Autonomous bioisosteric replacement for multi-property optimization in drug design

Kim, Hyeongwoo; Moon, Seokhyun; Zhung, Wonho; Kim, Shinwoo; Lim, Jaechang; Kim, Woo Youn

doi:10.1038/s41467-026-75512-9

DeepBioisostere

Autonomous Bioisosteric Replacement for Multi-Property Optimization in Drug Design

Hyeongwoo Kim^1,*, Seokhyun Moon^1,*, Wonho Zhung¹, Shinwoo Kim¹, Jaechang Lim², Woo Youn Kim^1,2,†

¹KAIST ²HITS Inc.

^*Equal contribution · ^†Corresponding author

Nature Communications 2026

Paper Code Data PyPI 🤗 Weights

DeepBioisostere proposes replacement fragments that preserve biological activity while moving several molecular properties toward their targets — including bioisosteres that do not appear in the training data.

Abstract

Bioisosteric replacement substitutes one molecular fragment for a chemically or biologically analogous one, tuning a compound's properties without disrupting its activity. Existing in silico approaches largely depend on an expert to nominate the modification site, and struggle to move more than one property at a time.

DeepBioisostere performs the replacement end to end. It selects the site itself, chooses the incoming fragment, and decides how to attach it, conditioned on a target change in several properties at once. Trained on matched molecular pairs drawn from experimental bioassay data, it proposes replacements that lie outside any pre-established substitution table, and we demonstrate its use in computational hit-to-lead optimization.

How the model works

A single forward pass makes three dependent decisions. No modification site is supplied, and no substitution rules are consulted.

Where to cut

The molecule is decomposed along BRICS bonds and every resulting subgraph is scored as a removal candidate.

What to put back

The vacated site is matched against a library of 140,096 insertion fragments, scored jointly with the removal choice.

How to attach it

A fragment can often be joined in more than one orientation, so the attachment points are scored before the product is composed.

Conditioning is applied to the property change rather than its absolute value, so one checkpoint serves any starting molecule and several targets can be commanded together.

Multi-property conditioning

Three input molecules are each given the same pair of opposing commands: hold molecular weight, move logP by −1 and by +1. The generated populations separate along logP while both remain centred on the input's molecular weight.

Joint density of molecular weight against logP for molecule 1 under two opposing logP targets — Figure 2 Joint distributions of molecular weight and logP for the molecules generated under each command. The black dot marks the input. The two contour sets separate along logP; the molecular-weight marginals remain superposed.

Joint density of molecular weight against logP for molecule 2 under two opposing logP targets — Figure 2 Joint distributions of molecular weight and logP for the molecules generated under each command. The black dot marks the input. The two contour sets separate along logP; the molecular-weight marginals remain superposed.

Table 1 Mean achieved change against the commanded target, over 3,000 test molecules.
Commanded	Achieved	Held property
logP −1	−0.84 ± 0.43	M_w −0.28 ± 12.6
logP +1	+0.83 ± 0.38	M_w +0.04 ± 12.8
QED +0.1	+0.07 ± 0.09	M_w −0.11 ± 12.1
QED +0.2	+0.10 ± 0.10	M_w −0.57 ± 12.0
SA −0.5	−0.42 ± 0.43	QED −0.03 ± 0.07
SA −1	−0.72 ± 0.52	QED −0.03 ± 0.07

In every row the commanded property moves and the held property does not, so the two are steered independently rather than traded against each other.

The target determines the modification site

The same molecule under the two opposing commands is cut in two different places. Removing the lipophilic group lowers logP and removing the polar one raises it, so the site the model selects tracks the direction it was asked to move.

Molecule 1 with the polar oxyacetic acid arm marked for removal — Figure 2, removal panels Molecule 1 under each of the two commands: the polar oxyacetic acid arm goes first, the lipophilic cycloheptylmethyl group second. Bold atoms mark the fragment selected for removal. Site selection is part of what the conditioning controls, not a fixed preprocessing step.

Molecule 1 with the lipophilic cycloheptylmethyl group marked for removal — Figure 2, removal panels Molecule 1 under each of the two commands: the polar oxyacetic acid arm goes first, the lipophilic cycloheptylmethyl group second. Bold atoms mark the fragment selected for removal. Site selection is part of what the conditioning controls, not a fixed preprocessing step.

Compatibility with the chemical environment

Two molecules, A and B, share a leaving fragment. The model's own outputs for each (M′_A, M′_B) are compared against a graft (M″_B) in which the fragments chosen for A are transplanted onto B's scaffold.

Density of the logP change for the model's own outputs and for the grafted control — Figure 3 100 generated insertions per molecule per target. logP is close to additive over fragments, so a fragment carries its logP effect with it and the transplant is indistinguishable. QED is a whole-molecule property: the model's own outputs peak near +0.07 while the graft spreads out and drifts negative. The gap is the part a lookup keyed on the leaving fragment could not supply.

Density of the QED change for the model's own outputs and for the grafted control — Figure 3 100 generated insertions per molecule per target. logP is close to additive over fragments, so a fragment carries its logP effect with it and the transplant is indistinguishable. QED is a whole-molecule property: the model's own outputs peak near +0.07 while the graft spreads out and drifts negative. The gap is the part a lookup keyed on the leaving fragment could not supply.

Replacements beyond established rules

For one molecule and one fixed leaving fragment, 100 replacements were generated. The proposals are chemically diverse rather than enumerations of a substitution table, and several do not appear among the training pairs.

QED distribution of the top 50 generated molecules against the original — Figure 4 QED of the top 50 generated molecules; the proposed fragments themselves are the grid shown at the top of this page. The dashed line marks the input molecule, and most proposals fall at or above it.

Hit-to-lead case study

The inputs are raw outputs of a structure-based generative model — compounds that score well against a pocket but are awkward to synthesize. Applying DeepBioisostere downstream improves synthetic accessibility and drug-likeness together, without cost to the docking score.

Case A: the input compound with its outgoing fragment marked, and the generated product with the incoming fragment marked — Figure 5a The outgoing fragment is marked on the input, the incoming one on the product. SA 3.43 → 2.88, QED 0.73 → 0.83, docking −7.40 → −7.60 kcal mol⁻¹.

Case B: the input compound with its outgoing fragment marked, and the generated product with the incoming fragment marked — Figure 5b A second compound from the same pipeline. SA 3.52 → 2.57, QED 0.45 → 0.54, docking −5.50 → −5.60 kcal mol⁻¹.

Lower SA indicates an easier synthesis, higher QED a more drug-like compound, and a more negative docking score a better predicted pose.

Availability

The implementation is distributed through PyPI, with source on GitHub under the MIT licence. Trained weights and the fragment library are hosted on Hugging Face and are retrieved on first use, so neither has to be obtained separately.

The analysis notebooks, the per-figure source data, the docking provenance and the reproduction scripts are archived on Zenodo. That record is self-contained: every figure on this page can be regenerated from it without this repository.

Citation

@article{kim2026deepbioisostere,
  title   = {Autonomous bioisosteric replacement for multi-property
             optimization in drug design},
  author  = {Kim, Hyeongwoo and Moon, Seokhyun and Zhung, Wonho and
             Kim, Shinwoo and Lim, Jaechang and Kim, Woo Youn},
  journal = {Nature Communications},
  year    = {2026},
  doi     = {10.1038/s41467-026-75512-9},
}