Inverse Design of Substituted BODIPY

We provide python codes to inverse design BODIPY molecules, as discussed in Ref-1.

DesignBodipy_Bayes.py can be used to design molecules using Bayesian optimization based on Gaussian process regression.

DesignBodipy_GA.py can be used for genetic algoritm (GA) optimization

Both programs use a trained kernel ridge regression machine learning (KRR-ML) model to evaluate the S₀→S₁ excitation energy.

Example run: Bayesian Optimization

$ python3 DesignBodipy_Bayes.py <target(eV)>

Additional parameters can be sought using --help argument. Given below are all possible flags.

---

Flag/ arg	Description	Default [range]	Compulsory
[target]	Target S0->S1 value, in eV. Positional argument, non optional.	-	✓
--group, -g	# of substitutions in target BODIPY.	2 [2, 7]	✗
--data -d	Location of datafiles to be used in KRR ML, contains descriptor and coefficients.	./data	✗
--restart, -r	# of evaluations for single EI evaluation. More evaluations give more robust minima, with higher computation cost.	5 [1, ∞]	✗
--exploration, -x	Exploitation vs Exploration parameter	0.01 (0,100)	✗
--seed, -s	Number of initial evaluations to build Gaussian Process surrogate. More evaluations might help converging faster.	5 [1, ∞]	✗
--iter, -i	Maximum number of iterations.	200 [1, ∞]	✗

---

Once run, it will run for iter times and print successive improvements towards obtaining target molecule. An example run is shown below:

$ python3 DesignBodipy_Bayes.py 2.7

Searching for 2D BODIPY near 2.700000 eV
Reading ML model from ./data
Iterations 200; Initial evaluations 5
Bayesian opt. parameters:
 Exploration/Exploitation param: 0.010000; Eval. per EI: 5
=================================================================
ITER    POS            GROUPS             S0S1(eV)        Target
=================================================================
0       1 6             28 27           3.337201        2.700000
1       3 5             29 30           3.184931        2.700000
2       5 6             27 22           3.183506        2.700000
13      4 5             30 25           2.999981        2.700000
18      5 7             23 19           2.952890        2.700000
38      2 5             15 25           2.866237        2.700000
83      5 4             34 6            2.709659        2.700000
=================================================================

Example run: GA Optimization

$ python3 DesignBodipy_GA.py <target(eV)>

Requirements:

1. Python3.6 and above
2. Numpy
3. Scipy (scipy.optimize.minimize for iter minimization)
4. Scikit-learn (for Gaussian Process)
5. MOPAC for calculating minimum energy geometry at the PM7 level
6. OBabel for file conversion
7. QML for calculating the SLATM descriptor using the PM7 geometry

Relevant resources

A publicly accessible web interface hosting a trained machine learning (ML) model to predict S₀ → S₁ excitation energy of BODIPYs is available at https://moldis.tifrh.res.in/db/bodipy.

Reference:

[1] Data-Driven Modeling of S0 -> S1 Transition in the Chemical Space of BODIPYs: High-Throughput Computation, Machine Learning Modeling and Inverse Design,
Amit Gupta, Sabyasachi Chakraborty, Debashree Ghosh, Raghunathan Ramakrishnan
submitted (2021) arxiv
Dataset: https://moldis-group.github.io/BODIPYs/
Dataset DOI: 10.6084/m9.figshare.16529214.v1