The effectiveness of the HDO method is demonstrated using two types of optimizing tasks on datasets from Shields’s study (Task 1) and on in-house data accumulated using HTE equipment (Task 2).
Task 1: A dataset comprising reaction conditions pertaining to the Suzuki-Miyaura reaction (1a) was released by Shields. They conducted high-throughput experiments on the class of Suzuki-Miyaura cross-coupling reactions. Twelve couplings of electrophiles and nucleophiles across the combinations of 11 ligands, 7 bases, and 4 solvents were considered, thereby resulting in combinations for a total of 3,696 reactions with a product. Buchwald-Hartwig reaction (2a-2e): They also conducted high-throughput experiments on the class of Pd-catalyzed Buchwald-Hartwig C-N cross-coupling reactions. They experimented with combinations of 3 aryl halides, 4 catalyst, 3 bases, and 22 additives for a total of 792 reactions per target product, of which there were five types. Arylation (3a): They studied the arylation of imidazoles, a key step in the commercial synthesis of the JAK2 inhibitor BMS-91154344,45. They selected a subspace consisting of 1,728 reactions including 12 ligands, 4 bases, 4 solvents, 3 temperatures, and 3 concentrations as a tractable set of experiments to be used as the ground truth. The data for the arylation reaction (3a) included results contributed by 50 expert chemists and engineers from academia and industry, who played the reaction optimization game. All of the data used in Task 1 are accessible from https://github.com/b-shields/edbo/ .
Task 2: Because HDO provides more general evaluations compared with skilled synthetic experts, a total of 22 experiments were conducted on four named reactions (details of Task 2 appear in Table 1). In this task, HDO was evaluated by comparing the yield results of the reaction conditions proposed by five experts in organic synthesis. A search space for each of the Suzuki-Miyaura (4a-4j), Buchwald-Hartwig (5a-5h), Ullmann (6a, 6b), and Chan-Lam reactions (7a, 7b) was defined by the experts and details thereof are provided in Table 1. Details are described in Supplement Information (S1 and S2).
Table 1. Details of the two performance benchmarking tasks
Summary of task sets with definition of search space
|
|
|
Task1 (vs. baselines, 50 humans)
|
Task2 (vs. only 5 humans)
|
|
Target conditions /Reaction type
|
Suzuki-Miyaura (1a)
|
Buchwald-Hartwig (2a-2e)
|
Arylation (3a)
|
Suzuki-Miyaura (4a-4j)
|
Buchwald-Hartwig (5a-5h)
|
Ullmann Reaction (6a, 6b)
|
Chan-lam Reaction (7a, 7b)
|
Target number of conditions
|
Reactant1
|
3
|
-
|
-
|
-
|
-
|
-
|
-
|
Reactant2
|
4
|
3
|
-
|
-
|
-
|
-
|
-
|
Additive
|
-
|
22
|
-
|
-
|
-
|
-
|
-
|
Catalyst
|
-
|
4
|
-
|
9
|
9
|
7
|
7
|
Base
|
7
|
3
|
4
|
8
|
9
|
13
|
3
|
Solvent
|
4
|
-
|
4
|
4
|
5
|
5
|
4
|
Ligand
|
11
|
-
|
12
|
12
|
12
|
9
|
-
|
Concentration
|
-
|
-
|
3
|
-
|
-
|
-
|
-
|
Temperature
|
-
|
-
|
3
|
-
|
-
|
-
|
-
|
Search space area
|
3,696
|
792
|
1,728
|
3,456
|
4,860
|
4,095
|
84
|
> 95% (yields)
|
1.92%
|
0.48%
|
0.58%
|
-
|
-
|
-
|
-
|
Number of target products
|
1
|
5
|
1
|
10
|
8
|
2
|
2
|
Reactions for training
|
1,227,756
|
8,541
|
49,625
|
158,605
|
17,705
|
10,518
|
2,694
|
Task 1: Optimization of reaction conditions to benchmark the performance
We investigated the effectiveness of the proposed method for optimizing the chemical conditions on the Suzuki-Miyaura, Buchwald-Hartwig, and Arylation reactions. We first targeted the number of trials (NT) to identify the top 1, 5, and 10% target yields in the entire search space of reaction conditions as a performance measurement. We calculated the average improvement rate (AIR) over a random selection of NTrs as performance indicator to allow for comparison with other baselines (Eq. (6)). Here, N is the number of epochs required for the overall optimization process and we used 50 epochs for all the models. In each optimization model, NTnc is the NT in the c category of baseline models (Random Forest (RF)46, BO42, humans42, MPNN47, and HDO), respectively. AIR represents the average improvement in the performance of a model over the naïve approach that one might use as a starting point for experimental optimization. In other words, AIR is the extent (percentage) to which the model differs from the naïve model. AIR 0 means the same number of trials with random selection. When AIR has a negative value, the number of trails to find the combination of conditions with a target yield exceeds the random selection.
