Two cohorts of African giant pouched rats (C. ansorgei) obtained from APOPOs breeding colony served as subjects. The Experienced group (n = 6; mean age 4.4 years, SD = 0.19; 3 females) had prior experience detecting unrelated scents in a similar apparatus; the Naïve group (n = 5; mean age 0.4 years, SD = 0.001; 2 females) was naïve to all procedures. The sample size was constrained by the number of trained animals available within the APOPO breeding program. Although modest, this cohort was sufficient to demonstrate proof-of-principle. Rats were single- or pair-housed with a same-sex littermate within kennels designated for either male or female rats at APOPO’s Training and Research Headquarters on the campus of the Sokoine University of Agriculture in Tanzania. Home-cages were equipped with non-aromatic wood shavings, a clay sleeping pot, and an untreated hardwood climbing/gnawing structure as enrichment. Rats were maintained on a natural (roughly 12:12 h) light cycle with experimental sessions conducted during the daylight period. Water was available ad libitum. Animals were maintained at their free-feeding bodyweights with restricted access to food during daily training sessions. If the rat’s daily ration of food was not earned during training, supplemental food (fruit, vegetables, and 20 g of chow pellets; Specialty Feeds, Maintenance Food, Glen Forrest, Australia) was provided no less than 2 hours after the daily session. Additional protein (sundried fish, peanuts) and produce (e.g., mango, watermelon, tomato) was provided prior to days in which sessions were not conducted (e.g., on Friday). All experiments were conducted under approval of the Institutional Committee for Research Involving Animals at the Sokoine University of Agriculture (SUA/RES/APOPO/2018), following national standards and guidelines. All procedures complied with Tanzanian national guidelines and EU Directive 2010/63/EU. Rats received daily health monitoring with weekly vet oversight, and humane endpoints were strictly applied.

A custom engineered fully-automated line cage (FALCON; L210 × W41 × H52 cm, Figure 1) with glass walls and an aluminium lid and floor, mounted on four 96 cm tall legs was used for all experiments. See Supplementary Video S1 (available at OSF); consent was obtained from all personnel shown in the video, and training sessions posed no health risks requiring PPE. Ten circular sample holes (diameter 30 mm) were evenly spaced along the floor. Aluminium sample bars (L192 × W45 × H8 cm; L × H × W) could be positioned into a hinged bracket directly underneath the sample holes. Each bar had ten circular spaces (diameter 40 mm) which could be loaded with 20 mL borosilicate glass sample containers (Lenz Laborglasinstrumente; diameter 30 mm) and aligned with the ten holes along the floor of the cage. Each sample hole was fitted with an infrared photoelectric sensor which emitted a continuous auditory beep when broken, and an aluminium lid that could slide open to reveal the sample below. The first hole opened upon the start of the session and each subsequent hole opened after the infrared beam was broken by the rat inserting its nose in the hole immediately preceding it. The sample-hole lid was programmed to automatically close 500 ms after the rat removed its nose from the hole. On the left side of the cage, a pellet dispenser and magazine (ENV-203-94, MedAssociates, Georgia, VT) delivered pellets (5TCY OmniTreatTM) via a 20 cm plastic tube. The FALCON floor, walls and pellet dispenser were thoroughly cleaned between each rat session with 70% methylated spirits.

Material impacted with PHC was used as target samples from four different sources. The first was a Synthetically Aged crude oil sand mixture that was created by mixing 2.7 kg of crude oil with 14 kg of clean quartz ‘Play’ sand sourced from Ace Hardware. During this process the majority of the volatile fraction of the crude oils volatized, leaving a mixture more representative of a past crude oil spill which is fairly stable in both volatility and biodegradability. This resulted in a PHC concentration of 8.40% (gas chromatography analysis using Chevron modified method based on EPA 8015 for Total PHC C4-C44). Samples were extracted using dichloromethane (DCM), and the resulting extracts were then analysed using gas chromatography with flame ionization detection (GC-FID). Targets from the remaining three sources involved material as it was collected (i.e., naturally mixed with sand or soil) from three active remediation sites: California A at 2.10% PHC concentration, California B at 4.30%, and Illinois at 6.50%. Targets were presented in either undiluted form as described above or mixed with additional clean soil (described below) to dilute the PHC concentration (Table 1).

