<h4>Background</h4> HIV transmission is characterised by a low per-act probability, a relatively high proportion of multiple variant transmission events, and a plateauing of transmission risk at high viral loads. No existing mechanistic model can simultaneously recapitulate all of these observations, thereby limiting our ability to predict unobserved transmission phenomena and evaluate prevention strategies. <h4>Methods</h4> We developed a suite of mathematical models that encode an empirically plausible set of transmission mechanisms and then fit these models within a Bayesian framework to available epidemiological data to identify which set of mechanisms are sufficient to recapitulate the data. Following formal model comparison, we embedded the best-fit model into a phylodynamic framework and calibrated it using Approximate Bayesian Computation, to assess whether phylogenetic trees from individual transmission pairs were both consistent with the model and informative. Finally, we further validated our most likely model against two large prospective studies (PARTNER1 and STEP). <h4>Results</h4> Our calibrated model predicts that for each systemic infection, approximately four to five transient infections occur—exposure events in which viral replication occurs but is stochastically extinguished—consistent with indirect empirical evidence from the STEP vaccine trial. The model predicts a transmission rate of fewer than 0.05 systemic infections per 100 couple-years follow up from individuals with undetectable viral load, providing a mechanistic basis for the negligible risk observed in the PART-NER1 study. The model also predicts a strong link between the number of viral particles transmitted and the number of variants establishing infection, modulated by the transmitter’s infection stage. Recalibrating for men who have sex with men indicated that higher transmission rates in this population are explained by a single parameter: a greater probability of permissive c SinceΘis uniform, the probability that the transmitter is in stages ∈ {E, A, L}is: Pr(S = s|V = v) = τs τE +τ A(vA) +τ L .(4) 2.2 Intermittent susceptibility model In the intermittent susceptibility model, HIV infection can only occur during a small frac- tion of sexual exposures—specifically those that occur when conditions are permissive for transmission. Our intermittent susceptibility model follows the same structure as described elsewhere [7], with two modifications: first, viral loads vary between individuals in the pop- ulation in all infection stages, consistent with epidemiological data [31, 32]. Second, for mathematical tractability, we approximate the number of virions transmitted to the exposed partner using a Poisson distribution (rather than a binomial distribution) because the viral load in the genital mucosa,v, is large and the per-virion acquisition probability,p, is small. In this model, the probability thatN1 virions establish infection during a single bility combined with infrequent permissive conditions for infection, likely reflecting stringent transmission barriers that fluctuate through time. Second, the relatively high probability of multiple founder variants was attributed to relaxed bottlenecks during these rare per- missive exposures allowing multiple virions to establish infection simultaneously. Third, the plateauing of transmission rates at high viral loads was explained by target cell limitation at the infection site. Fourth, stage-specific differences in both transmission probability and the frequency of multiple founder variants were attributed to variations in the probabil- ity that infected cells establish systemic infection. Beyond reconciling these observations, the calibrated model yields novel predictions: transient infections occur approximately five times more frequently than systemic infections, viral suppression reduces transmission risk by several hundredfold, and higher transmission rates in MSM populations are primarily explained by a greater probability of permissive conditions for infection, that typically occur infrequently. To our knowledge, this is the first model to reconcile all these key epidemiological obser- vations, and it yields additional insights into HIV transmission. The model predicts a close association between a transmitter’s viral load and the number of virions infecting exposed partners, with the transmitter’s infection stage influencing the number of transmitted vari- ants. These patterns are broadly consistent with previous theoretical predictions [7], with two notable differences: first, our model permits a wider range of particle counts to cross the transmission barrier, reflecting a trade-off between particle counts and the per-cell establish- ment probability; second, during early infection, our model predicts a higher frequency of two to three founding variants [7], a consequence of fitting to the elevated relative hazard of multiple founders observed at this stage. These findings offer a mechanistic interpretation of stage-specific transmission risk. The probability that an infected cell establishes systemic infection peaks during early infection, aligning with clinical observations that transmission risk during this stage exceeds what would be expected from elevated viral load alone [40]. This elevated establishment proba- bility may reflect the transmission of highly infectious founder-like variants that are subse- quently lost through immune-driven evolution and antibody neutralisation [41–43]. Consis- tent with this, transmitters in the asymptomatic stage showed markedly lower per-cell estab- 19 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 23, 2026. ; https://doi.org/10.64898/2026.03.20.26348904doi: medRxiv preprint