Table 2. Comparison of reaction optimization performance with the baselines on Task 1.
Table 2. Summary of evaluation
Reaction
|
target
|
Top%
|
Target yield
|
RandomSearch
|
RF
|
BO
|
MPNN
|
Human (50 experts)
|
HDO (Ours)
|
|
|
|
|
ANTr
|
Average Improvement Rate over RS (AIR)
|
Suzuki- Miyaura
|
1(a)
|
1
|
96.20
|
98.15
|
0.1844
|
0.5885
|
0.6589
|
-
|
0.6685
|
5
|
92.52
|
19.55
|
0.0925
|
0.3750
|
0.4740
|
-
|
0.4740
|
10
|
88.19
|
9.51
|
0.2460
|
0.4440
|
0.6900
|
-
|
0.6900
|
Buchwald- Hartwig
|
2(a)
|
1
|
52.67
|
99.15
|
0.2244
|
0.7585
|
-0.1985
|
-
|
0.6642
|
5
|
47.94
|
18.54
|
0.1175
|
0.4245
|
-0.1725
|
-
|
0.3945
|
10
|
44.65
|
8.99
|
0.1060
|
0.4850
|
0.3850
|
-
|
0.4420
|
2(b)
|
1
|
83.09
|
97.54
|
0.2046
|
0.5846
|
0.5149
|
-
|
0.6885
|
5
|
79.06
|
19.54
|
0.0745
|
0.4395
|
0.3220
|
-
|
0.5505
|
10
|
73.58
|
9.15
|
0.1850
|
0.4550
|
0.3850
|
-
|
0.5850
|
2(c)
|
1
|
94.34
|
98.01
|
0.2844
|
0.5545
|
-0.0846
|
-
|
0.5585
|
5
|
86.76
|
19.44
|
0.0675
|
0.3725
|
-0.0670
|
-
|
0.3425
|
10
|
81.27
|
8.95
|
0.1890
|
0.5420
|
-0.2510
|
-
|
0.3850
|
2(d)
|
1
|
65.46
|
98.18
|
0.2549
|
0.6185
|
-0.1785
|
-
|
0.6745
|
5
|
52.56
|
19.24
|
0.2175
|
0.1940
|
0.1440
|
-
|
0.2210
|
10
|
49.15
|
8.97
|
0.2850
|
0.5010
|
0.3500
|
-
|
0.5850
|
2(e)
|
1
|
97.56
|
99.54
|
0.3085
|
0.6785
|
-0.2485
|
-
|
0.6986
|
5
|
91.05
|
19.54
|
0.2245
|
0.3945
|
0.1245
|
-
|
0.4125
|
10
|
86.25
|
8.98
|
0.1850
|
0.3850
|
0.1200
|
-
|
0.4850
|
Arylation
|
3(a)
|
1
|
91.21
|
98.66
|
0.1846
|
0.7485
|
0.4785
|
0.6382
|
0.7285
|
5
|
76.54
|
19.81
|
0.2420
|
0.4210
|
0.3345
|
0.4225
|
0.4560
|
10
|
59.11
|
9.43
|
0.4440
|
0.5850
|
0.4210
|
0.6750
|
0.6950
|
Bold numbers indicate the highest performing model for each reaction target.
