PROTEUS: Methods for Quantum-Accelerated Molecular Discovery

1Introduction and scope

Drug discovery is a search problem over an astronomically large space under a punishing failure rate. PROTEUS narrows that search in silico so that scarce wet-lab effort is spent only on molecules worth making.

The drug-like chemical space is estimated at up to 10⁶⁰ molecules; a discovery campaign can afford to synthesize only a few dozen. The job of computation is not to replace the assay but to raise the prior probability that a synthesized molecule is potent, selective, safe, and makeable. PROTEUS targets four decision points where that prior can be raised the most.

1.1Design philosophy

1. Learn the empirical, compute the fundamental. Sequence-to-structure and structure-activity relationships are learned from data; binding energetics, where electron correlation dominates, are computed with quantum chemistry.
2. Chemists in the loop. Every stage exposes an interpretable output and a gate at which a human decides. PROTEUS proposes; medicinal chemists dispose.
3. Uncertainty everywhere. No prediction ships without a calibrated confidence, so borderline calls are escalated rather than trusted blindly.
4. The assay is ground truth. Models are judged prospectively against measurement, and every measurement is fed back to improve them.

1.2Scope and non-goals

This document covers small-molecule discovery against a defined protein target with a known or predictable binding site. It does not cover biologics, and it does not claim that quantum hardware today outperforms classical methods on every task, §10 is explicit about where it does not. PROTEUS is a discovery accelerator, not an autonomous chemist.

2Pipeline overview

The four methods form a loop. Each narrows the candidate set and enriches it, and assay data from synthesized leads flows back to retrain the learned components.

Sections 3–6 describe each method. Section 7 describes the active-learning loop that connects them to the lab. Sections 8–9 cover data and validation; Section 10 is a candid account of limitations.

3Fold — structure and pockets

Fold turns a target's amino-acid sequence into a three-dimensional structure with per-residue confidence, then locates and ranks the pockets a small molecule could bind.

3.1Inputs and representation

The input is the target sequence, augmented by a multiple-sequence alignment (MSA) that encodes evolutionary covariation, residue pairs that co-vary across homologs tend to be in contact, and optional structural templates. The model maintains two coupled representations: a per-residue (single) representation and a residue-pair representation that carries geometric relationships.

3.2Structure model

An attention-based trunk (Evoformer-style) iteratively refines the single and pair representations, exchanging information between them, before a structure module places atoms in 3D. Attention over residues takes the standard form, biased by the pair representation b_ij:

The structure module predicts a backbone frame (a rotation and translation) per residue plus side-chain torsion angles, and the whole trunk is run recyclingly so later passes refine earlier ones.

3.3Confidence

Two confidence signals ship with every structure. pLDDT is a per-residue estimate of local accuracy (0–100); PAE (predicted aligned error) estimates the positional error between residue pairs and reveals which domains are placed reliably relative to one another. Both are used downstream: a pocket sitting in a low-confidence region is treated with appropriate caution.

3.4Pocket detection

Cavities on the predicted surface are enumerated geometrically and scored by volume, enclosure, hydrophobicity, and evolutionary conservation into a druggability score. The top-ranked pockets define the conditioning target for Forge (§4).

4Forge — generative design

Forge invents new molecules inside a target pocket, optimizing potency, selectivity, physicochemical properties, and synthesizability at once, and proposes only molecules that pass a retrosynthetic check.

4.1Pocket-conditioned diffusion

Molecules are generated by an equivariant diffusion model that builds a 3D molecular graph conditioned on the pocket. Training corrupts a known ligand by adding Gaussian noise over T steps; generation reverses the process. The forward step is

and a neural network ε_θ, made SE(3)-equivariant so predictions rotate and translate with the input, learns to denoise. Sampling then integrates the reverse process conditioned on the pocket P:

Because generation happens inside the pocket, the geometric complementarity that governs binding is built in rather than filtered for afterward.

4.2Multi-objective optimization

A raw generative model produces valid but not necessarily useful molecules. Forge steers generation toward a weighted objective over interpretable properties, then fine-tunes the sampler with reinforcement learning against that reward:

where QED is quantitative drug-likeness and SA is a synthetic-accessibility score. Weights w_k are set per program, a CNS target and an oncology target want different profiles.