Clean soil collected from the same field locations served as control and dilutant for California B and Illinois field tests, however, was not available from the California A field location. Non-target samples and dilutant for California A and the Synthetically Aged material were a wide variety of soil types sourced from sixteen locations in Morogoro Region, Tanzania, air-dried and stored in sealed plastic buckets. The different soils were selected based on varying locations, composition, organic matter, soil type, and color. These were used as either i) clean soil, ii) dilutant for PHC samples, or iii) control for the presence of quartz and the tumbling and mixing procedure of diluted target samples by mixing the soil with equivalent volumes of quartz sand (i.e., the same volume of synthetically aged California material used to create target samples was matched for non-PHC controls). An additional control sample (‘Mix’) was created by combining various control mixtures (varying concentrations of quartz sand) within a single storage bucket.

To prepare sample mixtures, soil and PHC were placed in aluminium containers on a rock-tumbler for multiple 15 min cycles and crushed using a metal fork between tumbling rounds to disperse the oil and soil uniformly. Five mL of the mixture was inserted into a glass sample container and sealed prior to presentation to the rats. A subset of Synthetically Aged samples were sent for Total PHC laboratory analysis to Chevron Environmental Analysis Lab in Richmond, California to confirm our procedures achieved the desired PHC concentration (Supplementary Table S1) and validate rat performance results.

The primary independent variable was PHC concentration of soil samples, which was a within-subject factor that varied across training and tests. The dependent variable was the rats’ indication behavior (indicate vs. reject) when exposed to the samples.

Rats were trained to detect PHCs in soil through a multi-stage training program. Each group (Naïve and Experienced) progressed through the same training stages, with some adjustments made to account for age, experience and timeline differences. By the end of each stage, rats were able to perform increasingly complex detection tasks with proficient discrimination capabilities.

Stage 1. Clicker: Rats were trained to associate a click sound with food delivery. They progressed to the next stage once they consistently approached the food magazine within 2 seconds of the click across at least ten trials in a single session.

We evaluated rat ability to distinguish PHC from other naturally occurring organic materials, which can result in false-positives on conventional PHC laboratory analysis. An equal number of four soils, including familiar Soils 1 and 2 along with two novel soils pseudo-randomly selected from Soils 7-12 were presented during each of six sessions. Unlike all other training samples, Soils 7-12 were not sifted to remove excess organic matter (see Oil and Soil Samples section) and it was therefore hypothesized that they contained additional baseline quantities of organic materials which was further increased with the addition of random volumes of sticks, mosses, grass, etc. Ten targets were presented per session, including five samples each of 3.40% and 0.50% per soil (either one or two from each concentration and soil type combination, alternated each session; two blinds from familiar soil samples). One sample from each combination of soil type and concentration was mixed with additional organic materials (moss or compost soil). Non-targets were 8 subthreshold PHC 0.01% (two from each of the four soil types), 44 clean soil samples, as collected (not mixed with quartz sand; eleven from each soil type), 20 samples containing 3.40% control (ten each from Soils 1 and 2), and 18 Mix samples.

In a series of tests, we presented the rats with novel soil samples collected from the three active field sites (see Oil and Soil Samples section). For each field site, rats were first presented with undiluted samples as they were collected from the field, followed by a session in which these samples were diluted with additional soil to produce a range of PHC concentrations.

Only Experienced rats were tested with undiluted field samples sourced from California (California A). This test was conducted in a single session with ten total targets, including five local 3.40% samples (three of Soil 4 and two of Soil 3, one blind) and five undiluted California A (one blind). Non-targets were identical to the pre-test.

During the subsequent test (Test 2), a greater range of dilutions was presented. For Experienced rats, there was one session and the ten targets included two undiluted California A, one sample each of Soils 3 and 4 at 2.50% (1 blind), and six total samples containing 0.50% PHC (four California A, including one blind, and one sample each of Soils 3 and 4). Non-targets included 5 each of Soils 5 and 6; 13 each of controls 0.50% and 0.10% from Soils 3 and 4, 4 Mix; and 8 each of subthreshold 0.10% PHC from Soils 3, 4, and California A. Naïve rats completed two identical test sessions with the ten targets comprised of five undiluted California A (one blind) and five local 2.50% PHC samples (two each from Soils 1 and 2 plus one additional blind, alternated between soil type per session). Non-targets included five California A samples diluted to 0.10% PHC, 34 each of Soils 1 and 2 (sixteen control samples at 50:50 mixtures, sixteen control 3.40%, and two Synthetic sub-threshold 0.01% PHC), and seventeen Mix.

