaFold3 integration covered
• 2026-06-24 — Compiled new sources (unigenx-foundation-model, esm-protein-lm-survey-2026, isomorphic-isodde-human-trials)