4.3Synthesizability gate

Every proposed molecule is passed through a retrosynthesis model that attempts to find a plausible route from purchasable building blocks. Molecules without a route are rejected. This keeps Forge honest: what it proposes, a medicinal chemist can actually make.

Generation loop (simplified)for pocket in targets:
    cands = diffusion.sample(pocket, n=10_000)     # pocket-conditioned
    cands = [m for m in cands if valid(m) and novel(m)]
    scored = [(m, reward(m)) for m in cands]      # multi-objective
    routes = retrosynthesis.plan(top_k(scored, 2000))
    return [m for m, r in routes if r.feasible]     # synthesizable only

4.4Novelty and validity

Forge reports validity (chemically sensible structures), uniqueness (non-duplicated), and novelty (Tanimoto distance from the training set and known actives), so a reviewer can tell genuine invention from recall of known chemotypes.

5Bind — quantum binding energy

Whether a molecule binds is decided by electrons. Bind computes binding energetics with a variational quantum eigensolver on the chemically active region, where classical density-functional theory makes its largest, least predictable errors.

5.1Binding free energy

The quantity of interest is the binding free energy, related to the dissociation constant by

Reaching chemical accuracy (about 1 kcal/mol) on ΔG matters because a factor of ten in potency is roughly 1.4 kcal/mol, so errors larger than that reorder your candidates.

5.2Active-region embedding

A full protein–ligand complex is far too large for a quantum computer. Bind isolates the small, strongly correlated active region (the binding-site residues and the ligand's interacting groups) and treats it quantum-mechanically while embedding it in the classical environment via a QM/MM and projection-based (embedding) scheme. Only the active region's electronic structure is sent to the quantum solver.

5.3Variational quantum eigensolver

The active region's electronic Hamiltonian, in second-quantized form, is

mapped to qubits (Jordan–Wigner or Bravyi–Kitaev). VQE prepares a parameterized trial state and uses the variational principle, the expectation of H is an upper bound on the true ground-state energy, so minimizing it over the parameters approaches the answer:

The trial state uses a unitary coupled-cluster (UCCSD) ansatz built from the Hartree–Fock reference:

A classical optimizer proposes parameters; the quantum backend estimates the energy by measurement; the loop repeats until convergence, the same hybrid pattern used throughout PROTEUS.

5.4Error mitigation and validation

On real hardware, Bind applies readout-error calibration, zero-noise extrapolation, and reference-molecule checks, and validates against high-accuracy classical references (coupled-cluster, and full configuration interaction where tractable) and experimental affinities. Where quantum execution is not yet advantageous, the same interface runs a classical correlated method and the router records which was used.

6Screen — ADMET and developability

Most candidates die on safety and developability, not potency. Screen predicts dozens of endpoints with calibrated uncertainty and triages millions of molecules to the handful worth synthesizing.

6.1Endpoints and models

Screen predicts absorption, distribution, metabolism, excretion, and toxicity endpoints with a shared multitask graph neural network over the molecular graph, so related endpoints share representation and small-data endpoints borrow strength from large ones.

6.2Calibrated uncertainty

A point prediction without a confidence is dangerous in triage. Screen uses conformal prediction to attach statistically valid uncertainty: for a target coverage 1−α, it returns a prediction set (or interval) guaranteed to contain the truth at that rate on exchangeable data.

Molecules whose interval straddles a decision threshold are escalated to a human rather than silently passed or failed.

6.3Triage cascade and applicability domain

Filters run cheapest-first, potency, then selectivity, then ADMET, then developability, so expensive predictions only run on survivors. Each prediction is checked against the model's applicability domain: a molecule too far from the training distribution is flagged as an extrapolation, not scored with false confidence.

7The active-learning loop

PROTEUS is not a one-shot predictor. Each campaign generates assay data, and that data is the most valuable signal in the system. The loop chooses what to make next to learn the most.

7.1Closing the loop

Synthesized molecules are assayed for potency, selectivity, and key ADMET liabilities. Results are written back to the training corpus with full provenance, and Fold, Forge, and Screen are periodically retrained. Bind, being first-principles, is not retrained but is recalibrated against measured affinities.

