Figure 2. Dopamine responses to footshocks, but not safety cues, predict future behavior during negative reinforcement.

(A) Supervised machine learning was used to define whether dopamine signals around each behavioral event predicted current and/or future trial behavior. Support vector machine (SVM) algorithms were used. Datasets from dopamine recordings were coded depending on whether animals made a response or not in each trial. Data were then split into a training set and a testing set where the training set was used to construct an optimized hyperplane for behavior prediction and the testing set was used to test the accuracy of those predictions in an iterative fashion. (B-E) The SVM algorithm accurately predicted the current trial behavioral response based on the dopamine response to the positive reinforcement discriminative cue (S^d,sucrose; >75% accuracy; unpaired t-test ordered vs scrambled controls, t₃₈=5.34, p<0.0001; N=69 trials) with few errors (~20% false go and false no go predictions; opposite dot color). (F-I) For negative reinforcement, the algorithm was unable to use the S^d,shock dopamine response to predict whether animals would or would not respond in the current trial (~45% accuracy; unpaired t-test ordered vs scrambled controls, t₃₈=0.000, p>0.99; N=69 trials). (J) During negative reinforcement, when animals did not respond during the S^d,shock they received a footshock. The SVM algorithm was able to predict whether an animal would nose-poke during the S^d,shock or not on the next trial based on the dopamine response to the shock itself (independent t-test, t₁₉=7.95, p<0.0001; N=30 trials; see Figure S3 for additional analyses). (K) The SVM algorithm was unable to predict behavior in the subsequent trial based on the dopamine response to the safety cue on trials when animals did not respond correctly during S^d,shock, suggesting that dopamine responses to the safety cue are not error-based learning signals (>75% accuracy; independent t-test, t₁₉=0.55, p=0.59; ~50% accuracy; N=30 trials; see Figure S3 for additional analyses). Data represented as mean ± S.E.M., **** p < 0.0001; ns, not significant.