Unanticipated shipping delays caused by the COVID-19 pandemic affected delivery of oil samples used in Test 3. As a result, rats were given a 5.5 months break after completing Test 2 and before conducting Test 3. During this break, no training of any type was conducted while the rats continued to receive care and housing as described above. Following this break, rats completed all training stages in rapid sequence (see SI), however, sub-threshold samples were included. Additional soil types (13–22) were also introduced during this training.

Test 3 then introduced a second field sample from California (California B). Within a single test session, twelve targets were presented, including six familiar (Soil 20) samples of 2.5% PHC (one blind) and six undiluted California B (two blinds). The 88 non-targets included 32 clean California B and 56 samples of Soil 20 (twelve sub-threshold 0.10% PHC, twelve Mix, and eight each of controls 0.00%, 0.10%, 2.50%, and 3.40%). Test 4 immediately followed during the subsequent session and involved identical composition of targets and non-targets, but California B was replaced with the novel Illinois field sample.

Test 5 presented both California B and Illinois field samples (Supplementary Table S3) within a single session. As with Tests 3 and 4, this session included 12 total targets comprised of 6 undiluted field samples (3 from each source) and 6 that had been diluted to approximately 2.50% PHC (3 from each source). The 88 remaining samples within the session were non-targets, including 32 clean field samples from each site and 6 samples from each site that had been diluted to either 0.01% or 0.10% PHCs.

We then evaluated whether additional training could calibrate rats to site-specific differences in PHC scent profiles for greater delineation accuracy (as similarly required with existing technologies). Rats were presented with twelve targets per session, including four undiluted samples from each site (California B and Illinois, one blind each) and four California B diluted to 2.50%. Among the 88 non-targets were 29 clean samples from each site and 30 California B diluted to 0.10%. Combining the Experienced and Naïve groups, seven out of nine rats were required to indicate at least ten targets while committing no more than five false alarms over two consecutive sessions to advance to final testing.

Four final test sessions were then conducted to assess the impact of training with two different field sites on delineation accuracy for those sites. As with all prior tests, each session included 100 soil samples with only twelve representing targets. Sessions 1 and 2 presented a greater range of PHC concentrations within California B soil. Among the twelve targets were two undiluted samples from each site (California B and Illinois), two California B diluted to 2.50%, and three each (including one blind) of California B diluted to 1.00% and 0.50%. The 88 non-targets included 22 clean samples from each site, 22 California B diluted to 0.10%, and 22 California B diluted to 0.01%. Similar samples were presented during Sessions 3 and 4, however, diluted field samples only contained Illinois soil and only undiluted and clean California B samples were included.

All data were analyzed using R Studio. Generalized linear mixed models (GLMMs) were used to analyse the rats’ responses, with Indication (indicate vs. reject) as the outcome variable and Rat ID included as a random effect. Depending on the distribution of the data, models assumed either a binomial or Poisson structure, with arcsine transformation applied where necessary to address complete separation. When multiple rat groups were included, Rat ID was nested within Group. Model assumptions were checked through inspection of residuals and simplified random effects structures were used if singular fit issues occurred (Barr, 2013). Fixed effects varied according to the hypothesis tested. For example, models included Concentration when assessing detection thresholds, Stage (pre-vs. post-training) when comparing learning effects, Source (organic vs. PHCs or field vs. familiar soils) when comparing different sample types, and Blinding (blind vs. non-blind) when assessing potential experimenter influence. Interactions between factors were included when relevant. Post hoc comparisons were carried out using generalized linear hypothesis tests (GLHT) for main effects and emmeans for interactions, with Tukey’s Honestly Significant Difference (HSD) used to correct for multiple comparisons.