\(AIR=1/N{\sum }_{n=1}^{n}\frac{\left({NT}_{n}^{r}-{NT}_{n}^{c}\right)}{{NT}_{n}^{r}}\)
(6) The results in Table 2 indicate that HDO significantly outperformed the base models across the categories for the top 1, 5, and 10% optimization tasks. While MPNN actually produced negative top 1% AIRs over RS in some cases, these scores showed the limitations of an optimization approach based on only exploitation. In the case of 2(a) and 2(c), HDO has difficulties with the top 1, 5, and 10% tests over BO, and this could be the result of poor performance by MPNN; excluding this target case, the HDO outperformed the baseline. A few interesting reaction type trends can be seen across the model types. For instance, in the Suzuki-Miyaura reaction type, the training dataset for MPNN contains approximately 4,200,000 reaction conditions, whereas the Buchwald-Hartwig and Arylation datasets contain approximately 40,600. Therefore, MPNN models the found best combination faster than others. Similarly, the HDO model also performed well in the Suzuki-Miyaura reaction over BO, which still has a cold start problem. Likewise, for the top 5 and 10% yield searching cases in Arylation reaction 3(a), the AIR of HDO was higher than that of the human experts and BO, although BO is the best for the top 1%. Overall, the proposed model HDO found the best combination of conditions for a high yield compare to other optimization algorithms and human experts. (The results of 3(a) are shown in detail in Supplementary Information S4 sector) | Task 2: Validation of the HDO against five human chemists
This task was designed to additionally verify the performance of HDO in terms of general organic synthesis experiments compared to chemists. Task 1 involved a comparison of various optimization models and repeat simulations on the basis of Shields’s dataset, which includes yield results Fig. 3 Average cumulative maximum observed yields by the HDO (blue curve) and the average yield of the combination of conditions proposed by five experts (black dotted line). HDO required an average of 4.7 experiments to find the conditions with the same average yield obtained with the combination of conditions proposed by the five experts for the 22 reactions.
corresponding to all combinations of reaction conditions. However, it is difficult to obtain results from 1,000–4,000 different combinations of the conditions for one reaction. We therefore conducted experiments to create various additional target molecules to assess the stability of the HDO performance in comparison with that of humans. In Task 2, we used 22 different reactions to determine the number of experimental trials required to reach the yield results referenced by five specialists in the field of chemical synthesis. In these experiments, the four named reactions (the Suzuki-Miyaura, Buchwald-Hartwig, Chan-Lam, and Ullmann reactions) were used to synthesize the 22 target products. In addition, all the synthesis experiments included in Task 2 were also conducted on the HTE platform. In these tests, we focused on optimizing four aspects of the reaction context (catalyst, solvent, base, ligands) for each individual reaction. The results are plotted in Fig. 3. For the all 22 reactions, the best combination of reagents proposed by the five synthesis experts had a yield value of 64.48 on average. In comparison, HDO achieved this yield value in 4.7 experiments (i.e., by using 4.7 combinations of the reaction conditions) and average yield values of 82.18 were obtained by repeating the experiments 50 times (Fig. 3). The performance varied for each of the four aforementioned named reactions. Because the Reaxys database contains 158,605 training data for the Suzuki-Miyaura reaction, the HDO based on MPNN quickly identified the reaction conditions that delivered the yield of the combination of reaction conditions recommended by experts in an average of 4.22 times for 10 reactions. However, as is evident from the cumulative number of experiments, synthesis with the Suzuki-Miyaura reaction is easier than with the other named reactions. Therefore, the experts also tended to swiftly determine an effective combination of conditions for this reaction. In the eight experiments based on the Buchwald-Hartwig reaction, both HDO and the experts experienced difficulties to identify reaction conditions with high yield. Except for the reaction shown in 5(c) in Fig. 5, all of the reactions yielded poor results, yet even in these difficult situations, HDO found expert-level yields after 1.9 trails on average. For the Ullmann and Chan-Lam reactions, HDO required an average of 7.15 and 3.84 attempts, respectively, to identify the combination of conditions proposed by the experts. Details are provided in Fig. 4.
Finally, Fig. 5 presents examples of fragments of each of the named reactions. In the Suzuki-Miyaura reaction in 4(a), HDO proposed the same combination of conditions with the same yield value as the experts, and required a single experiment. In examples 5(c) and 7(a) for the Buchwald-Hartwig and Ullmann reactions, respectively, HDO obtained higher yield results than the experts with different combinations of reaction conditions and requiring two experimental trials in both cases. These are examples of the optimal combination of conditions found in the initial five experiments proposed by the MPNN models. In these cases, similar experiments were included in the training data, which are well-predicted examples. In marked contrast, the Chan-Lam coupling reaction in 6(b) required 20 experimental trails to identify reaction conditions with a conversion yield similar to that of the reference. Nevertheless, the proposed HDO algorithm demonstrated optimization performance comparable to that of the experts and was shown to have reliable and efficient navigation capabilities for a variety of combinations of reaction conditions.