7.2Acquisition

Which molecules to synthesize next is a design-of-experiments question. PROTEUS selects a batch that balances exploitation (high predicted value) against exploration (high model uncertainty), maximizing an acquisition function over the candidate pool:

where μ is predicted value, σ its uncertainty, κ trades off the two, and a redundancy penalty keeps the batch chemically diverse so a synthesis round is not spent on near-duplicates.

8Data and reproducibility

Method quality is bounded by data quality and by the discipline of the splits used to measure it. PROTEUS is explicit about both.

8.1Training data

8.2Splits and leakage

Random splits flatter models by leaking near-neighbors between train and test. PROTEUS reports on the splits that matter for real deployment:

• Scaffold split for chemistry models, so the test set contains genuinely different molecular frameworks.
• Temporal split for structures and assays, training only on data available before a cutoff and testing on what came after, mirroring prospective use.
• Target split for binding, holding out entire protein families to test generalization, not interpolation.

8.3Reproducibility

Data snapshots, feature definitions, model weights, and configs are versioned together, and every reported number is tied to a specific pipeline version and split definition so it can be reproduced.

9Validation and benchmarks

The only validation that counts is prospective: does a molecule the model liked behave as predicted when it is made and measured? We report retrospective benchmarks to calibrate expectations and prospective results to prove them.

9.1Metrics

Structure is scored by backbone accuracy and pocket recovery; generative design by validity, novelty, and hit enrichment over a baseline library; binding by mean absolute error against reference and experiment; ADMET by area under the curve and by calibration error, not accuracy alone.

The verdict column is stated plainly: quantum helps on the correlated binding step in simulation and on small active regions today; the learned methods carry most of the retrospective lift now.

10Limitations and current maturity

This section is deliberately blunt. Drug discovery has a long history of computational overpromising; credibility with medicinal chemists depends on candor.

• NISQ reality. Quantum hardware is noisy and small. Bind's quantum advantage is established on modest active regions and in simulation; large systems still rely on classical correlated methods, and PROTEUS routes to them automatically.
• Distribution loading and scaling. Efficient preparation of the required quantum states, and scaling active regions, are active research problems that can erode theoretical advantage.
• Generative models hallucinate. A synthesizability gate and human review are mandatory precisely because generative models can propose confident nonsense.
• Prediction is not proof. Structure prediction can be wrong in flexible regions; ADMET models extrapolate poorly off-distribution. The applicability-domain check exists because of this.
• The assay is ground truth. Every number here is a prior to be tested at the bench, not a substitute for it. Simulated figures in this document are illustrative of targets, not measured claims.

11Responsible use and intellectual property

Generative chemistry is dual-use. PROTEUS is operated for therapeutic discovery under partner agreements, with safeguards against the design of harmful compounds and controls appropriate to the chemistry involved. Partner data and generated matter remain the partner's intellectual property; deployment is in a secure enclave, and models trained on partner data are not shared across partners. Provenance is retained so inventorship and the origin of every proposed molecule can be established.

12Roadmap

—References and further reading

1. Jumper et al. Highly accurate protein structure prediction with AlphaFold. Nature 596 (2021).
2. Abramson et al. Accurate structure prediction of biomolecular interactions (AlphaFold 3). Nature (2024).
3. Hoogeboom et al. Equivariant Diffusion for Molecule Generation in 3D. ICML 2022.
4. Schneuing et al. Structure-based Drug Design with Equivariant Diffusion Models. 2022.
5. Olivecrona et al. Molecular de novo design through deep reinforcement learning (REINVENT). J. Cheminformatics 9 (2017).
6. Peruzzo et al. A variational eigenvalue solver on a photonic quantum processor. Nature Communications 5 (2014).
7. Romero et al. Strategies for quantum computing molecular energies using the unitary coupled cluster ansatz. QST 4 (2019).
8. Sun & Chan. Quantum embedding theories. Acc. Chem. Res. 49 (2016).
9. Angelopoulos & Bates. A Gentle Introduction to Conformal Prediction. 2021.
10. Su et al. Comparative Assessment of Scoring Functions (CASF). J. Chem. Inf. Model. (2019).

This document describes methods and design intent. All performance figures are illustrative and simulated for demonstration; production characteristics are established per program and validated experimentally. © Ace Hacker Research & Development Institute.