We first trained two cohorts of rats to determine if they could reliably distinguish between the smell of clean (non-PHC) soil and soil containing a range (2.50%–4.20%) of PHC concentrations (Table 2; Figure 2). After an average of 78 daily training sessions (SD = 57.13; or approximately 3.6 months; Figure 3), rats reliably evaluated 100 soil samples in a line-up within an average of 23 min (SD = 7.92). Across the final three training sessions, rats indicated 94.85% (SD = 0.22) of PHC samples and only 1.68% (SD = 0.13) of clean samples.

A general linear mixed model (GLMM) suggested no significant differences in rat detection behavior to control concentrations of quartz sand in clean soil (ranging from 0.00%–4.20% and Mix; GLMM: X ² = 1.85, p = 0.76) therefore all controls were grouped into a single category (clean 0.00% PHC). A significant difference in detection accuracy was found between PHC concentrations (0.00%, 2.50%, 3.40% and 4.20%; GLMM X ² = 351.64, p < 0.001), with all target concentrations significantly differing from 0.00% (Emmeans: all p <0 .001) but no differences in detection accuracy between concentrations of PHC impacted samples (p ≥ 0.47).

Trainer knowledge did not significantly affect detection accuracy (blinded petroleum hydrocarbon impacted soil [PHIS]: M = 93.33%, SD = 0.25; known PHIS: M = 97.92%, SD = 0.14; GLMM: F(1, 298) = 3.48, p = 0.06). Additionally, an outlier analysis (defined as no more than two standard deviations from the mean) did not identify any subjects.

Background soil type (Soils 1–4) also did not influence rat detection accuracy (GLMM: X ² = 2.34, p = 0.50) nor did it interact with Concentration (0.00%–4.20%; X ² = 11.13, p = 0.27). Thus soil composition had little effect on detection. Taken together, these results provide strong evidence rats can be trained to reliably signal the presence of PHCs within soil (see Supplementary Table S2 for individual scores).

To determine if rats can distinguish between concentrations of PHCs, we then introduced soil samples containing less than or equal to 0.50% PHCs. Control concentrations did not differ (GLMM: X² = 9.90, p = 0.13) and were grouped as 0.00%. There was a main effect of Concentration (GLMM: X² = 5192.30, p < 0.001; Figure 4). Rats readily detected PHIS at 0.50% PHCs (M = 98.33%, SD = 0.13; Figure 1D) with accuracy comparable to higher concentrations (Emmeans: all p = 0.99). Detection declined at 0.10% (M = 84.0%, SD = 0.37; p ≤ 0.01 vs. higher concentrations), though still significantly above lower levels (all p < 0.001). Finally, we found significantly fewer indications of 0.01% (M = 25.33%, SD = 0.44) compared to all other PHC samples but that they still indicated these samples significantly more than clean soil (0% PHC; M = 1.75%, SD = 0.13; all p < 0.001).

As found at the conclusion of training, there was no interaction between Soil Type and Concentration (GLMM: X ² = 6.11, p = 0.98) nor main effect of Soil Type (X ² = 2.82, p = 0.42), thus soil type again had minimal influence.

Following 48 training sessions in which rats were reinforced for indicating soil containing PHC levels ≥ 0.50% but no others, the same delineation test was repeated with the Experienced group only. During this second (post-training) test, all rats indicated 100% (SD = 0.00) of PHIS containing ≥0.50% PHCs while the average rat also hit 1.11% (SD = 0.10) or just 2 out of 180 soil samples that would otherwise be considered clean by our example threshold. Control concentrations again showed no differences (X ² = 4.60, p = 0.47), and were grouped as 0.00%. A significant interaction between Stage and Concentration (GLMM: X ² = 1008.91, p < 0.001) reflected a decrease in sub-threshold indications compared to the first delineation test: 0.01% (M = 0.00%; Emmeans: p < 0.001; Figure 2) and 0.10% (M = 16.67, SD = 0.38; p <0 .001). Subtle, though non-significant improvement in accuracy was also shown for 0.00% (M = 0.21, SD = 0.05; p = 0.09) and 4.20% (M = 100%; p = 0.09) concentrations. Soil Type again had no effect (LM: F (3, 1976) = 0.61, p = 0.61) nor interaction with Concentration (F (15, 1976) = 0.19, p = 0.99).

Adopting a rat team strategy (≥4/6 rats in agreement), the team of six rats achieved perfect delineation: 100% of the 20 PHIS samples identified and none of the 180 clean samples incorrectly flagged. Dilution accuracy was validated by lab analysis (mean deviation = 0.093%, range −0.79%–0.29%; Supplementary Table S1) with no misclassification. Collectively, these results suggest explicit training enabled accurate delineation of PHC thresholds.

We evaluated if additional hydrocarbons from non-petroleum organic sources influenced rat delineation accuracy by introducing organic material (moss, grasses, compost material) to both clean and impacted samples prior to rat evaluation. Control conditions did not differ (X ² = 4.29, p = 0.12), and were grouped as 0.00%.

Overall, rats indicated 97.22% (SD = 0.16) targets and 0.86% (SD = 0.09) of nontargets with only PHC added, and 90.42% (SD = 0.29) targets and 2.12% (SD = 0.14) nontargets with additional organic material (Figure 5). There was a main effect of Concentration (GLMM: X ² = 721.29, p < 0.001). All concentrations significantly differed except 0.00% vs. 0.01% (p = 0.76), with lower concentrations eliciting fewer hits (GLHT: p ≤ 0.006).

Rats hit more 0.00% samples when organics were added (Emmeans: b = 1.10, SE = 0.28, z = 3.89, p = 0.002). At the 0.50% threshold, the reverse occurred: fewer hits with organics vs. PHC only (b = 1.48, SE = 0.42, z = 3.52, p < 0.001). No differences emerged at 0.01% or 3.40% (both p ≥ 0.99).

Applying the same rat team strategy with a 4-rat cut-off, all false indications to 0.00% and 0.01% samples were eliminated. Indications on 3.40% were increased to 100% and there was also a slight increase in indications of 0.50%–83.33%. Thus while organics increased false positives for individuals, a team approach mitigated this effect.

We tested if rat accuracy was influenced by crude oil source using soil from three remediation sites (Figures 6, 7). Sample Type (source/concentration) predicted responses in all cases.

Two rats were excluded from analysis (one for failing to complete both sessions, one as an outlier >2 SD below mean performance).

For the remaining rats, when diluted samples were introduced, 98.81% (SD = 0.11) California B and 92.86% (SD = 0.26) Illinois targets were indicated, and 11.04% (SD = 0.31) California B and 2.92% (SD = 0.17) Illinois nontargets were indicated (Figure 8).

Concentration significantly predicted responses (X ² = 2988.92, p < 0.001): higher concentrations were indicated more than lower ones (p ≤ 0.03) (Figure 9), except undiluted and 2.50% which did not differ (p = 0.98). An interaction with Source (X² = 234.54, p < 0.001) was driven by significantly more false indications to 0.10% California B than Illinois samples (p < 0.001). A 5-rat cut-off eliminated most errors but still produced 25% false indications to 0.10% California B, while maintaining perfect discrimination at all other concentrations.

This discrepancy could reflect differences in either VOC profiles or soil substrate. GCMS analysis confirmed clear differences in volatiles (Supplementary Table S3): Illinois samples contained more light hydrocarbons, but California B may have had a higher proportion of the specific volatiles rats learned during training. Alternatively, sandy California B soils may have released volatiles more readily than clay-based Illinois soils, which would slow diffusion.

To test these alternatives, rats were explicitly trained on California B before being re-tested across additional dilutions (0.50, 1.00%). Post-training, indications to California B samples decreased overall (pre: 21.6%, post: 15.0%; X² = 101.70, p < 0.001), with a Stage × Concentration interaction (X² = 284.46, p < 0.001). Indications to 0.10% fell significantly (p < 0.001), though they remained higher than 0.01% and 0.00% (both p < 0.001). By contrast, indications to Illinois samples increased after California B training (pre: 13.7%, post: 23.6%; X² = 9.94, p = 0.002), driven by false positives at 0.10%.

Thus, while explicit training reduced errors on California B, it increased errors on Illinois, supporting the explanation that soil substrate and VOC release dynamics—rather than source-specific odor signatures alone—contributed to performance differences.